include llama.cpp since there were some edits

llama.cpp/ggml/src/ggml-openvino/.clang-format

@@ -0,0 +1,149 @@
+---
+# Override root .clang-format
+AlignConsecutiveAssignments: false
+AlignConsecutiveDeclarations: false
+AccessModifierOffset: -4
+
+Language:        Cpp
+AlignAfterOpenBracket: Align
+AlignArrayOfStructures: Left
+AlignConsecutiveBitFields: AcrossComments
+AlignConsecutiveMacros: AcrossComments
+# AlignConsecutiveShortCaseStatements: AcrossComments
+AlignEscapedNewlines: Left # LeftWithLastLine
+AlignOperands:   Align
+AlignTrailingComments:
+  Kind: Always
+  OverEmptyLines: 1
+AllowAllArgumentsOnNextLine: true
+AllowAllParametersOfDeclarationOnNextLine: false
+# AllowBreakBeforeNoexceptSpecifier: OnlyWithParen
+AllowShortBlocksOnASingleLine: Never
+AllowShortCaseLabelsOnASingleLine: false
+AllowShortFunctionsOnASingleLine: Inline
+AllowShortIfStatementsOnASingleLine: Never
+AllowShortLambdasOnASingleLine: Inline
+AllowShortLoopsOnASingleLine: false
+AlwaysBreakBeforeMultilineStrings: true
+# Treat CUDA keywords/attributes as "attribute macros" and avoid breaking lines inside them
+AttributeMacros:
+  - __host__
+  - __device__
+  - __global__
+  - __forceinline__
+  - __launch_bounds__
+BinPackArguments: true
+BinPackParameters: false # OnePerLine
+BitFieldColonSpacing: Both
+# BreakAdjacentStringLiterals: true
+BreakAfterAttributes: Never
+BreakBeforeBinaryOperators: None
+BreakBeforeInlineASMColon: OnlyMultiline
+BreakBeforeTernaryOperators: false
+# BreakBinaryOperations: Never
+BreakConstructorInitializers: AfterColon
+# BreakFunctionDefinitionParameters: false
+BreakInheritanceList: AfterComma
+BreakStringLiterals: true
+# BreakTemplateDeclarations: Yes
+ColumnLimit:     120
+CommentPragmas:  '^ IWYU pragma:'
+CompactNamespaces: false
+ConstructorInitializerIndentWidth: 4
+ContinuationIndentWidth: 4
+DerivePointerAlignment: false
+DisableFormat:   false
+EmptyLineBeforeAccessModifier: Leave
+EmptyLineAfterAccessModifier: Never
+ExperimentalAutoDetectBinPacking: false
+FixNamespaceComments: true
+IncludeBlocks:   Regroup
+IncludeCategories:
+  - Regex:           '".*"'
+    Priority:        1
+    SortPriority:    0
+  - Regex:           '^<.*\.h>'
+    Priority:        2
+    SortPriority:    0
+  - Regex:           '^<.*'
+    Priority:        3
+    SortPriority:    0
+  - Regex:           '.*'
+    Priority:        4
+    SortPriority:    0
+IncludeIsMainRegex: '([-_](test|unittest))?$'
+IncludeIsMainSourceRegex: ''
+IndentAccessModifiers: false
+IndentExternBlock: NoIndent
+IndentGotoLabels: false
+IndentPPDirectives: AfterHash
+IndentWidth:     4
+IndentWrappedFunctionNames: false
+InsertBraces:    true # NOTE: may lead to incorrect formatting
+InsertNewlineAtEOF: true
+JavaScriptQuotes: Leave
+JavaScriptWrapImports: true
+KeepEmptyLinesAtTheStartOfBlocks: false
+LambdaBodyIndentation: Signature
+LineEnding: LF
+MacroBlockBegin: ''
+MacroBlockEnd:   ''
+MaxEmptyLinesToKeep: 1
+NamespaceIndentation: None
+ObjCBinPackProtocolList: Auto
+ObjCBlockIndentWidth: 4
+ObjCSpaceAfterProperty: true
+ObjCSpaceBeforeProtocolList: true
+PPIndentWidth: -1
+PackConstructorInitializers: CurrentLine
+PenaltyBreakAssignment: 2
+PenaltyBreakBeforeFirstCallParameter: 1
+PenaltyBreakComment: 300
+PenaltyBreakFirstLessLess: 120
+PenaltyBreakString: 1000
+PenaltyBreakTemplateDeclaration: 10
+PenaltyExcessCharacter: 1000000
+PenaltyReturnTypeOnItsOwnLine: 200
+PointerAlignment: Middle
+QualifierAlignment: Left
+#QualifierOrder: ['static', 'inline', 'friend', 'constexpr', 'const', 'volatile', 'type', 'restrict']
+RawStringFormats:
+  - Language:        Cpp
+    Delimiters:
+      - cc
+      - CC
+      - cpp
+      - Cpp
+      - CPP
+      - 'c++'
+      - 'C++'
+    CanonicalDelimiter: ''
+ReferenceAlignment: Middle
+ReflowComments:  false # IndentOnly
+SeparateDefinitionBlocks: Always
+SortIncludes:    CaseInsensitive
+SortUsingDeclarations: LexicographicNumeric
+SpaceAfterCStyleCast: true
+SpaceAfterLogicalNot: false
+SpaceAfterTemplateKeyword: true
+SpaceBeforeAssignmentOperators: true
+SpaceBeforeCpp11BracedList: false
+SpaceBeforeCtorInitializerColon: true
+SpaceBeforeInheritanceColon: true
+SpaceBeforeParens: ControlStatements
+SpaceBeforeRangeBasedForLoopColon: true
+SpaceInEmptyBlock: false
+SpaceInEmptyParentheses: false
+SpacesBeforeTrailingComments: 2
+SpacesInAngles:  Never
+SpacesInLineCommentPrefix:
+  Minimum: 1
+  Maximum: -1
+SpacesInParentheses: false
+SpacesInSquareBrackets: false
+SpaceBeforeSquareBrackets: false
+Standard:        c++17
+TabWidth:        4
+UseTab:          Never
+WhitespaceSensitiveMacros: ['STRINGIZE']
+...

llama.cpp/ggml/src/ggml-openvino/CMakeLists.txt

@@ -0,0 +1,33 @@
+find_package(OpenVINO REQUIRED COMPONENTS Runtime Threading)
+find_package(OpenCL REQUIRED)
+
+message(STATUS "Found OpenVINO: ${OpenVINO_DIR} (found version \"${OpenVINO_VERSION}\")")
+
+file(GLOB_RECURSE GGML_HEADERS_OPENVINO "*.h" "*.hpp")
+file(GLOB_RECURSE GGML_SOURCES_OPENVINO "*.cpp")
+
+ggml_add_backend_library(ggml-openvino
+    ${GGML_SOURCES_OPENVINO}
+    ${GGML_HEADERS_OPENVINO}
+)
+
+target_link_libraries(ggml-openvino PRIVATE openvino::runtime openvino::threading OpenCL::OpenCL)
+
+# the OpenVINO RTTI macros take one argument and leave __VA_ARGS__ empty, which -Wpedantic reports
+if (CMAKE_CXX_COMPILER_ID MATCHES "Clang" OR CMAKE_CXX_COMPILER_ID STREQUAL "IntelLLVM")
+    target_compile_options(ggml-openvino PRIVATE -Wno-gnu-zero-variadic-macro-arguments)
+elseif (CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
+    target_compile_options(ggml-openvino PRIVATE -Wno-pedantic)
+endif()
+
+if (WIN32)
+    target_link_libraries(ggml-openvino PRIVATE psapi)
+endif()
+
+if (GGML_OPENVINO)
+    if (CMAKE_SYSTEM_PROCESSOR STREQUAL "aarch64")
+    elseif (CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64" OR CMAKE_SYSTEM_PROCESSOR STREQUAL "amd64" OR CMAKE_SYSTEM_PROCESSOR STREQUAL "AMD64")
+    else()
+        message(FATAL_ERROR "OpenVINO: OpenVINO toolkit supports x86-64 and arm64 but not ${CMAKE_SYSTEM_PROCESSOR}")
+    endif()
+endif()

llama.cpp/ggml/src/ggml-openvino/README.md

@@ -0,0 +1,44 @@
+# Compiled model cache
+
+`GGML_OPENVINO_COMPILED_MODEL_CACHE_DIR` exports compiled CPU/GPU graphs with their weights. It bypasses the plugin-level `GGML_OPENVINO_CACHE_DIR` and uses `OPTIMIZE_SPEED`, so weightless caching is disabled.
+
+One directory can hold blobs for different models and compilation settings. Run each intended workload once to export its dynamic graph:
+
+```sh
+GGML_OPENVINO_DEVICE=GPU \
+GGML_OPENVINO_NATIVE_SOFTPLUS=1 \
+GGML_OPENVINO_DISABLE_KV_SLICE=1 \
+GGML_OPENVINO_REQUANT_KQUANT=q4_asym64_all \
+GGML_OPENVINO_COMPILED_MODEL_CACHE_DIR=/path/to/qwen-cache \
+./build/ReleaseOV/bin/llama-bench -m /path/to/model.gguf -r 1
+```
+
+`GGML_OPENVINO_SPILL_DIR` remains optional for this first run. Wait for the `model cache WROTE` message and completion of the workload before stopping it. Compatible prefill and decode graphs share one blob and manifest. A graph with different ports or incompatible shapes gets an exact entry instead; interrupted exports are not cache hits.
+
+On later runs, supply the same compilation settings and enable `GGML_OPENVINO_COMPILED_MODEL_CACHE_ONLY=1`:
+
+```sh
+GGML_OPENVINO_DEVICE=GPU \
+GGML_OPENVINO_NATIVE_SOFTPLUS=1 \
+GGML_OPENVINO_DISABLE_KV_SLICE=1 \
+GGML_OPENVINO_REQUANT_KQUANT=q4_asym64_all \
+GGML_OPENVINO_COMPILED_MODEL_CACHE_DIR=/path/to/qwen-cache \
+GGML_OPENVINO_COMPILED_MODEL_CACHE_ONLY=1 \
+./build/ReleaseOV/bin/llama-bench -m /path/to/model.gguf -r 1
+```
+
+Cache-only mode allocates backend address space without filling weight pages. On Windows, this also uses system commit capacity. The model-buffer size in the loader log is this virtual size. Weight uploads only record source identity; they do not read or requantize the weights. Graph conversion and compilation are skipped. Runtime buffers are still allocated and populated normally.
+
+Cache-only mode uses the settings provided by the current process. Keep these values exactly the same, including set versus unset: `GGML_OPENVINO_REQUANT_KQUANT`, `GGML_OPENVINO_NATIVE_SOFTPLUS`, `GGML_OPENVINO_DISABLE_KV_SLICE`, `GGML_OPENVINO_MANUAL_GQA_ATTN`, `GGML_OPENVINO_STATEFUL_EXECUTION`, `GGML_OPENVINO_DISABLE_KV_STATE_RELAYOUT`, `GGML_OPENVINO_DISABLE_REMOTE_OUTPUTS`, `GGML_OPENVINO_REDUCE_COMPILE_MEM`, `GGML_OPENVINO_MEMORY_OPTIMIZE`, `GGML_OPENVINO_PROFILING`, and `GGML_OPENVINO_DEBUG_NODE`. On GPU, also repeat `GGML_OPENVINO_MOE_OP=0` if used. `GGML_OPENVINO_SPILL_DIR` is optional on the first run and ignored in cache-only mode; host-weight release is disabled in cache-only mode.
+
+A missing or incompatible graph fails with an error instead of compiling with absent weights. The fingerprint uses the dynamic graph's topology, ports, model parameters, weights, settings, and OpenVINO version; changing only dynamic token or KV sizes does not require a new entry. A different workload can still require another graph; populate it first without cache-only mode.
+
+## Restrictions
+
+- Cache-only mode requires Linux or Windows and mmap loading (`--load-mode mmap`, or the default when all selected devices support mmap). Do not use tensor validation or mlock when trying to avoid weight reads.
+- On Windows, the backend commits virtual memory for its buffers without touching weight pages. Large models can still reach the system commit limit.
+- The model must execute entirely on OpenVINO, with dynamic CPU/GPU graphs and in-process caching enabled. Static/NPU execution and CPU fallback are unsupported in cache-only mode.
+- GGUF metadata, tokenizer data, tensor descriptors, and graph construction are still needed. The llama.cpp loader is unchanged: depending on its prefetch settings, it may request pages with `MAP_POPULATE` or read-ahead on Linux, or `PrefetchVirtualMemory` on Windows. Non-mmap loading also reads the payload before the backend sees it.
+- File identity, size, modification/change timestamps, tensor offsets, graph structure, settings, and OpenVINO version identify cache entries. Linux uses device/inode and Windows uses volume serial/file index. Replacing, copying, or modifying a GGUF invalidates its entries. This avoids reading weight bytes and ties the cache to the local source files. Keep those files unchanged throughout loading and inference.
+- Use the same target device and compatible OpenVINO/plugin installation. Import support depends on the plugin; the tested CPU plugin cannot import MoE graphs containing `GatherMatmulCompressed`. GPU MoE and CPU dense graph imports were tested.
+- Blobs contain weights and can approach model size for each compiled graph. Import still reads those blobs and initializes the device.

llama.cpp/ggml/src/ggml-openvino/ggml-decoder.h

@@ -0,0 +1,507 @@
+#pragma once
+
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+#include "ggml.h"
+#include "openvino/decoder.h"
+
+#include <cstdint>
+#include <cstring>
+#include <map>
+#include <memory>
+#include <openvino/core/partial_shape.hpp>
+#include <optional>
+#include <set>
+#include <string>
+#include <vector>
+
+struct ModelParams {
+    int ctx = -1;
+    int ctx_per_seq = -1;
+    int ctx_per_seq_swa = -1;
+    int n_seq = 1;
+    int n_heads_kv = -1;
+    // Per-layer KV head count. gemma-4 12B interleaves 8 x 256 sliding layers with 1 x 512
+    // full-attention layers, so no single scalar describes every layer. Keyed by layer, not by
+    // layer TYPE, because the SWA classification depends on the context size (extents tie at a
+    // small -c) while the head count does not.
+    std::map<int, int> n_heads_kv_per_layer;
+    int head_size = -1;
+    int state_size = -1;  // for SSM molels, eg qwen35
+    int32_t rope_params[16] = {};
+    int n_rs_slots = -1;
+    bool has_rs_rollback = false;
+    bool mixed_rope_params = false;
+    bool is_cacheless_attn = false;
+    std::vector<int> swa_layers;
+    // The sliding-window mask tensor, identified in compute_llm_params() by grouping attention
+    // layers on the mask they consume. Only used to tell the two masks apart when naming OV
+    // parameters -- both carry the same tensor name. Null when the graph has a single mask.
+    const ggml_tensor * swa_mask = nullptr;
+
+    std::vector<std::string> kv_names;
+    size_t kv_buffer_ctx_id = 0;
+
+    bool same_rope_params(const ModelParams & other) const {
+        return mixed_rope_params == other.mixed_rope_params &&
+               memcmp(rope_params, other.rope_params, sizeof(int32_t) * 16) == 0;
+    }
+
+    bool can_reuse_dynamically(const ModelParams & other) const {
+        return same_rope_params(other) && n_rs_slots == other.n_rs_slots &&
+               has_rs_rollback == other.has_rs_rollback;
+    }
+
+    bool can_reuse_statically(const ModelParams & other) const {
+        return same_rope_params(other) && ctx == other.ctx && n_rs_slots == other.n_rs_slots &&
+               has_rs_rollback == other.has_rs_rollback;
+    }
+
+    bool kv_buffer_changed(const ModelParams & other) const { return kv_buffer_ctx_id != other.kv_buffer_ctx_id; }
+};
+
+struct ComputeParams {
+    int n_seq_active = 1;
+    int seq_active_start = 0;
+    int attention_size = -1;
+    int attention_size_swa = -1;
+    int attention_size_static = -1;  // encoder/cross-attn KV fill level (whisper)
+    // Sliding window width, read back from the band of ggml's own SWA mask. ggml never passes
+    // n_swa down to a backend, but fill_mask() bakes it into the mask contents, so the widest
+    // unmasked row recovers it. Shorter than n_swa while the sequence is still short, which is
+    // harmless: every causal pair is inside the window then anyway.
+    int swa_window = -1;
+    int input_len = -1;
+    int token_len_per_seq = -1;
+    int past_kv_len = -1;
+    int output_len = 1;
+
+    int cache_rs_reset_idx = -1;
+    int cache_rs_reset_len = -1;
+    // SSM/DeltaNet models otionally clear cache_r and cache_s of certain slots in the cgraph
+    // 3: [ 18432,     4,     1,     1] RESHAPE              cache_r_l0 (reshaped)
+    //    [ 18432,     4,     1,     1]            0: NONE        cache_r_l0
+    // 4: [ 18432,     1,     1,     1] VIEW                 cache_r_l0 (reshaped) (view)
+    //    [ 18432,     4,     1,     1]            0: RESHAPE     cache_r_l0 (reshaped)
+    // 5: [ 18432,     1,     1,     1] SCALE                cache_r_l0 (reshaped) (view) (view)
+    //    [ 18432,     1,     1,     1]            0: VIEW        cache_r_l0 (reshaped) (view)
+
+    int s_copy_active_slot_len = -1;
+    // SSM/DeltaNet models otionally reorder slots of state cache, to make the active slots contiguous
+    // leaf_5 is the inp->s_copy in llama-graph.cpp, eg if there are 8 slots in total and slot 3 and 7
+    // are active in the current batch, leaf_5 will be [3, 7, 5, 6, 4]
+    //  6: [     2,     1,     1,     1] VIEW                  (view)
+    //      [     2,     1,     1,     1]            0: NONE        leaf_5
+    //  7: [ 18432,     2,     1,     1] GET_ROWS             conv_states-0
+    //      [ 18432,     4,     1,     1]            0: RESHAPE     cache_r_l0 (reshaped)
+    //      [     2,     1,     1,     1]            1: VIEW         (view)
+    //  8: [     0,     1,     1,     1] VIEW                  (view)
+    //      [     2,     1,     1,     1]            0: NONE        leaf_5
+    //  9: [ 18432,     0,     1,     1] GET_ROWS             node_9
+    //      [ 18432,     4,     1,     1]            0: RESHAPE     cache_r_l0 (reshaped)
+    //      [     0,     1,     1,     1]            1: VIEW         (view)
+    // 10: [ 18432,     0,     1,     1] VIEW                 cache_r_l0 (view)
+    //      [ 18432,     4,     1,     1]            0: NONE        cache_r_l0
+    // 11: [ 18432,     0,     1,     1] CPY                  cache_r_l0 (view) (copy of )
+    //      [ 18432,     0,     1,     1]            0: GET_ROWS    node_9
+    //      [ 18432,     0,     1,     1]            1: VIEW        cache_r_l0 (view)
+
+    struct RsWriteback {
+        int slot_begin = 0;  // first cache slot written by the CPY
+        int src_begin = -1;  // first source column copied by a conv-state CPY
+    };
+
+    std::map<std::string, RsWriteback> rs_writebacks;
+    // Destination slot offset of each state cache writeback CPY node, keyed by node name. It
+    // changes with the batch (kv head, active sequence count) and, with rollback enabled
+    // (cparams.n_rs_seq > 0), the conv state is written back once per snapshot slot. Passed to the
+    // cached model as a runtime input. Dynamic models also receive the source-side offset; static
+    // models use a fixed end-anchored offset in the translator.
+};
+
+// defined below; declared here because GgmlOvDecoder uses it inline
+std::optional<int> extract_layer_from_name(const std::string & name);
+
+// detects the MoE expert-plane-sum ADD chain (see definition); used by supports_op too
+bool is_moe_expert_sum_add(const ggml_tensor * node);
+
+class GgmlOvDecoder : public ov::frontend::ggml::GgmlDecoder {
+public:
+    static std::string get_tensor_name(const ggml_cgraph * cgraph, const ggml_tensor * tensor);
+    struct NodeInfo {
+        ggml_tensor * node;
+        std::string node_name;
+        std::string node_op_type;
+        std::map<std::string, ggml_tensor *> node_inputs;
+        std::map<std::string, std::vector<std::pair<std::string, ggml_tensor *>>> node_inputs_views;
+        std::vector<std::string> node_inputs_names;
+        int node_op_case = 0;
+        void * data_addr;
+    };
+
+    // Graph decoder
+    GgmlOvDecoder(ggml_cgraph * cgraph,
+                  ModelParams & model_params,
+                  ComputeParams & compute_params,
+                  std::map<std::string, std::shared_ptr<ov::Node>> & model_weights,
+                  bool is_static,
+                  bool is_stateful = false,
+                  bool model_is_splitted = false,
+                  bool is_prefill = false,
+                  int prefill_chunk_size = 256);
+
+    // Naive graph decoder
+    GgmlOvDecoder(ggml_cgraph * cgraph, std::map<std::string, std::shared_ptr<ov::Node>> & model_weights);
+
+    virtual ov::Any get_attribute(const std::string & name) const override {
+        return nullptr;
+        GGML_UNUSED(name);
+    }
+
+    virtual ov::PartialShape get_input_shape(int node_idx, const std::string & name) const override;
+
+    virtual std::vector<size_t> get_input_stride(int node_idx, const std::string & name) const override;
+
+    virtual size_t get_view_input_size(int node_idx, const std::string & name) const override;
+
+    virtual size_t get_view_input_offset(int node_idx, const std::string & name, size_t view_index) const override;
+
+    virtual size_t get_view_input_src_offset(int node_idx, const std::string & name, size_t view_index) const override;
+
+    virtual std::vector<size_t> get_view_input_stride(int node_idx,
+                                                      const std::string & name,
+                                                      size_t view_index) const override;
+
+    virtual std::vector<size_t> get_view_input_src_stride(int node_idx,
+                                                          const std::string & name,
+                                                          size_t view_index) const override;
+
+    virtual ov::Shape get_view_input_ggml_shape(int node_idx,
+                                                const std::string & name,
+                                                size_t view_index) const override;
+
+    virtual ov::Shape get_view_input_src_ggml_shape(int node_idx,
+                                                    const std::string & name,
+                                                    size_t view_index) const override;
+
+    virtual ov::PartialShape get_view_input_ov_shape(int node_idx,
+                                                     const std::string & name,
+                                                     size_t view_index) const override;
+
+    virtual ov::PartialShape get_view_input_src_ov_shape(int node_idx,
+                                                         const std::string & name,
+                                                         size_t view_index) const override;
+
+    virtual std::string get_view_input_name(int node_idx, const std::string & name, size_t view_index) const override;
+
+    virtual std::string get_view_input_src_name(int node_idx,
+                                                const std::string & name,
+                                                size_t view_index) const override;
+
+    virtual ov::element::Type get_input_type(int node_idx, const std::string & name) const override;
+
+    virtual size_t get_input_size() const override;
+
+    virtual size_t get_input_size(int node_idx) const override;
+
+    virtual void get_input_node(size_t input_port_idx,
+                                std::string & producer_name,
+                                std::string & producer_output_port_name,
+                                size_t & producer_output_port_index) const override {
+        GGML_UNUSED(input_port_idx);
+        GGML_UNUSED(producer_name);
+        GGML_UNUSED(producer_output_port_name);
+        GGML_UNUSED(producer_output_port_index);
+    }
+
+    virtual std::vector<std::string> get_input_names(int node_idx) const override;
+
+    virtual ov::PartialShape get_output_shape(int node_idx) const override;
+
+    virtual ov::element::Type get_output_type(int node_idx) const override;
+
+    virtual std::vector<size_t> get_output_stride(int node_idx) const override;
+
+    virtual int32_t * get_input_op_params(int node_idx, const std::string & name) const override;
+
+    virtual int32_t * get_output_op_params(int node_idx) const override;
+
+    virtual size_t get_output_op_offset(int node_idx) const override;
+
+    virtual std::vector<std::string> get_output_names(int node_idx) const override;
+
+    virtual std::string get_inplace_op_src(int node_idx) const override;
+
+    virtual bool is_view_like_alias_of(int node_idx, const std::string & view_src_name) const override;
+
+    virtual const std::string & get_op_type() const override;
+
+    virtual const std::string & get_op_type(int node_idx) const override;
+
+    virtual const std::string & get_op_name() const override;
+
+    virtual const std::string & get_op_name(int node_idx) const override;
+
+    virtual int32_t get_op_dynamic_dim(int node_idx) const override;
+
+    virtual void visit_subgraph(
+        std::function<void(std::shared_ptr<GgmlDecoder>, int node_idx)> node_visitor) const override;
+
+    ggml_tensor * get_input_ggml_tensor(const std::string & name) const { return m_inputs.at(name); }
+
+    virtual int get_op_case(int node_idx) const override { return m_node_info_list[node_idx].node_op_case; }
+
+    virtual const std::map<std::string, ov::frontend::ggml::ModelInputInfo> & get_model_inputs() const override {
+        return m_model_inputs;
+    }
+
+    virtual const std::map<std::string, ov::frontend::ggml::ModelExtraInputInfo> & get_model_extra_inputs() const override {
+        return m_model_extra_inputs;
+    }
+
+    virtual const std::map<std::string, std::shared_ptr<ov::Node>> & get_model_weights() const override {
+        return m_model_weights;
+    }
+
+    virtual std::set<std::string> get_model_output_names() const override { return m_model_output_names; }
+
+    const std::map<std::string, ggml_tensor *> & get_model_outputs() const { return m_model_outputs; }
+
+    virtual int get_ctx_size() const { return m_model_params.ctx; }
+
+    virtual int get_ctx_per_seq() const { return m_model_params.ctx_per_seq; }
+
+    virtual int get_ctx_per_seq_swa() const { return m_model_params.ctx_per_seq_swa; }
+
+    virtual int get_n_seq() const { return m_model_params.n_seq; }
+
+    virtual int is_swa_layer(int layer) const override {
+        return std::find(m_model_params.swa_layers.begin(), m_model_params.swa_layers.end(), layer) !=
+               m_model_params.swa_layers.end();
+    }
+
+    // KV head count for one layer. Sliding and full layers can differ (gemma-4 12B), so callers
+    // that reinterpret a KV buffer must use this and not the model-level n_heads_kv.
+    int get_n_heads_kv_for_layer(int layer) const {
+        auto it = m_model_params.n_heads_kv_per_layer.find(layer);
+        return it != m_model_params.n_heads_kv_per_layer.end() ? it->second : m_model_params.n_heads_kv;
+    }
+
+    // Same, for a KV cache tensor: its layer comes from the leaf name (cache_k_l<N>).
+    int get_n_heads_kv_for_tensor(const ggml_tensor * kv_tensor) const {
+        if (auto layer = extract_layer_from_name(std::string(kv_tensor->name)); layer.has_value()) {
+            return get_n_heads_kv_for_layer(layer.value());
+        }
+        return m_model_params.n_heads_kv;
+    }
+
+    int get_past_kv_len() const { return m_compute_params.past_kv_len; }
+
+    int get_input_len() const { return m_compute_params.input_len; }
+
+    virtual int32_t * get_rope_params() const override { return const_cast<int32_t *>(m_model_params.rope_params); }
+
+    virtual bool has_mixed_rope_params() const override { return m_model_params.mixed_rope_params; }
+
+    virtual int get_ssm_state_size() const override { return m_model_params.state_size; }
+
+    virtual std::map<std::string, std::string> get_kv_param_res_names() const override;
+
+    virtual bool is_static() const override { return m_is_static; }
+
+    virtual bool is_stateful() const override { return m_is_stateful; }
+
+    int get_static_n_tokens() const { return m_is_prefill ? m_prefill_chunk_size : 1; }
+
+    virtual bool is_splited_model() const override { return m_model_is_splitted; }
+
+    ov::PartialShape get_graph_input_shape(const ggml_tensor * op,
+                                           const ggml_tensor * input,
+                                           int dynamic_dim_index = -1) const;
+
+    static void dump_cgraph(const ggml_cgraph * cgraph, std::string & filename);
+
+    static std::shared_ptr<ov::Node> create_weight_node(ggml_tensor * tensor, bool naive = false);
+
+    static std::map<std::string, std::shared_ptr<ov::Node>> create_weight_nodes(ggml_cgraph * cgraph,
+                                                                                bool naive = false);
+
+    // Collect just the set of weight-tensor names referenced by the graph, without
+    // building (or requantizing) any OV weight nodes. Used by topology checks like
+    // is_model_splitted that only need name membership.
+    static std::set<std::string> collect_weight_names(ggml_cgraph * cgraph);
+
+    const ggml_tensor * get_tensor_used_op(const ggml_tensor * tensor) const;
+
+    const ggml_tensor * get_tensor_from_name(const std::string & name) const;
+
+    void clear_model_weights() { m_model_weights.clear(); }
+
+    static std::pair<ModelParams, ComputeParams> compute_llm_params(ggml_cgraph * cgraph, bool is_static);
+
+    ModelParams get_model_params() const { return m_model_params; }
+
+    ComputeParams get_compute_params() const { return m_compute_params; }
+
+    void set_model_params(const ModelParams & model_params) { m_model_params = model_params; }
+
+    void set_compute_params(const ComputeParams & compute_params) { m_compute_params = compute_params; }
+
+    bool m_is_static = false;
+    bool m_is_stateful = false;
+    bool m_is_prefill = false;
+    bool m_naive = false;
+    int m_prefill_chunk_size = 0;
+    bool m_model_is_splitted = false;  // label the cgraph is splited or not
+
+    static ov::Shape get_shape(const ggml_tensor * tensor);
+    static std::vector<size_t> get_stride(const ggml_tensor * tensor);
+    static ov::element::Type get_ov_type(const ggml_tensor * tensor);
+    static std::string compute_op_type(const ggml_tensor * node);
+    void add_extra_inputs();
+
+    void update_io(ggml_cgraph * cgraph);
+    bool is_bound_to(const ggml_cgraph * cgraph) const;
+
+    static bool is_inp_tok(const ggml_tensor * tensor, const ggml_tensor * op) {
+        return op->op == GGML_OP_GET_ROWS && tensor == op->src[1] && op->src[0]->op == GGML_OP_NONE;
+    }
+
+    static bool is_inp_pos(const ggml_tensor * tensor, const ggml_tensor * op) {
+        return op->op == GGML_OP_ROPE && tensor == op->src[1];
+    }
+
+    // IMROPE and VISION pack 4 stacked position planes (t/h/w/e) into inp_pos, each of length
+    // n_tokens; other modes carry a single position per token.
+    static int get_inp_pos_n_planes(const ggml_tensor * op) {
+        const int mode = op->op_params[2];
+        return (mode == GGML_ROPE_TYPE_IMROPE || mode == GGML_ROPE_TYPE_VISION || (mode & GGML_ROPE_TYPE_MROPE)) ? 4 : 1;
+    }
+
+    static bool is_inp_emb(const ggml_tensor * tensor, const ggml_tensor * op) {
+        return tensor->op == GGML_OP_GET_ROWS && op->op == GGML_OP_RMS_NORM;
+    }
+
+    static bool is_inp_mask(const ggml_tensor * tensor, const ggml_tensor * op) {
+        return op->op == GGML_OP_CPY || (op->op == GGML_OP_FLASH_ATTN_EXT && tensor == op->src[3]) ||
+               (op->op == GGML_OP_SOFT_MAX && tensor == op->src[1]);
+    }
+
+    static bool is_inp_mean(const ggml_tensor * tensor, const ggml_tensor * op) {
+        return op->op == GGML_OP_MUL_MAT && tensor == op->src[1] && tensor->op == GGML_OP_NONE &&
+               (tensor->flags & GGML_TENSOR_FLAG_INPUT) && tensor->type == GGML_TYPE_F32 &&
+               op->src[0] != nullptr && op->src[0]->op != GGML_OP_NONE;
+    }
+
+    // per-token embedding scale [1, n_tokens] (inp->scale_rows in llama-graph.cpp)
+    static bool is_inp_scale_rows(const ggml_tensor * tensor, const ggml_tensor * op) {
+        return op->op == GGML_OP_MUL && tensor == op->src[1] && strcmp(tensor->name, "inp_scale_rows") == 0;
+    }
+
+    static bool is_rope_freqs_weight(const ggml_tensor * tensor, const ggml_tensor * op) {
+        return op->op == GGML_OP_ROPE && tensor == op->src[2];
+    }
+
+    inline static bool is_recurrent_cache(const ggml_tensor * tensor) {
+        return tensor != nullptr && (strncmp(tensor->name, "cache_r_l", strlen("cache_r_l")) == 0 ||
+                                     strncmp(tensor->name, "cache_s_l", strlen("cache_s_l")) == 0 ||
+                                     strncmp(tensor->name, "cache_ple_r_l", strlen("cache_ple_r_l")) == 0);
+    }
+
+    inline static bool is_cache(const ggml_tensor * tensor, const ggml_tensor * op) {
+        return is_recurrent_cache(tensor) || is_kvcache(tensor, op);
+    }
+
+    inline static bool is_kvcache(const ggml_tensor * tensor, const ggml_tensor * op) {
+        if (tensor == nullptr || is_recurrent_cache(tensor)) {
+            return false;
+        }
+        return (tensor->buffer != nullptr && tensor->buffer->usage == GGML_BACKEND_BUFFER_USAGE_ANY) ||
+               (op != nullptr && op->op == GGML_OP_SET_ROWS && op->src[2] == tensor);
+    }
+
+    inline static bool is_conv_state_writeback(const ggml_tensor * node) {
+        return node->op == GGML_OP_CPY && node->view_src != nullptr && is_recurrent_cache(node->view_src) &&
+               node->src[0] != nullptr && node->src[0]->op == GGML_OP_VIEW && node->src[0]->src[0] != nullptr &&
+               node->src[0]->src[0]->op == GGML_OP_CONCAT && node->src[1] != nullptr &&
+               node->src[1]->op == GGML_OP_VIEW && node->src[1]->view_src == node->view_src;
+    }
+
+    static bool is_kv_idx(const ggml_tensor * tensor, const ggml_tensor * op) {
+        return op->op == GGML_OP_SET_ROWS && op->src[1] == tensor;
+    }
+
+    bool is_swa_mask(const ggml_tensor * tensor) const {
+        return m_model_params.swa_mask != nullptr && tensor == m_model_params.swa_mask;
+    }
+
+    static bool is_output_idx(const ggml_tensor * tensor, const ggml_tensor * op) {
+        return op->op == GGML_OP_GET_ROWS && tensor == op->src[1] && op->src[0]->op != GGML_OP_NONE &&
+               op->src[1]->op == GGML_OP_NONE;
+    }
+
+    // the state permutation index input used in SSM/DeltaNet models (inp->s_copy in llama-graph.cpp)
+    static bool is_inp_s_copy(const ggml_tensor * tensor, const ggml_tensor * op) {
+        return op->op == GGML_OP_GET_ROWS && tensor == op->src[1] &&
+               op->src[0]->buffer->usage == GGML_BACKEND_BUFFER_USAGE_ANY;
+    }
+
+    std::string get_graph_input_ov_name(const ggml_tensor * tensor, const ggml_tensor * op) const {
+        if (is_inp_pos(tensor, op)) {
+            return "inp_pos";
+        }
+        if (is_inp_emb(tensor, op)) {
+            return "embd";
+        }
+        if (is_inp_mask(tensor, op)) {
+            // Give the two attention masks distinct OV parameter names.
+            //
+            // An interleaved-SWA model builds one full-attention mask and one sliding-window mask,
+            // but build_attn_inp_kq_mask() names them identically, so keying a parameter off
+            // tensor->name alone makes the second mask OVERWRITE the first in m_model_inputs: both
+            // attention types then read a single parameter, and the windowed layers silently run
+            // against an unbanded mask. Disambiguate using the SWA layer set computed in
+            // compute_llm_params(), which classifies by mask tensor identity rather than by name.
+            //
+            // When no SWA layer was found there is only one mask in play, so the plain name is
+            // correct and no _swa parameter is created.
+            if (m_model_params.swa_layers.empty()) {
+                return "self_kq_mask";
+            }
+            return is_swa_mask(tensor) ? "self_kq_mask_swa" : "self_kq_mask";
+        }
+        return tensor->name;
+    }
+
+private:
+    void set_input_output();
+    int compute_op_case(const ggml_tensor * node) const;
+    bool node_is_used_as_src(const int node_idx);
+    void compute_model_inputs();
+    void compute_model_outputs();
+
+    // True if tensor is the inp->s_copy index leaf gathered by a recurrent state cache GET_ROWS
+    // (possibly through a VIEW), so it gets a dynamic [1,1,1,-1] graph-input shape.
+    bool is_s_copy_leaf(const ggml_tensor * tensor) const;
+
+    // Infer and propagate dynamic-dimension indices for all tensors in the GGML graph.
+    void compute_node_dynamic_dims();
+
+    void validate_cgraph() const;
+
+    ggml_cgraph * m_cgraph = nullptr;
+    std::map<std::string, ggml_tensor *> m_inputs;
+
+    std::map<std::string, ov::frontend::ggml::ModelInputInfo> m_model_inputs;
+    std::map<std::string, ov::frontend::ggml::ModelExtraInputInfo> m_model_extra_inputs;
+    std::map<std::string, std::shared_ptr<ov::Node>> m_model_weights;
+    std::map<std::string, ggml_tensor *> m_model_outputs;
+    std::set<std::string> m_model_output_names;
+    std::vector<NodeInfo> m_node_info_list;
+    std::map<ggml_tensor *, int> m_node_dynamic_dims;
+
+    ModelParams m_model_params;
+    ComputeParams m_compute_params;
+};
+
+void print_tensor_address_map(const ggml_cgraph * cgraph);

llama.cpp/ggml/src/ggml-openvino/ggml-openvino-extra.cpp

@@ -0,0 +1,669 @@
+#include "ggml-openvino-extra.h"
+
+#include "ggml-impl.h"
+#include "ggml.h"
+#include "model-cache.h"
+
+#include <algorithm>
+#include <cstdlib>
+#include <cstring>
+#include <openvino/runtime/intel_gpu/ocl/ocl.hpp>
+#include <openvino/runtime/intel_npu/level_zero/level_zero.hpp>
+#include <openvino/runtime/properties.hpp>
+#include <mutex>
+#include <optional>
+
+ov::Core & ov_singleton_core() {
+    static ov::Core core;
+    return core;
+}
+
+static bool has_prefix(const std::string & s, const std::string & prefix) {
+    return s.size() >= prefix.size() && std::equal(prefix.begin(), prefix.end(), s.begin());
+}
+
+static bool is_virtual_routing_device(const std::string & device_name) {
+    return has_prefix(device_name, "AUTO") || has_prefix(device_name, "MULTI") || has_prefix(device_name, "HETERO");
+}
+
+static std::vector<std::string> ov_enumerate_devices() {
+    std::vector<std::string> result;
+
+    for (const auto & device : ov_singleton_core().get_available_devices()) {
+        if (!is_virtual_routing_device(device)) {
+            result.push_back(device);
+        }
+    }
+
+    if (result.empty()) {
+        result.push_back("CPU");
+    }
+
+    std::sort(result.begin(), result.end());
+    result.erase(std::unique(result.begin(), result.end()), result.end());
+    return result;
+}
+
+std::string ggml_openvino_get_device_description(const std::string & device_name) {
+    std::string description = device_name;
+    try {
+        description = ov_singleton_core().get_property(device_name, ov::device::full_name);
+    } catch (...) {
+        return device_name;
+    }
+
+    if (has_prefix(device_name, "NPU")) {
+        try {
+            const std::string arch = ov_singleton_core().get_property(device_name, "DEVICE_ARCHITECTURE").as<std::string>();
+            if (!arch.empty()) {
+                description += " (NPU " + arch + ")";
+            }
+        } catch (...) {
+        }
+    }
+
+    return description;
+}
+
+// requested: GGML_OPENVINO_DEVICE, nullptr if unset. available_devices is never empty (see ov_enumerate_devices)
+static std::string resolve_openvino_device_name(const std::vector<std::string> & available_devices,
+                                                const char * requested) {
+    auto available = [&](const std::string & name) {
+        return std::find(available_devices.begin(), available_devices.end(), name) != available_devices.end();
+    };
+    if (requested == nullptr) {
+        return available("CPU") ? "CPU" : available_devices.front();
+    }
+    if (!available(requested)) {
+        // No fallback to CPU (easy to miss) and no GPU -> GPU.0 alias (with iGPU + dGPU, GPU.0 is often the
+        // wrong one). List the devices here: --list-devices initializes this backend and would abort too.
+        std::string list;
+        for (const std::string & name : available_devices) {
+            list += "\n  " + name + ": " + ggml_openvino_get_device_description(name);
+        }
+        GGML_ABORT("GGML OpenVINO Backend: GGML_OPENVINO_DEVICE=%s is not available. "
+                   "Set it to one of the available OpenVINO devices:%s",
+                   requested, list.c_str());
+    }
+    return requested;
+}
+
+// =====================================================
+// Device Configuration Implementations
+// =====================================================
+
+void ggml_openvino_device_config::init() {
+    static std::mutex mutex;
+    std::lock_guard<std::mutex> lock(mutex);
+    if (initialized) {
+        return;
+    }
+    // Set up front: a failed OpenCL setup below is not retried on every call
+    initialized = true;
+
+    // All recognized GGML_OPENVINO_* env vars. Their values are cached here
+    // once at backend init time and read back via ggml_openvino_getenv_str()
+    // (raw string) or ggml_openvino_getenv_int() (integer / boolean toggle).
+    static constexpr const char * env_var_names[] = {
+        // String values (use ggml_openvino_getenv_str)
+        "GGML_OPENVINO_DEVICE",
+        "GGML_OPENVINO_CACHE_DIR",
+        "GGML_OPENVINO_SPILL_DIR",
+        "GGML_OPENVINO_DEBUG_NODE",
+        "GGML_OPENVINO_COMPILED_MODEL_CACHE_DIR",
+        "GGML_OPENVINO_COMPILED_MODEL_CACHE_ONLY",
+        "GGML_OPENVINO_NPU_COMPILE_CONFIG",
+        // Integer values (use ggml_openvino_getenv_int)
+        "GGML_OPENVINO_PREFILL_CHUNK_SIZE",
+        // Boolean toggles (treated as int flags via ggml_openvino_getenv_int)
+        "GGML_OPENVINO_STATEFUL_EXECUTION",
+        "GGML_OPENVINO_PROFILING",
+        "GGML_OPENVINO_DUMP_CGRAPH",
+        "GGML_OPENVINO_DUMP_IR",
+        "GGML_OPENVINO_DEBUG_INPUT",
+        "GGML_OPENVINO_DEBUG_OUTPUT",
+        // Force the static (NPU-shape) compute path on any device, e.g. GGML_OPENVINO_DEVICE=CPU,
+        // to test the static-shape translation without NPUW/real NPU hardware in the loop.
+        "GGML_OPENVINO_FORCE_STATIC",
+        "GGML_OPENVINO_PRINT_CGRAPH_TENSOR_ADDRESS",
+        "GGML_OPENVINO_ENABLE_CACHE",
+        "GGML_OPENVINO_DISABLE_CACHE",
+        "GGML_OPENVINO_DISABLE_KV_SLICE",
+        "GGML_OPENVINO_ENABLE_FALLBACK",
+        "GGML_OPENVINO_MANUAL_GQA_ATTN",
+        "GGML_OPENVINO_DISABLE_ELTWISE_RANK_ALIGN",
+        "GGML_OPENVINO_MOE_OP",
+        "GGML_OPENVINO_MEMORY_OPTIMIZE",
+        "GGML_OPENVINO_RELEASE_WEIGHTS",
+        "GGML_OPENVINO_REDUCE_COMPILE_MEM",
+        "GGML_OPENVINO_LOG_UNSUPPORTED_OPS",
+        "GGML_OPENVINO_LOG_SWA_LAYERS",
+        "GGML_OPENVINO_NATIVE_SOFTPLUS",
+        "GGML_OPENVINO_DISABLE_REMOTE_OUTPUTS",
+        "GGML_OPENVINO_REQUANT_KQUANT",
+        "GGML_OPENVINO_DISABLE_KV_STATE_RELAYOUT",
+        // Build the precise (but O(n_nodes)) graph cache key. Needed by op tests.
+        "GGML_OPENVINO_FULL_GRAPH_KEY",
+    };
+
+    for (const char * const & env_var : env_var_names) {
+        auto * env = getenv(env_var);
+        if (env) {
+            environment_variables[env_var] = env;
+        }
+    }
+
+    available_devices = ov_enumerate_devices();
+    device_name = resolve_openvino_device_name(available_devices, ggml_openvino_getenv_str("GGML_OPENVINO_DEVICE"));
+    is_npu = has_prefix(device_name, "NPU");
+
+    ggml_openvino_model_cache_init();
+
+    const char * cache_dir = ggml_openvino_getenv_str("GGML_OPENVINO_CACHE_DIR");
+    if (has_prefix(device_name, "NPU")) {
+        compile_config = {
+            {"NPU_COMPILER_DYNAMIC_QUANTIZATION", "YES"   },
+            {"NPU_USE_NPUW",                      "YES"   },
+            {"NPUW_DEVICES",                      "NPU"   },
+            {"NPUW_FOLD",                         "YES"   },
+            {"NPUW_WEIGHTS_BANK",                 "shared"},
+            {"NPUW_FUNCALL_FOR_ALL",              "YES"   },
+            {"NPUW_FUNCALL_ASYNC",                "YES"   },
+            {"NPUW_DQ",                           "YES"   },
+            {"NPUW_DQ_FULL",                      "NO"    },
+        };
+        if (cache_dir && strlen(cache_dir) > 0) {
+            compile_config["NPUW_CACHE_DIR"] = cache_dir;
+            compile_config.insert(ov::cache_mode(ov::CacheMode::OPTIMIZE_SIZE));
+        }
+        const char * compilation_mode_params =
+            ggml_openvino_getenv_str("GGML_OPENVINO_NPU_COMPILE_CONFIG");
+        if (compilation_mode_params && strlen(compilation_mode_params) > 0) {
+            compile_config["NPU_COMPILATION_MODE_PARAMS"] = compilation_mode_params;
+        }
+    } else if (cache_dir && strlen(cache_dir) > 0) {
+        compile_config.insert(ov::cache_dir(cache_dir));
+        compile_config.insert(ov::cache_mode(ov::CacheMode::OPTIMIZE_SIZE));
+    }
+
+    if (ggml_openvino_getenv_int("GGML_OPENVINO_PROFILING") >= 2) {
+        compile_config.insert(ov::enable_profiling(true));
+    }
+
+    // Initialize remote context with queue sharing for GPU
+    if (has_prefix(device_name, "GPU")) {
+        // Use the OpenCL context OpenVINO created for this device, so GPU.N gets its own device
+        cl_context cl_ctx;
+        try {
+            auto ov_ctx = ov_singleton_core().get_default_context(device_name).as<ov::intel_gpu::ocl::ClContext>();
+            cl_ctx = ov_ctx.get();
+        } catch (const std::exception & e) {
+            // The consumers of the remote context have no host fallback, and OpenVINO
+            // already reported the device as present.
+            GGML_ABORT("ggml-openvino: failed to get the OpenCL context for %s: %s", device_name.c_str(), e.what());
+        }
+
+        cl_int err;
+        cl_device_id cl_device;
+        err = clGetContextInfo(cl_ctx, CL_CONTEXT_DEVICES, sizeof(cl_device), &cl_device, nullptr);
+        if (err != CL_SUCCESS) {
+            GGML_ABORT("ggml-openvino: failed to get the OpenCL device for %s: %d", device_name.c_str(), err);
+        }
+
+        cl_platform_id cl_platform;
+        err = clGetDeviceInfo(cl_device, CL_DEVICE_PLATFORM, sizeof(cl_platform), &cl_platform, nullptr);
+        if (err != CL_SUCCESS) {
+            GGML_ABORT("ggml-openvino: failed to get the OpenCL platform for %s: %d", device_name.c_str(), err);
+        }
+
+        cl_mem_fill_fn =
+            (clEnqueueMemFillINTEL_fn) clGetExtensionFunctionAddressForPlatform(cl_platform, "clEnqueueMemFillINTEL");
+        cl_mem_cpy_fn =
+            (clEnqueueMemcpyINTEL_fn) clGetExtensionFunctionAddressForPlatform(cl_platform, "clEnqueueMemcpyINTEL");
+
+        cl_ulong device_max_alloc = 0;
+        err = clGetDeviceInfo(cl_device, CL_DEVICE_MAX_MEM_ALLOC_SIZE, sizeof(device_max_alloc), &device_max_alloc,
+                              nullptr);
+        if (err == CL_SUCCESS) {
+            max_alloc_size = device_max_alloc;
+        } else {
+            // not fatal, ggml then allocates one buffer
+            GGML_LOG_WARN("Failed to get OpenCL max allocation size: %d\n", err);
+        }
+
+        const cl_queue_properties profiling_properties[] = {
+            CL_QUEUE_PROPERTIES,
+            CL_QUEUE_PROFILING_ENABLE,
+            0,
+        };
+        const cl_queue_properties * queue_properties =
+            ggml_openvino_getenv_int("GGML_OPENVINO_PROFILING") >= 2 ? profiling_properties : nullptr;
+        cl_queue = clCreateCommandQueueWithProperties(cl_ctx, cl_device, queue_properties, &err);
+        if (err != CL_SUCCESS) {
+            GGML_ABORT("ggml-openvino: failed to create the OpenCL queue for %s: %d", device_name.c_str(), err);
+        }
+
+        // Create OpenVINO remote context with queue sharing
+        remote_context = ov::intel_gpu::ocl::ClContext(ov_singleton_core(), cl_queue);
+    } else if (has_prefix(device_name, "NPU")) {
+        // remote tensor is not used for NPU yet
+        // remote_context = ov_singleton_core().get_default_context(device_name);
+    }
+}
+
+ggml_openvino_device_config::~ggml_openvino_device_config() {
+    if (cl_queue != nullptr) {
+        clReleaseCommandQueue(cl_queue);
+        cl_queue = nullptr;
+    }
+}
+
+// Get the global device config singleton
+ggml_openvino_device_config & ggml_openvino_get_device_config() {
+    static ggml_openvino_device_config config;
+    return config;
+}
+
+// Initialize device config (call during backend init)
+void ggml_openvino_init_device_config() {
+    ggml_openvino_get_device_config().init();
+}
+
+// Get the device name
+const std::string & ggml_openvino_get_device_name() {
+    return ggml_openvino_get_device_config().device_name;
+}
+
+std::vector<std::string> ggml_openvino_get_available_devices() {
+    auto & config = ggml_openvino_get_device_config();
+    config.init();
+    return config.available_devices;
+}
+
+// Get the value of a GGML_OPENVINO_* env var as a string. Returns
+// default_value when the var is unset or set to an empty string.
+const char * ggml_openvino_getenv_str(const char * var, const char * default_value) {
+    auto & env_map = ggml_openvino_get_device_config().environment_variables;
+    auto it = env_map.find(var);
+    return (it == env_map.end() || it->second.empty()) ? default_value : it->second.c_str();
+}
+
+// Get the value of a GGML_OPENVINO_* env var as an int (via std::atoi).
+// Returns default_value (0) when the var is unset or empty. Used for both
+// integer settings (e.g. GGML_OPENVINO_PREFILL_CHUNK_SIZE) and boolean
+// toggles: "0" disables, any non-zero integer enables.
+int ggml_openvino_getenv_int(const char * var, int default_value) {
+    const char * v = ggml_openvino_getenv_str(var, nullptr);
+    return v ? std::atoi(v) : default_value;
+}
+
+bool ggml_openvino_reduce_compile_mem_enabled() {
+    const char * reduce_compile_mem = ggml_openvino_getenv_str("GGML_OPENVINO_REDUCE_COMPILE_MEM");
+    if (reduce_compile_mem != nullptr) {
+        return ggml_openvino_getenv_int("GGML_OPENVINO_REDUCE_COMPILE_MEM") != 0;
+    }
+    return ggml_openvino_getenv_int("GGML_OPENVINO_MEMORY_OPTIMIZE") != 0;
+}
+
+bool ggml_openvino_release_weights_enabled() {
+    const char * release_weights = ggml_openvino_getenv_str("GGML_OPENVINO_RELEASE_WEIGHTS");
+    if (release_weights != nullptr) {
+        return ggml_openvino_is_gpu() && ggml_openvino_getenv_int("GGML_OPENVINO_RELEASE_WEIGHTS") != 0;
+    }
+    return ggml_openvino_is_gpu() && ggml_openvino_getenv_int("GGML_OPENVINO_MEMORY_OPTIMIZE") != 0;
+}
+
+// Check if running on NPU
+bool ggml_openvino_is_npu() {
+    return ggml_openvino_get_device_config().is_npu;
+}
+
+bool ggml_openvino_is_gpu() {
+    return has_prefix(ggml_openvino_get_device_name(), "GPU");
+}
+
+size_t ggml_openvino_max_alloc_size() {
+    return ggml_openvino_get_device_config().max_alloc_size;
+}
+
+// Get the remote context for the current device (returns empty optional for CPU)
+std::optional<ov::RemoteContext> ggml_openvino_get_remote_context() {
+    return ggml_openvino_get_device_config().remote_context;
+}
+
+// Get the compile config for the current device
+const ov::AnyMap & ggml_openvino_get_compile_config() {
+    return ggml_openvino_get_device_config().compile_config;
+}
+
+// Get the OpenCL command queue for GPU operations
+cl_command_queue ggml_openvino_get_cl_queue() {
+    return ggml_openvino_get_device_config().cl_queue;
+}
+
+// Get the clEnqueueMemFillINTEL function pointer
+clEnqueueMemFillINTEL_fn ggml_openvino_get_clEnqueueMemFillINTEL() {
+    return ggml_openvino_get_device_config().cl_mem_fill_fn;
+}
+
+// Get the clEnqueueMemcpyINTEL function pointer
+clEnqueueMemcpyINTEL_fn ggml_openvino_get_clEnqueueMemcpyINTEL() {
+    return ggml_openvino_get_device_config().cl_mem_cpy_fn;
+}
+
+// Get requantization type for a tensor type (returns nullopt if no requant needed)
+std::optional<ExtraQuantType> ggml_openvino_get_requant_type(const ggml_tensor * tensor, bool no_requant) {
+    if (no_requant) {
+        return std::nullopt;
+    }
+    if (strncmp(tensor->name, "token_embd.weight", 17) == 0) {
+        return ((ggml_openvino_is_npu() && tensor->type == GGML_TYPE_Q6_K) ? ExtraQuantType::F16 :
+                                                                             ExtraQuantType::Q8_0_C);
+    }
+    if (strncmp(tensor->name, "output.weight", 13) == 0) {
+        return ExtraQuantType::Q8_0_C;
+    }
+    if (ggml_openvino_is_npu()) {
+        return ExtraQuantType::Q4_0_128;
+    }
+    // By default Q6_K/Q5_K are requantized to Q8_0_C, which *inflates* 6- and 5-bit weights to 8
+    // while the rest of the model stays at 4 bits, and Q4_K keeps its native group-32 layout
+    // (an f16 scale plus an f16 zero point per 32 weights = 0.125 B/weight of metadata).
+    // Decode of a large model is bandwidth-bound, so both cost throughput.
+    //
+    // GGML_OPENVINO_REQUANT_KQUANT selects a 4-bit target instead. Names are
+    // q4_<sym|asym><group>[_all]: <sym|asym> says whether a per-group zero point is kept, <group>
+    // is the group size, and the _all suffix sends Q4_K down the same path (without it only
+    // Q6_K/Q5_K are touched):
+    //   q4_sym128      Q6_K/Q5_K -> Q4_0_128 (u4, group 128, symmetric)
+    //   q4_sym128_all  and Q4_K too -- drops Q4_K's per-32 zero point, which costs some accuracy
+    //   q4_asym64      Q6_K/Q5_K -> Q4_1_64 (u4, group 64, asymmetric)
+    //   q4_asym64_all  Q6_K/Q5_K and Q4_K -> Q4_1_64 (u4, group 64, asymmetric)
+    //   native         no requantization at all (keep Q6_K/Q5_K as they are)
+    //
+    // q4_asym64 leaves Q4_K at its native group 32. Use q4_asym64_all to keep the group size uniform.
+    const char * rq = ggml_openvino_getenv_str("GGML_OPENVINO_REQUANT_KQUANT");
+    auto is_opt = [rq](const char * name) {
+        return rq && strcmp(rq, name) == 0;
+    };
+    const bool sym128 = is_opt("q4_sym128");
+    const bool sym128_all = is_opt("q4_sym128_all");
+    const bool asym64_all = is_opt("q4_asym64_all");
+    const bool asym64 = is_opt("q4_asym64");
+
+    if (tensor->type == GGML_TYPE_Q4_K) {
+        if (sym128_all) {
+            return ExtraQuantType::Q4_0_128;
+        }
+        if (asym64_all) {
+            return ExtraQuantType::Q4_1_64;
+        }
+    }
+    // MoE expert weights (3D, ne[2] = n_expert) stored as Q5_1/Q8_0 are the expert-side
+    // equivalent of Q6_K/Q5_K: kept at 8 bits by default while the rest of the model is at 4
+    // (gemma-4 26B-A4B keeps its down projection there). Send them to 4 bits under the same
+    // option, at group 64 rather than 128: the down expert has k=704, which 64 divides
+    // (704/64 = 11) and 128 does not.
+    if (tensor->ne[2] > 1 && (tensor->type == GGML_TYPE_Q5_1 || tensor->type == GGML_TYPE_Q8_0)) {
+        if (sym128 || sym128_all) {
+            return ExtraQuantType::Q4_0_64;
+        }
+        if (asym64 || asym64_all) {
+            return ExtraQuantType::Q4_1_64;
+        }
+        // TODO: temporary workaround for a known OpenVINO GPU-plugin bug -- remove once the
+        // plugin computes grouped 8-bit GatherMatmulCompressed correctly. This costs accuracy
+        // (5/8-bit -> 4-bit) on any model it applies to, so it must not outlive the bug.
+        //
+        // On GPU these would otherwise stay in their native *grouped 8-bit* layout, which the GPU
+        // plugin's GatherMatmulCompressed computes incorrectly -- gemma-4 26B-A4B (whose down
+        // projection is Q5_1) produces garbage, while the same graph is correct on CPU. It is
+        // specific to grouped 8 bit: the gate/up experts are grouped u4 *with* a zero point and
+        // are fine, and Qwen3.5 / granite are fine because their Q5_K/Q6_K down projections
+        // already requantize to per-channel Q8_0_C (grouped=0). Sending these to grouped 4 bit
+        // avoids the broken layout and restores correct output.
+        // Opt out with GGML_OPENVINO_REQUANT_KQUANT=native.
+        if (ggml_openvino_is_gpu() && !is_opt("native")) {
+            return ExtraQuantType::Q4_0_64;
+        }
+    }
+    switch (tensor->type) {
+    case GGML_TYPE_Q6_K:
+    case GGML_TYPE_Q5_K:
+        if (sym128 || sym128_all) {
+            return ExtraQuantType::Q4_0_128;
+        }
+        if (asym64 || asym64_all) {
+            return ExtraQuantType::Q4_1_64;
+        }
+        if (is_opt("native")) {
+            return std::nullopt;
+        }
+        return ExtraQuantType::Q8_0_C;
+    default:
+        return std::nullopt;
+    }
+}
+
+// =====================================================
+// Extracted Layout Calculation
+// =====================================================
+
+ggml_openvino_extracted_layout ggml_openvino_get_extracted_layout(const ggml_tensor * tensor, bool use_bias) {
+    ggml_openvino_extracted_layout layout = {};
+    layout.is_symmetric = false;
+
+    if (!ggml_is_quantized(tensor->type)) {
+        return layout;
+    }
+
+    // Most quantized weights use the existing 2D extraction path. 3D expert weights for
+    // MUL_MAT_ID (MoE) are also supported, either as MXFP4 (packed, dedicated branch below) or via the
+    // generic sizing math below, which is shape-agnostic (based on total element count). Only reject 4D.
+    if (tensor->ne[3] != 1) {
+        return layout;
+    }
+
+    // 3D MoE expert weights that are not requantized (see below) always use the exact f16
+    // zero-point extraction (see extract_quantized_weights), which needs a wider zp slot than
+    // the packed integer zero point -- must be kept in sync with that function so the buffer
+    // sizing here matches what process_weight_tensor actually writes.
+    const bool for_gather_matmul = tensor->ne[2] > 1;
+
+    int64_t n_elements = ggml_nelements(tensor);
+    const size_t alignment = 64;  // Good for SIMD
+
+    if (tensor->type == GGML_TYPE_MXFP4 && (tensor->ne[2] > 1 || tensor->ne[3] > 1)) {
+        layout.weights_per_block = 32;
+        layout.is_symmetric = true;
+        layout.weights_size = ggml_nbytes(tensor);
+        layout.weights_offset = 0;
+        layout.total_size = layout.weights_size;
+        return layout;
+    }
+
+    // Check if requantization is needed (NPU-specific)
+    auto requant_type = ggml_openvino_get_requant_type(tensor, use_bias);
+    if (requant_type.has_value()) {
+        layout.is_requant = true;
+        layout.requant_type = requant_type;
+
+        // Special case: requant to F16 - just store F16 weights, no scales/zp
+        if (requant_type.value() == ExtraQuantType::F16) {
+            layout.weights_size = n_elements * sizeof(uint16_t);  // F16 = 2 bytes
+            layout.total_size = layout.weights_size;
+            layout.weights_offset = 0;
+            // No scales/zp for F16
+            return layout;
+        }
+
+        // Requant to different quantized format (e.g., Q4_0_128)
+        switch (requant_type.value()) {
+        case ExtraQuantType::Q4_0_128:
+            layout.is_u4 = true;
+            layout.weights_per_block = 128;
+            layout.is_symmetric = true;
+            break;
+        case ExtraQuantType::Q4_1_64:
+            layout.is_u4 = true;
+            layout.weights_per_block = 64;
+            layout.is_symmetric = false;
+            break;
+        case ExtraQuantType::Q4_0_64:
+            layout.is_u4 = true;
+            layout.weights_per_block = 64;
+            layout.is_symmetric = true;
+            break;
+        case ExtraQuantType::Q4_0_C:
+            layout.is_u4 = true;
+            layout.weights_per_block = tensor->ne[0];
+            layout.is_symmetric = true;
+            break;
+        case ExtraQuantType::Q8_0_32:
+            layout.is_u4 = false;
+            layout.weights_per_block = 32;
+            layout.is_symmetric = true;
+            break;
+        case ExtraQuantType::Q8_0_C:
+            layout.is_u4 = false;
+            layout.weights_per_block = tensor->ne[0];
+            layout.is_symmetric = true;
+            break;
+        case ExtraQuantType::Q8_1_C:
+            layout.is_u4 = false;
+            layout.weights_per_block = tensor->ne[0];
+            break;
+        default:
+            GGML_ABORT("Code of re-quantizing to channel-wise is not updated");
+        }
+
+        if (layout.is_requant) {
+            // Calculate sizes for requantized format
+            layout.weights_size = layout.is_u4 ? (n_elements / 2) : n_elements;
+            int64_t n_blocks = n_elements / layout.weights_per_block;
+            layout.scales_size = n_blocks * sizeof(uint16_t);
+            // For symmetric quantization, no zp needed (weights stored as signed)
+            if (layout.is_symmetric) {
+                layout.zp_size = 0;
+            } else {
+                layout.zp_size = layout.is_u4 ? ((n_blocks + 1) / 2) : n_blocks;
+            }
+
+            layout.weights_offset = 0;
+            layout.scales_offset = ((layout.weights_size + alignment - 1) / alignment) * alignment;
+            layout.zp_offset = layout.scales_offset + ((layout.scales_size + alignment - 1) / alignment) * alignment;
+            layout.total_size = layout.zp_offset + layout.zp_size;
+            layout.total_size = std::max(layout.total_size, ggml_nbytes(tensor));
+            return layout;
+        }
+    }
+
+    // Normal extraction (no requant) - determine format based on tensor type
+    layout.is_u4 = false;
+    layout.weights_per_block = 32;
+    layout.is_symmetric = false;
+
+    switch (tensor->type) {
+    case GGML_TYPE_MXFP4:
+    case GGML_TYPE_Q4_0:
+        layout.is_u4 = true;
+        layout.is_symmetric = true;
+        break;
+
+    case GGML_TYPE_Q4_1:
+    case GGML_TYPE_Q4_K:
+        layout.is_u4 = true;
+        break;
+
+    case GGML_TYPE_Q8_0:
+        layout.is_symmetric = true;
+        break;
+
+    case GGML_TYPE_Q5_1:
+        // u8 weights (5-bit values), asymmetric (scale + zero point)
+        break;
+
+    case GGML_TYPE_Q6_K:
+        layout.weights_per_block = 16;
+        layout.is_symmetric = true;
+        break;
+
+    case GGML_TYPE_Q5_K:
+        break;
+
+    default:
+        // Unsupported quantization type
+        return layout;
+    }
+
+    // Calculate sizes
+    // Weights: U4 = n_elements/2 bytes, U8 = n_elements bytes
+    layout.weights_size = layout.is_u4 ? (n_elements / 2) : n_elements;
+
+    // Scales: F16 per block, except MXFP4 which stores one E8M0 byte per block.
+    int64_t n_blocks = n_elements / layout.weights_per_block;
+    layout.scales_size = n_blocks * (tensor->type == GGML_TYPE_MXFP4 ? sizeof(uint8_t) : sizeof(uint16_t));
+    // For symmetric quantization, no zp needed (weights stored as signed). Asymmetric
+    // for_gather_matmul (3D MoE expert) weights use an exact f16 zero point (see
+    // extract_quantized_weights/make_int8_weights/make_int4_weights), which needs one f16 per
+    // block instead of a packed u4/u8 integer zero point.
+    if (layout.is_symmetric) {
+        layout.zp_size = 0;
+    } else if (use_bias || for_gather_matmul) {
+        layout.zp_size = n_blocks * sizeof(uint16_t);
+    } else {
+        layout.zp_size = layout.is_u4 ? ((n_blocks + 1) / 2) : n_blocks;
+    }
+
+    // Layout in buffer: [weights | scales | zp] with alignment
+    layout.weights_offset = 0;
+    layout.scales_offset = ((layout.weights_size + alignment - 1) / alignment) * alignment;
+    layout.zp_offset = layout.scales_offset + ((layout.scales_size + alignment - 1) / alignment) * alignment;
+    layout.total_size = layout.zp_offset + layout.zp_size;
+    layout.total_size = std::max(layout.total_size, ggml_nbytes(tensor));
+
+    return layout;
+}
+
+ggml_openvino_tensor_extra * ggml_openvino_create_tensor_extra(const ggml_tensor * tensor, bool is_remote) {
+    ov::Shape shape;
+    for (int i = GGML_MAX_DIMS - 1; i >= 0; --i) {
+        shape.push_back(static_cast<size_t>(tensor->ne[i]));
+    }
+
+    ov::element::Type element_type;
+    switch (tensor->type) {
+    case GGML_TYPE_F32:
+        element_type = ov::element::f32;
+        break;
+    case GGML_TYPE_F16:
+        element_type = ov::element::f16;
+        break;
+    case GGML_TYPE_BF16:
+        element_type = ov::element::bf16;
+        break;
+    case GGML_TYPE_I32:
+        element_type = ov::element::i32;
+        break;
+    case GGML_TYPE_I64:
+        element_type = ov::element::i64;
+        break;
+    default:
+        // GGML_LOG_WARN("%s: unsupported tensor type for ov::Tensor: %s\n", __func__, ggml_type_name(tensor->type));
+        return nullptr;
+    }
+
+    auto remote_context = ggml_openvino_get_remote_context();
+
+    std::shared_ptr<ov::Tensor> ov_tensor;
+    if (is_remote) {
+        GGML_ASSERT(ggml_openvino_is_gpu());
+        auto gpu_context = remote_context->as<ov::intel_gpu::ocl::ClContext>();
+        auto usm_tensor = gpu_context.create_tensor(element_type, shape, tensor->data);
+        ov_tensor = std::make_shared<ov::intel_gpu::ocl::USMTensor>(std::move(usm_tensor));
+    } else {
+        ov_tensor = std::make_shared<ov::Tensor>(element_type, shape, tensor->data);
+    }
+
+    return new ggml_openvino_tensor_extra(ov_tensor);
+}

llama.cpp/ggml/src/ggml-openvino/ggml-openvino-extra.h

@@ -0,0 +1,235 @@
+#pragma once
+
+#include "ggml.h"
+#include "openvino/runtime/core.hpp"
+
+#define CL_TARGET_OPENCL_VERSION 300
+#include <CL/cl.h>
+
+#include <cstdlib>
+#include <memory>
+#include <openvino/core/node.hpp>
+#include <openvino/runtime/remote_context.hpp>
+#include <openvino/runtime/tensor.hpp>
+#include <optional>
+#include <string>
+
+// ExtraQuantType enum - defines requantization target formats
+// Q4_1_64: u4, group 64, *true* asymmetric (per-group scale and zero point). Note that
+// Q4_0_128/Q4_0_C are symmetric despite taking the unsigned branch of quantize_q4_0 -- that branch
+// pins zp to 8 with d = max/-8, which is algebraically symmetric.
+enum class ExtraQuantType { F16, Q4_0_C, Q8_1_C, Q4_0_128, Q4_0_64, Q8_0_C, Q8_0_32, Q4_1_64 };
+
+ov::Core & ov_singleton_core();
+
+// Get the remote context for the current device (returns empty optional for CPU)
+std::optional<ov::RemoteContext> ggml_openvino_get_remote_context();
+
+// Get the compile config for the current device
+const ov::AnyMap & ggml_openvino_get_compile_config();
+
+// Get the OpenCL command queue for GPU operations (returns nullptr for CPU/NPU)
+cl_command_queue ggml_openvino_get_cl_queue();
+
+// Intel USM extension function type
+typedef cl_int(CL_API_CALL * clEnqueueMemFillINTEL_fn)(cl_command_queue queue,
+                                                       void * dst_ptr,
+                                                       const void * pattern,
+                                                       size_t pattern_size,
+                                                       size_t size,
+                                                       cl_uint num_events_in_wait_list,
+                                                       const cl_event * event_wait_list,
+                                                       cl_event * event);
+
+typedef cl_int(CL_API_CALL * clEnqueueMemcpyINTEL_fn)(cl_command_queue queue,
+                                                      cl_bool blocking,
+                                                      void * dst_ptr,
+                                                      const void * src_ptr,
+                                                      size_t size,
+                                                      cl_uint num_events_in_wait_list,
+                                                      const cl_event * event_wait_list,
+                                                      cl_event * event);
+
+// Get the clEnqueueMemFillINTEL function pointer (returns nullptr if not available)
+clEnqueueMemFillINTEL_fn ggml_openvino_get_clEnqueueMemFillINTEL();
+
+// Get the clEnqueueMemcpyINTEL function pointer (returns nullptr if not available)
+clEnqueueMemcpyINTEL_fn ggml_openvino_get_clEnqueueMemcpyINTEL();
+
+// =====================================================
+// Global Device Configuration (singleton)
+// =====================================================
+// Initialized once during backend init from GGML_OPENVINO_DEVICE env var
+
+struct ggml_openvino_device_config {
+    std::string device_name = "CPU";
+    std::vector<std::string> available_devices;
+    bool is_npu = false;
+    bool initialized = false;
+    std::optional<ov::RemoteContext> remote_context;
+    size_t max_alloc_size = SIZE_MAX;
+    ov::AnyMap compile_config;
+    std::unordered_map<std::string, std::string> environment_variables;
+    cl_command_queue cl_queue = nullptr;
+    clEnqueueMemFillINTEL_fn cl_mem_fill_fn = nullptr;
+    clEnqueueMemcpyINTEL_fn cl_mem_cpy_fn = nullptr;
+
+    void init();
+    ~ggml_openvino_device_config();
+};
+
+// Get the global device config singleton
+ggml_openvino_device_config & ggml_openvino_get_device_config();
+
+// Initialize device config (call during backend init)
+void ggml_openvino_init_device_config();
+
+// Get the device name
+const std::string & ggml_openvino_get_device_name();
+
+// Get all available physical OpenVINO devices
+std::vector<std::string> ggml_openvino_get_available_devices();
+
+// Human-readable device name, e.g. "Intel(R) AI Boost (NPU 4000)"; the device id if unavailable
+std::string ggml_openvino_get_device_description(const std::string & device_name);
+
+// Environment variable accessors. All GGML_OPENVINO_* env vars are read once
+// during backend init and cached on the device config; consumers must go
+// through these helpers (never call ::getenv directly) so behavior stays
+// consistent and centralized.
+//
+// Use ggml_openvino_getenv_str() for string / path values
+// (e.g. GGML_OPENVINO_DEVICE, GGML_OPENVINO_CACHE_DIR). The optional
+// default_value is returned when the var is unset or empty.
+//
+// Use ggml_openvino_getenv_int() for boolean toggles and integer settings.
+// It returns std::atoi(value) when set, otherwise default_value. For
+// boolean use, `if (ggml_openvino_getenv_int(name))` is true iff the value
+// is a non-zero integer (so "0" disables, "1" enables).
+const char * ggml_openvino_getenv_str(const char * var, const char * default_value = nullptr);
+int ggml_openvino_getenv_int(const char * var, int default_value = 0);
+
+// Memory optimization toggles. GGML_OPENVINO_MEMORY_OPTIMIZE is an umbrella
+// switch; the fine-grained env vars still override it when explicitly set.
+bool ggml_openvino_reduce_compile_mem_enabled();
+bool ggml_openvino_release_weights_enabled();
+
+// Check if running on NPU
+bool ggml_openvino_is_npu();
+
+// Check if running on a GPU (GPU, GPU.0, GPU.1, ...)
+bool ggml_openvino_is_gpu();
+
+// Largest single memory object the device can allocate, SIZE_MAX when there is no known limit
+size_t ggml_openvino_max_alloc_size();
+
+// Host weight-buffer release (GGML_OPENVINO_RELEASE_WEIGHTS, GPU only).
+// register: record a host weight buffer (idempotent per data pointer).
+// release:  madvise(MADV_DONTNEED) all registered buffers, dropping their RSS.
+// released: true once release has run (used to fail-fast on post-release recompile).
+void ggml_openvino_register_weight_buffer(void * data, size_t size);
+void ggml_openvino_release_weight_buffers();
+bool ggml_openvino_weight_buffers_released();
+
+// Get requantization type for a tensor type (returns nullopt if no requant needed)
+std::optional<ExtraQuantType> ggml_openvino_get_requant_type(const ggml_tensor * tensor, bool no_requant = false);
+
+// =====================================================
+// OpenVINO Tensor Extra Types
+// =====================================================
+// These types are stored in tensor->extra by the OpenVINO backend buffer.
+// They allow:
+// 1. Pre-built ov::Constant nodes for weights (avoiding memcpy during graph construction)
+// 2. ov::Tensor wrappers for KV cache / compute tensors (for direct use with infer_request)
+
+// Base class for OpenVINO tensor extra data
+struct ggml_openvino_extra_base {
+    enum class Type { WEIGHT, QUANTIZED_WEIGHT, TENSOR };
+    Type type;
+    virtual ~ggml_openvino_extra_base() = default;
+protected:
+    explicit ggml_openvino_extra_base(Type t) : type(t) {}
+};
+
+// Extra data for F16/F32/BF16 weight tensors - stores the pre-built weight node
+struct ggml_openvino_weight_extra : public ggml_openvino_extra_base {
+    ov::Tensor weights;                     // The underlying weight data tensor
+    std::shared_ptr<ov::Node> weight_node;  // Pre-built OpenVINO weight node
+
+    ggml_openvino_weight_extra(ov::Tensor w, std::shared_ptr<ov::Node> n) :
+        ggml_openvino_extra_base(Type::WEIGHT),
+        weights(std::move(w)),
+        weight_node(std::move(n)) {}
+};
+
+// Extra data for quantized weight tensors - stores extracted weights/scales/zp and weight node
+struct ggml_openvino_quantized_weight_extra : public ggml_openvino_extra_base {
+    ov::Tensor weights;                     // U4 or U8 extracted weights
+    ov::Tensor scales;                      // F16 scales
+    ov::Tensor zp;                          // U4 or U8 zero points (same type as weights)
+    std::shared_ptr<ov::Node> weight_node;  // Pre-built OpenVINO weight subgraph
+
+    ggml_openvino_quantized_weight_extra(ov::Tensor w, ov::Tensor s, ov::Tensor z, std::shared_ptr<ov::Node> n) :
+        ggml_openvino_extra_base(Type::QUANTIZED_WEIGHT),
+        weights(std::move(w)),
+        scales(std::move(s)),
+        zp(std::move(z)),
+        weight_node(std::move(n)) {}
+};
+
+// Extra data for KV cache / compute tensors - stores ov::Tensor for infer_request
+struct ggml_openvino_tensor_extra : public ggml_openvino_extra_base {
+    std::shared_ptr<ov::Tensor> tensor;  // For direct use with infer_request
+
+    explicit ggml_openvino_tensor_extra(std::shared_ptr<ov::Tensor> t) :
+        ggml_openvino_extra_base(Type::TENSOR),
+        tensor(std::move(t)) {}
+};
+
+// =====================================================
+// Extracted Size Calculation for Quantized Tensors
+// =====================================================
+// For quantized tensors, we need extra space to store extracted weights, scales, and zero points.
+// Returns the total size needed in the buffer for extracted data.
+
+struct ggml_openvino_extracted_layout {
+    size_t total_size = 0;      // Total bytes needed
+    size_t weights_offset = 0;  // Offset to weights in buffer
+    size_t weights_size = 0;    // Size of weights in bytes
+    size_t scales_offset = 0;   // Offset to scales in buffer
+    size_t scales_size = 0;     // Size of scales in bytes
+    size_t zp_offset = 0;       // Offset to zero points in buffer
+    size_t zp_size = 0;         // Size of zero points in bytes (U4 or U8)
+    bool is_u4;                 // true for U4 weights, false for U8
+    int64_t weights_per_block;  // weights per scale/zp block
+    bool is_symmetric;          // true for symmetric quantization
+
+    // Requantization info
+    bool is_requant = false;                     // true if this tensor needs requantization
+    std::optional<ExtraQuantType> requant_type;  // target requant type if is_requant
+};
+
+// Calculate the buffer layout for extracted quantized data
+ggml_openvino_extracted_layout ggml_openvino_get_extracted_layout(const ggml_tensor * tensor, bool use_bias = false);
+
+ggml_openvino_tensor_extra * ggml_openvino_create_tensor_extra(const ggml_tensor * tensor, bool is_remote);
+
+// Check if a tensor's buffer uses remote (device) memory (e.g. GPU USM)
+bool ggml_openvino_buffer_is_remote(const ggml_tensor * tensor);
+
+// Register an extra with the tensor's OpenVINO buffer context for proper lifetime management.
+// This sets tensor->extra and tracks the extra in the buffer context for cleanup.
+void ggml_openvino_buffer_register_extra(ggml_tensor * tensor, ggml_openvino_extra_base * extra);
+
+// =====================================================
+// OpenVINO Backend Context and Interface
+// =====================================================
+struct ggml_backend_openvino_context {
+    int device = 0;
+    std::string name = "OpenVINO";
+    std::string description = "OpenVINO Backend Context";
+
+    std::shared_ptr<void> runtime_context = nullptr;
+
+    ggml_backend_openvino_context() = default;
+};

llama.cpp/ggml/src/ggml-openvino/ggml-quants.h

@@ -0,0 +1,58 @@
+#pragma once
+#include "ggml-openvino-extra.h"  // For ExtraQuantType
+#include "ggml.h"
+
+#include <openvino/core/node_output.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/runtime/tensor.hpp>
+
+static constexpr size_t GGML_QUANTIZATION_GROUP_SIZE = 32;
+
+inline const char * extra_quant_type_name(ExtraQuantType t) {
+    switch (t) {
+    case ExtraQuantType::F16:
+        return "F16";
+    case ExtraQuantType::Q4_0_C:
+        return "Q4_0_C";
+    case ExtraQuantType::Q4_0_128:
+        return "Q4_0_128";
+    case ExtraQuantType::Q8_0_C:
+        return "Q8_0_C";
+    case ExtraQuantType::Q8_0_32:
+        return "Q8_0_32";
+    case ExtraQuantType::Q8_1_C:
+        return "Q8_1_C";
+    case ExtraQuantType::Q4_0_64:
+        return "Q4_0_64";
+    case ExtraQuantType::Q4_1_64:
+        return "Q4_1_64";
+    default:
+        return "unknown";
+    }
+}
+
+// Result from process_weight_tensor containing the weight node and tensors.
+// For quantized weights, also contains the extracted layout and scale/zp tensors.
+struct OvWeight {
+    std::shared_ptr<ov::Node> weight_node;
+    ggml_openvino_extracted_layout layout;  // Only meaningful for quantized (layout.total_size > 0)
+    ov::Tensor weights;
+    ov::Tensor scales;
+    ov::Tensor zp;
+
+    bool is_quantized() const { return layout.scales_size > 0; }
+};
+
+// Process weight tensor and create an OpenVINO weight node
+// Handles F16/F32/BF16 and quantized weights, with optional requantization
+// If output_base_ptr is nullptr, allocates internal buffers (for decoder use)
+// If output_base_ptr is provided, uses pre-allocated buffers at specified offsets (for backend buffer use)
+// Returns OvWeight with the weight node and optional quantized tensors
+OvWeight process_weight_tensor(
+    const ggml_tensor * tensor,
+    const void * data,                 // Source data pointer (may differ from tensor->data)
+    void * output_base_ptr = nullptr,  // Base pointer for output buffers (or nullptr for internal allocation)
+    bool use_bias = false);            // Use an exact f16 zero point (vs. a rounded integer one);
+                                       // always used for for_gather_matmul (3D MoE expert) weights
+                                       // regardless of this flag, and also settable explicitly for
+                                       // test-backend-ops.

llama.cpp/ggml/src/ggml-openvino/model-cache.cpp

@@ -0,0 +1,484 @@
+#include "model-cache.h"
+
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+#include "ggml-impl.h"
+#include "ggml-openvino-extra.h"
+
+#include <cerrno>
+#include <algorithm>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <iomanip>
+#include <limits>
+#include <sstream>
+#include <openvino/core/version.hpp>
+#include <string>
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <vector>
+
+#if defined(_WIN32)
+#    define WIN32_LEAN_AND_MEAN
+#    ifndef NOMINMAX
+#        define NOMINMAX
+#    endif
+#    include <windows.h>
+#    include <psapi.h>
+#    include <direct.h>
+#    include <process.h>
+#else
+#    include <unistd.h>
+#endif
+#ifdef __linux__
+#    include <sys/sysmacros.h>
+#endif
+
+namespace {
+
+// 64-bit FNV-1a, the mixing primitive for all fingerprints here.
+inline uint64_t fnv1a(uint64_t h, const void * data, size_t n) {
+    const uint8_t * p = static_cast<const uint8_t *>(data);
+    for (size_t i = 0; i < n; ++i) {
+        h ^= p[i];
+        h *= 0x100000001b3ull;
+    }
+    return h;
+}
+
+inline uint64_t fnv1a_u64(uint64_t h, uint64_t v) {
+    return fnv1a(h, &v, sizeof(v));
+}
+
+constexpr uint64_t FNV_OFFSET = 0xcbf29ce484222325ull;
+
+// Fallback when source-file identity is unavailable outside cache-only mode.
+constexpr size_t WEIGHT_SAMPLE_BYTES = 4096;
+
+// Is this src a model weight, mirroring create_weight_nodes()'s selection:
+// non-view tensor whose buffer is USAGE_WEIGHTS or whose type is quantized.
+bool is_weight_src(const ggml_tensor * src) {
+    if (src == nullptr || src->view_src != nullptr || src->buffer == nullptr) {
+        return false;
+    }
+    return src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS || ggml_is_quantized(src->type);
+}
+
+// Weight metadata and source identity; do not read repacked or unloaded buffers.
+uint64_t weight_fingerprint(const ggml_tensor * t) {
+    uint64_t h = FNV_OFFSET;
+    h = fnv1a(h, t->name, strlen(t->name));
+    for (int i = 0; i < GGML_MAX_DIMS; ++i) {
+        h = fnv1a_u64(h, static_cast<uint64_t>(t->ne[i]));
+    }
+    h = fnv1a_u64(h, static_cast<uint64_t>(t->type));
+    const size_t nbytes = ggml_nbytes(t);
+    h = fnv1a_u64(h, nbytes);
+    return fnv1a_u64(h, ggml_backend_openvino_weight_fingerprint(t));
+}
+
+// Walk the cgraph and invoke fn(weight_tensor) for each distinct weight, in node
+// order. De-duplicates by tensor pointer so a weight used by several nodes is
+// fingerprinted once, deterministically.
+template <typename F>
+void for_each_weight(const ggml_cgraph * cgraph, F && fn) {
+    std::vector<const ggml_tensor *> seen;
+    for (int i = 0; i < cgraph->n_nodes; ++i) {
+        const ggml_tensor * node = cgraph->nodes[i];
+        for (int s = 0; s < GGML_MAX_SRC; ++s) {
+            const ggml_tensor * src = node->src[s];
+            if (!is_weight_src(src)) {
+                continue;
+            }
+            bool dup = false;
+            for (const auto * p : seen) {
+                if (p == src) {
+                    dup = true;
+                    break;
+                }
+            }
+            if (dup) {
+                continue;
+            }
+            seen.push_back(src);
+            fn(src);
+        }
+    }
+}
+
+std::string ov_version_string() {
+    const ov::Version v = ov::get_openvino_version();
+    return std::string(v.buildNumber ? v.buildNumber : "unknown");
+}
+
+std::string hex64(uint64_t v) {
+    char buf[17];
+    snprintf(buf, sizeof(buf), "%016llx", static_cast<unsigned long long>(v));
+    return std::string(buf);
+}
+
+// Portable mkdir for a single path component. Returns true if the directory
+// exists after the call (created now or already present).
+bool make_dir(const std::string & path) {
+#if defined(_WIN32)
+    int rc = _mkdir(path.c_str());
+#else
+    int rc = ::mkdir(path.c_str(), 0755);
+#endif
+    if (rc == 0 || errno == EEXIST) {
+        return true;
+    }
+    return false;
+}
+
+// Create `path` and any missing parents (like `mkdir -p`). Best-effort:
+// returns true only if the full directory exists afterwards.
+bool make_dirs(const std::string & path) {
+    if (path.empty()) {
+        return false;
+    }
+    std::string acc;
+    for (size_t i = 0; i < path.size(); ++i) {
+        const char c = path[i];
+        acc.push_back(c);
+        const bool sep = (c == '/'
+#if defined(_WIN32)
+                          || c == '\\'
+#endif
+        );
+        // Create each intermediate component (skip a leading "/" root).
+        if (sep && acc.size() > 1) {
+            std::string component = acc.substr(0, acc.size() - 1);
+            if (!make_dir(component)) {
+                return false;
+            }
+        }
+    }
+    return make_dir(path);
+}
+
+}  // namespace
+
+bool ggml_openvino_model_cache_only() {
+    return ggml_openvino_getenv_int("GGML_OPENVINO_COMPILED_MODEL_CACHE_ONLY") != 0;
+}
+
+static const char * cache_settings[] = {
+    "GGML_OPENVINO_REQUANT_KQUANT",
+    "GGML_OPENVINO_NATIVE_SOFTPLUS",
+    "GGML_OPENVINO_DISABLE_KV_SLICE",
+    "GGML_OPENVINO_MANUAL_GQA_ATTN",
+    "GGML_OPENVINO_STATEFUL_EXECUTION",
+    "GGML_OPENVINO_DISABLE_KV_STATE_RELAYOUT",
+    "GGML_OPENVINO_DISABLE_REMOTE_OUTPUTS",
+    "GGML_OPENVINO_REDUCE_COMPILE_MEM",
+    "GGML_OPENVINO_MEMORY_OPTIMIZE",
+    "GGML_OPENVINO_PROFILING",
+};
+
+void ggml_openvino_model_cache_init() {
+    const bool cache_only = ggml_openvino_model_cache_only();
+    const std::string dir = ggml_openvino_model_cache_dir();
+    if (dir.empty()) {
+        if (cache_only) {
+            GGML_ABORT("ggml-openvino: cache-only mode requires GGML_OPENVINO_COMPILED_MODEL_CACHE_DIR");
+        }
+        return;
+    }
+    if (cache_only && (ggml_openvino_is_npu() || ggml_openvino_getenv_int("GGML_OPENVINO_FORCE_STATIC") ||
+                       ggml_openvino_getenv_int("GGML_OPENVINO_DISABLE_CACHE") ||
+                       ggml_openvino_getenv_int("GGML_OPENVINO_ENABLE_FALLBACK"))) {
+        GGML_ABORT("ggml-openvino: cache-only mode requires dynamic CPU/GPU execution with caching and without fallback");
+    }
+#if !defined(__linux__) && !defined(_WIN32)
+    if (cache_only) {
+        GGML_ABORT("ggml-openvino: cache-only mmap identification requires Linux or Windows");
+    }
+#endif
+    if (cache_only) {
+        auto & config = ggml_openvino_get_device_config();
+        config.environment_variables.erase("GGML_OPENVINO_SPILL_DIR");
+        config.environment_variables["GGML_OPENVINO_RELEASE_WEIGHTS"] = "0";
+    }
+}
+
+uint64_t ggml_openvino_source_fingerprint(const void * data, size_t size, std::vector<ggml_openvino_source_mapping> & mappings) {
+    const uintptr_t address = reinterpret_cast<uintptr_t>(data);
+    auto contains = [&](const ggml_openvino_source_mapping & m) {
+        return address >= m.begin && address < m.end && size <= m.end - address;
+    };
+    auto fingerprint = [&](const ggml_openvino_source_mapping & m) {
+        return fnv1a_u64(m.identity, m.offset + address - m.begin);
+    };
+    for (const auto & m : mappings) {
+        if (contains(m)) {
+            return fingerprint(m);
+        }
+    }
+#ifdef __linux__
+    std::ifstream maps("/proc/self/maps");
+    std::string line;
+    while (std::getline(maps, line)) {
+        unsigned long long begin, end, offset, inode;
+        unsigned int dev_major, dev_minor;
+        char permissions[5];
+        int path_start = 0;
+        if (sscanf(line.c_str(), "%llx-%llx %4s %llx %x:%x %llu %n", &begin, &end, permissions,
+                   &offset, &dev_major, &dev_minor, &inode, &path_start) != 7 || inode == 0) {
+            continue;
+        }
+        ggml_openvino_source_mapping m{uintptr_t(begin), uintptr_t(end), offset, FNV_OFFSET};
+        if (!contains(m)) {
+            continue;
+        }
+        struct stat st;
+        const std::string path = line.substr(path_start);
+        if (stat(path.c_str(), &st) != 0 || !S_ISREG(st.st_mode) || uint64_t(st.st_ino) != inode ||
+            major(st.st_dev) != dev_major || minor(st.st_dev) != dev_minor) {
+            break;
+        }
+        m.identity = fnv1a_u64(m.identity, st.st_dev);
+        m.identity = fnv1a_u64(m.identity, st.st_ino);
+        m.identity = fnv1a_u64(m.identity, st.st_size);
+        m.identity = fnv1a_u64(m.identity, st.st_mtim.tv_sec);
+        m.identity = fnv1a_u64(m.identity, st.st_mtim.tv_nsec);
+        m.identity = fnv1a_u64(m.identity, st.st_ctim.tv_sec);
+        m.identity = fnv1a_u64(m.identity, st.st_ctim.tv_nsec);
+        mappings.push_back(m);
+        return fingerprint(m);
+    }
+#elif defined(_WIN32)
+    MEMORY_BASIC_INFORMATION memory;
+    if (VirtualQuery(data, &memory, sizeof(memory)) == sizeof(memory) && memory.Type == MEM_MAPPED) {
+        std::wstring name(MAX_PATH, L'\0');
+        DWORD length = 0;
+        while (name.size() <= 32768) {
+            length = GetMappedFileNameW(GetCurrentProcess(), const_cast<void *>(data), name.data(), static_cast<DWORD>(name.size()));
+            if (length == 0 || length < name.size() - 1) {
+                break;
+            }
+            name.resize(name.size() * 2);
+        }
+        if (length > 0 && length < name.size() - 1) {
+            name.resize(length);
+            const std::wstring path = L"\\\\?\\GLOBALROOT" + name;
+            HANDLE file = CreateFileW(path.c_str(), FILE_READ_ATTRIBUTES,
+                                      FILE_SHARE_READ | FILE_SHARE_WRITE | FILE_SHARE_DELETE, nullptr,
+                                      OPEN_EXISTING, FILE_ATTRIBUTE_NORMAL, nullptr);
+            if (file != INVALID_HANDLE_VALUE) {
+                BY_HANDLE_FILE_INFORMATION info;
+                FILE_BASIC_INFO basic;
+                const bool valid = GetFileInformationByHandle(file, &info) &&
+                                   GetFileInformationByHandleEx(file, FileBasicInfo, &basic, sizeof(basic));
+                CloseHandle(file);
+                if (valid) {
+                    const uint64_t file_size = (uint64_t(info.nFileSizeHigh) << 32) | info.nFileSizeLow;
+                    const uintptr_t begin = reinterpret_cast<uintptr_t>(memory.AllocationBase);
+                    if (file_size <= std::numeric_limits<uintptr_t>::max() - begin) {
+                        ggml_openvino_source_mapping m{begin, begin + static_cast<uintptr_t>(file_size), 0,
+                                                       fnv1a(FNV_OFFSET, "win32", 5)};
+                        if (contains(m)) {
+                            m.identity = fnv1a_u64(m.identity, info.dwVolumeSerialNumber);
+                            m.identity = fnv1a_u64(m.identity, (uint64_t(info.nFileIndexHigh) << 32) | info.nFileIndexLow);
+                            m.identity = fnv1a_u64(m.identity, file_size);
+                            m.identity = fnv1a_u64(m.identity, (uint64_t(info.ftLastWriteTime.dwHighDateTime) << 32) |
+                                                                 info.ftLastWriteTime.dwLowDateTime);
+                            m.identity = fnv1a_u64(m.identity, static_cast<uint64_t>(basic.ChangeTime.QuadPart));
+                            mappings.push_back(m);
+                            return fingerprint(m);
+                        }
+                    }
+                }
+            }
+        }
+    }
+#endif
+    if (ggml_openvino_model_cache_only()) {
+        GGML_ABORT("ggml-openvino: could not identify mapped GGUF weight; use --load-mode mmap");
+    }
+    uint64_t h = FNV_OFFSET;
+    const size_t head = std::min(size, WEIGHT_SAMPLE_BYTES);
+    h = fnv1a(h, data, head);
+    if (size > head) {
+        const size_t tail = std::min(size - head, WEIGHT_SAMPLE_BYTES);
+        h = fnv1a(h, static_cast<const uint8_t *>(data) + size - tail, tail);
+    }
+    return h;
+}
+
+std::string ggml_openvino_model_cache_dir() {
+    const char * dir = ggml_openvino_getenv_str("GGML_OPENVINO_COMPILED_MODEL_CACHE_DIR");
+    if (!dir || strlen(dir) == 0) {
+        return std::string();
+    }
+    std::string path(dir);
+    if (ggml_openvino_model_cache_only()) {
+        return path;
+    }
+    // Create the cache directory (and parents) on first use so callers don't
+    // have to pre-create it; a missing dir would otherwise silently disable the
+    // cache (manifest/blob writes fail with no directory to write into).
+    if (!make_dirs(path)) {
+        GGML_LOG_WARN("ggml-openvino: could not create model cache dir '%s' (errno=%d); caching disabled\n",
+                      path.c_str(), errno);
+        return std::string();
+    }
+    return path;
+}
+
+std::string ggml_openvino_model_cache_temp_path(const std::string & path) {
+#ifdef _WIN32
+    const int pid = _getpid();
+#else
+    const int pid = getpid();
+#endif
+    return path + ".tmp." + std::to_string(pid) + "." + std::to_string(ggml_time_us());
+}
+
+uint64_t ggml_openvino_model_fingerprint(const ggml_cgraph * cgraph,
+                                         const std::string & device,
+                                         bool fa,
+                                         const int32_t * rope_params,
+                                         int rope_len,
+                                         uint64_t extra_cfg,
+                                         const std::string & graph_signature) {
+    uint64_t h = FNV_OFFSET;
+    h = fnv1a_u64(h, 2);
+    h = fnv1a(h, graph_signature.data(), graph_signature.size());
+    for (const char * name : cache_settings) {
+        const char * value = ggml_openvino_getenv_str(name, "");
+        h = fnv1a(h, value, strlen(value) + 1);
+    }
+    if (const char * debug_nodes = ggml_openvino_getenv_str("GGML_OPENVINO_DEBUG_NODE")) {
+        h = fnv1a(h, "GGML_OPENVINO_DEBUG_NODE", sizeof("GGML_OPENVINO_DEBUG_NODE"));
+        h = fnv1a(h, debug_nodes, strlen(debug_nodes) + 1);
+    }
+    if (ggml_openvino_is_gpu() && ggml_openvino_getenv_int("GGML_OPENVINO_MOE_OP", 1) == 0) {
+        h = fnv1a(h, "GGML_OPENVINO_MOE_OP=0", sizeof("GGML_OPENVINO_MOE_OP=0"));
+    }
+
+    // Topology: node count + each node's op and name (cheap, and distinguishes
+    // graphs that share weights but differ structurally).
+    h = fnv1a_u64(h, static_cast<uint64_t>(cgraph->n_nodes));
+    for (int i = 0; i < cgraph->n_nodes; ++i) {
+        const ggml_tensor * node = cgraph->nodes[i];
+        h = fnv1a_u64(h, static_cast<uint64_t>(node->op));
+        h = fnv1a(h, node->name, strlen(node->name));
+    }
+
+    // Weights: the model identity.
+    for_each_weight(cgraph, [&](const ggml_tensor * t) { h = fnv1a_u64(h, weight_fingerprint(t)); });
+
+    // Device, model parameters, and backend configuration.
+    h = fnv1a(h, device.data(), device.size());
+    h = fnv1a_u64(h, fa ? 1u : 0u);
+    if (rope_params && rope_len > 0) {
+        h = fnv1a(h, rope_params, sizeof(int32_t) * static_cast<size_t>(rope_len));
+    }
+    h = fnv1a_u64(h, extra_cfg);
+    const std::string ver = ov_version_string();
+    h = fnv1a(h, ver.data(), ver.size());
+
+    return h;
+}
+
+std::string ggml_openvino_model_cache_blob_path(const std::string & dir, uint64_t fingerprint) {
+    return dir + "/" + hex64(fingerprint) + ".blob";
+}
+
+std::string ggml_openvino_model_cache_manifest_path(const std::string & dir, uint64_t fingerprint) {
+    return dir + "/" + hex64(fingerprint) + ".manifest";
+}
+
+bool ggml_openvino_model_cache_write_manifest(const std::string & path,
+                                              const ggml_cgraph * cgraph,
+                                              uint64_t fingerprint,
+                                              const std::vector<std::string> & inputs,
+                                              const std::vector<std::string> & outputs) {
+    std::ofstream f(path, std::ios::trunc);
+    if (!f.is_open()) {
+        return false;
+    }
+    f << "fingerprint " << hex64(fingerprint) << "\n";
+    f << "ov_version " << ov_version_string() << "\n";
+    for_each_weight(cgraph, [&](const ggml_tensor * t) {
+        f << t->name << " " << t->ne[0] << " " << t->ne[1] << " " << t->ne[2] << " " << t->ne[3] << " "
+          << static_cast<int>(t->type) << " " << hex64(weight_fingerprint(t)) << "\n";
+    });
+    f << "ports\n";
+    for (const auto * names : { &inputs, &outputs }) {
+        f << names->size() << '\n';
+        for (const auto & name : *names) {
+            f << std::quoted(name) << '\n';
+        }
+    }
+    return f.good();
+}
+
+bool ggml_openvino_model_cache_verify_manifest(const std::string & path,
+                                               const ggml_cgraph * cgraph,
+                                               uint64_t fingerprint,
+                                               std::vector<std::string> & inputs,
+                                               std::vector<std::string> & outputs) {
+    std::ifstream f(path);
+    if (!f.is_open()) {
+        return false;
+    }
+    std::string tag;
+    std::string val;
+    // header: fingerprint
+    if (!(f >> tag >> val) || tag != "fingerprint" || val != hex64(fingerprint)) {
+        return false;
+    }
+    // header: ov_version
+    if (!(f >> tag >> val) || tag != "ov_version" || val != ov_version_string()) {
+        return false;
+    }
+
+    // Build the expected per-weight lines from the live cgraph, then require an
+    // exact match (same set, same order) against the manifest.
+    std::vector<std::string> expected;
+    for_each_weight(cgraph, [&](const ggml_tensor * t) {
+        expected.push_back(std::string(t->name) + " " + std::to_string(t->ne[0]) + " " + std::to_string(t->ne[1]) +
+                           " " + std::to_string(t->ne[2]) + " " + std::to_string(t->ne[3]) + " " +
+                           std::to_string(static_cast<int>(t->type)) + " " + hex64(weight_fingerprint(t)));
+    });
+
+    size_t idx = 0;
+    std::string line;
+    std::getline(f, line);  // consume rest of ov_version line
+    while (idx < expected.size() && std::getline(f, line)) {
+        if (line.empty()) {
+            continue;
+        }
+        if (idx >= expected.size() || line != expected[idx]) {
+            return false;
+        }
+        ++idx;
+    }
+    if (idx != expected.size()) {
+        return false;
+    }
+    if (!std::getline(f, line)) {
+        return true;
+    }
+    if (line != "ports") {
+        return false;
+    }
+    for (auto * names : { &inputs, &outputs }) {
+        size_t count;
+        if (!(f >> count) || count > 100000) {
+            return false;
+        }
+        for (size_t i = 0; i < count; ++i) {
+            std::string name;
+            if (!(f >> std::quoted(name))) {
+                return false;
+            }
+            names->push_back(name);
+        }
+    }
+    f >> std::ws;
+    return f.eof();
+}

llama.cpp/ggml/src/ggml-openvino/model-cache.h

@@ -0,0 +1,57 @@
+#pragma once
+
+// Compiled blobs include weights. Cache-only execution skips weight uploads and graph compilation.
+
+#include "ggml.h"
+
+#include <cstdint>
+#include <string>
+#include <vector>
+
+bool ggml_openvino_model_cache_only();
+void ggml_openvino_model_cache_init();
+
+struct ggml_openvino_source_mapping {
+    uintptr_t begin;
+    uintptr_t end;
+    uint64_t offset;
+    uint64_t identity;
+};
+
+// Identify mmap weights without reading their pages. Cache mappings for one buffer lifetime.
+uint64_t ggml_openvino_source_fingerprint(const void * data, size_t size, std::vector<ggml_openvino_source_mapping> & mappings);
+uint64_t ggml_backend_openvino_weight_fingerprint(const ggml_tensor * tensor);
+
+// Returns the compiled-model cache directory, or empty if unset.
+std::string ggml_openvino_model_cache_dir();
+std::string ggml_openvino_model_cache_temp_path(const std::string & path);
+
+// Hash graph structure, source weight identities, configuration, and OpenVINO version.
+uint64_t ggml_openvino_model_fingerprint(const ggml_cgraph * cgraph,
+                                         const std::string & device,
+                                         bool fa,
+                                         const int32_t * rope_params,
+                                         int rope_len,
+                                         uint64_t extra_cfg,
+                                         const std::string & graph_signature);
+
+// Path to the compiled-blob file for a fingerprint (<dir>/<hex>.blob).
+std::string ggml_openvino_model_cache_blob_path(const std::string & dir, uint64_t fingerprint);
+
+// Path to the sidecar manifest (<dir>/<hex>.manifest) holding the per-weight
+// fingerprints, used to re-verify a hit before trusting the blob.
+std::string ggml_openvino_model_cache_manifest_path(const std::string & dir, uint64_t fingerprint);
+
+// Record weight metadata and source identities. Returns false on I/O error.
+bool ggml_openvino_model_cache_write_manifest(const std::string & path,
+                                              const ggml_cgraph * cgraph,
+                                              uint64_t fingerprint,
+                                              const std::vector<std::string> & inputs,
+                                              const std::vector<std::string> & outputs);
+
+// Require all weight metadata and source identities to match the manifest.
+bool ggml_openvino_model_cache_verify_manifest(const std::string & path,
+                                               const ggml_cgraph * cgraph,
+                                               uint64_t fingerprint,
+                                               std::vector<std::string> & inputs,
+                                               std::vector<std::string> & outputs);

llama.cpp/ggml/src/ggml-openvino/openvino/decoder.h

@@ -0,0 +1,133 @@
+#pragma once
+
+#include <cstdint>
+#include <map>
+#include <openvino/core/node.hpp>
+#include <openvino/core/partial_shape.hpp>
+#include <openvino/core/shape.hpp>
+#include <openvino/frontend/decoder.hpp>
+#include <set>
+#include <string>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+
+struct ModelInputInfo {
+    element::Type type;
+    PartialShape shape;
+};
+
+struct ModelExtraInputInfo {
+    element::Type type;
+    Shape shape;
+    int64_t value;
+    bool is_parameter;
+};
+
+class GgmlDecoder : public DecoderBase {
+public:
+    virtual ov::Any get_attribute(const std::string & name) const = 0;
+
+    virtual PartialShape get_input_shape(int node_idx, const std::string & name) const = 0;
+
+    virtual std::vector<size_t> get_input_stride(int node_idx, const std::string & name) const = 0;
+
+    virtual size_t get_view_input_size(int node_idx, const std::string & name) const = 0;
+
+    virtual size_t get_view_input_offset(int node_idx, const std::string & name, size_t view_index) const = 0;
+
+    virtual size_t get_view_input_src_offset(int node_idx, const std::string & name, size_t view_index) const = 0;
+
+    virtual std::vector<size_t> get_view_input_stride(int node_idx,
+                                                      const std::string & name,
+                                                      size_t view_index) const = 0;
+
+    virtual std::vector<size_t> get_view_input_src_stride(int node_idx,
+                                                          const std::string & name,
+                                                          size_t view_index) const = 0;
+
+    virtual Shape get_view_input_ggml_shape(int node_idx, const std::string & name, size_t view_index) const = 0;
+
+    virtual Shape get_view_input_src_ggml_shape(int node_idx, const std::string & name, size_t view_index) const = 0;
+
+    virtual PartialShape get_view_input_ov_shape(int node_idx, const std::string & name, size_t view_index) const = 0;
+
+    virtual PartialShape get_view_input_src_ov_shape(int node_idx,
+                                                     const std::string & name,
+                                                     size_t view_index) const = 0;
+
+    virtual std::string get_view_input_name(int node_idx, const std::string & name, size_t view_index) const = 0;
+
+    virtual std::string get_view_input_src_name(int node_idx, const std::string & name, size_t view_index) const = 0;
+
+    virtual element::Type get_input_type(int node_idx, const std::string & name) const = 0;
+
+    virtual size_t get_input_size() const = 0;
+
+    virtual size_t get_input_size(int node_idx) const = 0;
+
+    virtual void get_input_node(size_t input_port_idx,
+                                std::string & producer_name,
+                                std::string & producer_output_port_name,
+                                size_t & producer_output_port_index) const = 0;
+
+    virtual std::vector<std::string> get_input_names(int node_idx) const = 0;
+
+    virtual PartialShape get_output_shape(int node_idx) const = 0;
+
+    virtual element::Type get_output_type(const int node_idx) const = 0;
+
+    virtual std::vector<size_t> get_output_stride(int node_idx) const = 0;
+
+    virtual int32_t * get_input_op_params(int node_idx, const std::string & name) const = 0;
+
+    virtual int32_t * get_output_op_params(int node_idx) const = 0;
+
+    virtual size_t get_output_op_offset(int node_idx) const = 0;
+
+    virtual std::vector<std::string> get_output_names(int node_idx) const = 0;
+
+    virtual std::string get_inplace_op_src(int node_idx) const = 0;
+
+    virtual bool is_view_like_alias_of(int node_idx, const std::string & view_src_name) const = 0;
+
+    virtual const std::string & get_op_type() const = 0;
+
+    virtual const std::string & get_op_type(int node_idx) const = 0;
+
+    virtual const std::string & get_op_name() const = 0;
+
+    virtual const std::string & get_op_name(int node_idx) const = 0;
+
+    virtual void visit_subgraph(std::function<void(std::shared_ptr<GgmlDecoder>, int node_idx)> node_visitor) const = 0;
+
+    virtual int get_op_case(int node_idx) const = 0;
+
+    virtual const std::map<std::string, ModelInputInfo> & get_model_inputs() const = 0;
+    virtual const std::map<std::string, ModelExtraInputInfo> & get_model_extra_inputs() const = 0;
+    virtual const std::map<std::string, std::shared_ptr<ov::Node>> & get_model_weights() const = 0;
+    virtual std::set<std::string> get_model_output_names() const = 0;
+
+    virtual int32_t * get_rope_params() const = 0;
+
+    virtual bool has_mixed_rope_params() const = 0;
+
+    virtual int get_ssm_state_size() const = 0;
+
+    virtual std::map<std::string, std::string> get_kv_param_res_names() const = 0;
+
+    virtual bool is_static() const = 0;
+
+    virtual bool is_stateful() const = 0;
+
+    virtual bool is_splited_model() const = 0;
+
+    virtual int is_swa_layer(int layer) const = 0;
+
+    virtual int32_t get_op_dynamic_dim(int node_idx) const = 0;
+};
+
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/frontend.cpp

@@ -0,0 +1,28 @@
+#include "frontend.h"
+
+#include "input_model.h"
+#include "op_table.h"
+#include "translate_session.h"
+#include <openvino/core/type.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+
+FrontEnd::FrontEnd() {}
+
+std::shared_ptr<Model> FrontEnd::convert(const InputModel::Ptr & model, bool naive) {
+    auto ggml_model = ov::as_type_ptr<ggml::InputModel>(model);
+    FRONT_END_GENERAL_CHECK(ggml_model, "Invalid input model");
+    std::shared_ptr<Model> converted_model;
+    const auto & supported_ops = get_supported_ops();
+    {
+        TranslateSession translate_session(model, supported_ops, naive);
+        converted_model = translate_session.get_converted_model();
+    }
+    return converted_model;
+}
+
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/frontend.h

@@ -0,0 +1,22 @@
+// Copyright (C) 2018-2024 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include <openvino/frontend/frontend.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+
+class FrontEnd {
+public:
+    FrontEnd();
+
+    static std::shared_ptr<Model> convert(const InputModel::Ptr & model, bool naive = false);
+};
+
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/input_model.cpp

@@ -0,0 +1,17 @@
+#include "input_model.h"
+
+#include "decoder.h"
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+
+InputModel::InputModel(const std::shared_ptr<GgmlDecoder> & gdecoder) : m_decoder(gdecoder) {}
+
+const std::shared_ptr<GgmlDecoder> & InputModel::get_model_decoder() const {
+    return m_decoder;
+}
+
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/input_model.h

@@ -0,0 +1,29 @@
+#pragma once
+
+#include "decoder.h"
+
+#include <openvino/frontend/input_model.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+
+class FrontEnd;
+class GgmlDecoder;
+using ov::frontend::ggml::GgmlDecoder;
+
+class InputModel : public ov::frontend::InputModel {
+    friend class ::ov::frontend::ggml::FrontEnd;
+
+public:
+    explicit InputModel(const std::shared_ptr<GgmlDecoder> & gdecoder);
+
+    const std::shared_ptr<GgmlDecoder> & get_model_decoder() const;
+
+private:
+    std::shared_ptr<GgmlDecoder> m_decoder;
+};
+
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/node_context.h

@@ -0,0 +1,180 @@
+#pragma once
+
+#include "decoder.h"
+
+#include <cstdint>
+#include <openvino/frontend/node_context.hpp>
+#include <string>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+
+class TranslateSession;
+
+typedef std::map<std::string, Output<Node>> TensorMap;
+
+class NodeContext : public frontend::NodeContext {
+public:
+    NodeContext(const std::shared_ptr<GgmlDecoder> & decoder,
+                std::shared_ptr<TensorMap> & tensor_map,
+                int node_idx,
+                TranslateSession * translate_session = nullptr) :
+        ov::frontend::NodeContext(decoder->get_op_type(node_idx)),
+        m_decoder(decoder),
+        m_tensor_map(tensor_map),
+        m_node_idx(node_idx),
+        m_translate_session(translate_session) {
+        m_input_names = decoder->get_input_names(m_node_idx);
+        m_output_names = decoder->get_output_names(m_node_idx);
+    }
+
+    TranslateSession * get_translate_session() const { return m_translate_session; }
+
+    const std::vector<std::string> & get_input_names() const { return m_input_names; }
+
+    const std::vector<std::string> & get_output_names() const { return m_output_names; }
+
+    size_t get_input_size() const override { return m_decoder->get_input_size(m_node_idx); }
+
+    ov::element::Type get_input_type(size_t index) const {
+        return m_decoder->get_input_type(m_node_idx, m_input_names[index]);
+    }
+
+    PartialShape get_input_shape(size_t input_index) const {
+        return m_decoder->get_input_shape(m_node_idx, m_input_names[input_index]);
+    }
+
+    std::vector<size_t> get_input_stride(size_t index) const {
+        return m_decoder->get_input_stride(m_node_idx, m_input_names[index]);
+    }
+
+    std::string get_output_name() const { return m_output_names[0]; }
+
+    PartialShape get_output_shape() const { return m_decoder->get_output_shape(m_node_idx); }
+
+    int32_t * get_input_op_params(size_t index) const {
+        return m_decoder->get_input_op_params(m_node_idx, m_input_names[index]);
+    }
+
+    size_t get_view_input_size(size_t index) const {
+        return m_decoder->get_view_input_size(m_node_idx, m_input_names[index]);
+    }
+
+    size_t get_view_input_offset(size_t index, size_t view_index) const {
+        return m_decoder->get_view_input_offset(m_node_idx, m_input_names[index], view_index);
+    }
+
+    size_t get_view_input_src_offset(size_t index, size_t view_index) const {
+        return m_decoder->get_view_input_src_offset(m_node_idx, m_input_names[index], view_index);
+    }
+
+    std::vector<size_t> get_view_input_stride(size_t index, size_t view_index) const {
+        return m_decoder->get_view_input_stride(m_node_idx, m_input_names[index], view_index);
+    }
+
+    std::vector<size_t> get_view_input_src_stride(size_t index, size_t view_index) const {
+        return m_decoder->get_view_input_src_stride(m_node_idx, m_input_names[index], view_index);
+    }
+
+    ov::Shape get_view_input_ggml_shape(size_t index, size_t view_index) const {
+        return m_decoder->get_view_input_ggml_shape(m_node_idx, m_input_names[index], view_index);
+    }
+
+    ov::Shape get_view_input_src_ggml_shape(size_t index, size_t view_index) const {
+        return m_decoder->get_view_input_src_ggml_shape(m_node_idx, m_input_names[index], view_index);
+    }
+
+    ov::PartialShape get_view_input_ov_shape(size_t index, size_t view_index) const {
+        return m_decoder->get_view_input_ov_shape(m_node_idx, m_input_names[index], view_index);
+    }
+
+    ov::PartialShape get_view_input_src_ov_shape(size_t index, size_t view_index) const {
+        return m_decoder->get_view_input_src_ov_shape(m_node_idx, m_input_names[index], view_index);
+    }
+
+    std::string get_view_input_name(size_t index, size_t view_index) const {
+        return m_decoder->get_view_input_name(m_node_idx, m_input_names[index], view_index);
+    }
+
+    std::string get_view_input_src_name(size_t index, size_t view_index) const {
+        return m_decoder->get_view_input_src_name(m_node_idx, m_input_names[index], view_index);
+    }
+
+    int32_t get_op_dynamic_dim() const { return m_decoder->get_op_dynamic_dim(m_node_idx); }
+
+    int32_t * get_output_op_params() const { return m_decoder->get_output_op_params(m_node_idx); }
+
+    size_t get_output_op_offset() const { return m_decoder->get_output_op_offset(m_node_idx); }
+
+    ov::element::Type get_output_type() const { return m_decoder->get_output_type(m_node_idx); }
+
+    std::vector<size_t> get_output_stride() const { return m_decoder->get_output_stride(m_node_idx); }
+
+    Output<Node> get_input(int idx) const override {
+        // Check if this input is a VIEW
+        size_t view_input_size = m_decoder->get_view_input_size(m_node_idx, m_input_names[idx]);
+        if (view_input_size > 0) {
+            // This is a VIEW input, get the base tensor name (last element in the chain)
+            std::string base_name =
+                m_decoder->get_view_input_src_name(m_node_idx, m_input_names[idx], view_input_size - 1);
+            // Check if the VIEW has been resolved (translate_view produced a Slice)
+            auto view_it = m_tensor_map->find(m_input_names[idx]);
+            if (!base_name.empty() && view_it != m_tensor_map->end()) {
+                auto base_it = m_tensor_map->find(base_name);
+                // A multi-output translator can publish a VIEW directly without materializing
+                // its packed parent (GatedDeltaNet attention/state). In that case the VIEW is the
+                // authoritative value. The node comparison retains the existing resolved-VIEW path.
+                if (base_it == m_tensor_map->end() ||
+                    view_it->second.get_node_shared_ptr() != base_it->second.get_node_shared_ptr()) {
+                    return view_it->second;
+                }
+                return base_it->second;
+            }
+            if (!base_name.empty()) {
+                return m_tensor_map->at(base_name);
+            }
+        }
+        // Not a VIEW or failed to get base name, use the original logic
+        return m_tensor_map->at(m_input_names[idx]);
+    }
+
+    Output<Node> get_input(const std::string & name) const override {
+        if (m_tensor_map->find(name) == m_tensor_map->end()) {
+            throw std::runtime_error("'" + name + "' not found in tensor map.");
+        }
+        return m_tensor_map->at(name);
+    }
+
+    bool has_input(const std::string & name) const { return m_tensor_map->find(name) != m_tensor_map->end(); }
+
+    void put_shared(const std::string & name, const Output<Node> & value) const {
+        m_tensor_map->insert({name, value});
+    }
+
+    const std::string & get_name() const override { return m_decoder->get_op_name(m_node_idx); }
+
+    ov::Any get_attribute_as_any(const std::string & name) const override { return m_decoder->get_attribute(name); }
+
+    int get_op_case() const { return m_decoder->get_op_case(m_node_idx); }
+
+    bool is_static() const { return m_decoder->is_static(); }
+
+    bool is_stateful() const { return m_decoder->is_stateful(); }
+
+    int get_ssm_state_size() const { return m_decoder->get_ssm_state_size(); }
+
+private:
+    std::shared_ptr<GgmlDecoder> m_decoder;
+    std::shared_ptr<TensorMap> & m_tensor_map;
+    int m_node_idx;
+    TranslateSession * m_translate_session;
+    std::vector<std::string> m_input_names;
+    std::vector<std::string> m_output_names;
+};
+
+using CreatorFunction = std::function<ov::OutputVector(const ov::frontend::ggml::NodeContext &)>;
+
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/add.cpp

@@ -0,0 +1,67 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <memory>
+#include <openvino/op/add.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/reduce_sum.hpp>
+#include <openvino/op/unsqueeze.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_add(const NodeContext & context) {
+    num_inputs_check(context, 2, 2);
+
+    if (context.get_op_case() == 1) {
+        // MoE expert-plane sum (see is_moe_expert_sum_add): input 1 is a VIEW plane of the
+        // shared base tensor `experts` = [n_embd, n_expert_used, n_tokens, 1] (ggml order) ->
+        // [1, n_tokens, n_expert_used, n_embd] (OV order). The whole ADD chain is equivalent to
+        // reducing the expert axis (OV axis 2) of that base, so bypass the chain and the
+        // per-plane Slices entirely.
+        size_t view_size = context.get_view_input_size(1);
+        auto base_name = context.get_view_input_src_name(1, view_size - 1);
+        auto base = context.get_input(base_name);
+
+        // Stateful models drop the leading batch dim, so the base is rank 3 and both axes
+        // below shift down by one. Take them from the actual rank: the expert axis is always
+        // second from last, and the token axis is re-added just before it.
+        const auto base_rank = base.get_partial_shape().rank();
+        FRONT_END_OP_CONVERSION_CHECK(base_rank.is_static() && base_rank.get_length() >= 3,
+                                      "MoE expert sum needs a static rank of at least 3");
+        const int64_t rank = base_rank.get_length();
+
+        auto reduced = std::make_shared<ov::op::v1::ReduceSum>(
+            base, ov::op::v0::Constant::create(ov::element::i64, ov::Shape{1}, {rank - 2}), false);
+        auto res = std::make_shared<ov::op::v0::Unsqueeze>(
+            reduced, ov::op::v0::Constant::create(ov::element::i64, {1}, {rank - 3}));
+        return rename_outputs_with_suffix({res}, context.get_name());
+    }
+
+    auto input_0 = process_view_input_new(context, 0);
+    auto input_1 = process_view_input_new(context, 1);
+    // opset1::Add needs matching types (e.g. fused ADD_ADD mixes f16/f32); add in f32, cast once.
+    auto output_type = context.get_output_type();
+    if (input_0.get_element_type() != input_1.get_element_type()) {
+        if (input_0.get_element_type() != ov::element::f32) {
+            input_0 = std::make_shared<ov::op::v0::Convert>(input_0, ov::element::f32);
+        }
+        if (input_1.get_element_type() != ov::element::f32) {
+            input_1 = std::make_shared<ov::op::v0::Convert>(input_1, ov::element::f32);
+        }
+    }
+    ov::Output<ov::Node> res = std::make_shared<ov::op::v1::Add>(input_0, input_1);
+    if (res.get_element_type() != output_type) {
+        res = std::make_shared<ov::op::v0::Convert>(res, output_type);
+    }
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/add_id.cpp

@@ -0,0 +1,88 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <memory>
+#include <openvino/core/node.hpp>
+#include <openvino/core/node_output.hpp>
+#include <openvino/op/add.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/gather.hpp>
+#include <openvino/op/greater_eq.hpp>
+#include <openvino/op/maximum.hpp>
+#include <openvino/op/multiply.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/shape_of.hpp>
+#include <openvino/op/unsqueeze.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+static ov::Output<ov::Node> reshape_add_id_input_to_2d(const ov::Output<ov::Node> & input,
+                                                       const ov::PartialShape & input_shape,
+                                                       const std::vector<int> & dims) {
+    const auto & actual_shape = input.get_partial_shape();
+    if (actual_shape.rank().is_static() && actual_shape.rank().get_length() == 2) {
+        return input;
+    }
+
+    if (input_shape.rank().is_static() && input_shape.rank().get_length() == 2) {
+        return input;
+    }
+
+    auto shape = std::make_shared<ov::op::v3::ShapeOf>(input, ov::element::i64);
+    return std::make_shared<ov::op::v1::Reshape>(input, get_dimensions(shape, dims), false);
+}
+
+OutputVector translate_add_id(const NodeContext & context) {
+    num_inputs_check(context, 3, 3);
+
+    auto input = process_view_input_new(context, 0);
+    auto bias = process_view_input_new(context, 1);
+    auto ids = process_view_input_new(context, 2);
+
+    // OpenVINO uses reversed GGML dimensions:
+    //   input: [1, n_token, n_used, n_embd]
+    //   bias:  [1, 1, n_expert, n_embd]
+    //   ids:   [1, 1, n_token, n_used]
+    // Model bias constants may already be stored as [n_expert, n_embd].
+    bias = reshape_add_id_input_to_2d(bias, context.get_input_shape(1), {2, 3});
+    ids = reshape_add_id_input_to_2d(ids, context.get_input_shape(2), {2, 3});
+
+    if (ids.get_element_type() != ov::element::i32 && ids.get_element_type() != ov::element::i64) {
+        ids = std::make_shared<ov::op::v0::Convert>(ids, ov::element::i32);
+    }
+
+    // an id of -1 marks a skipped slot: gather row 0 for it and zero the bias with a mask
+    auto zero = ov::op::v0::Constant::create(ids.get_element_type(), ov::Shape{}, {0});
+    ov::Output<ov::Node> keep = std::make_shared<ov::op::v1::GreaterEqual>(ids, zero);
+    keep = std::make_shared<ov::op::v0::Convert>(keep, bias.get_element_type());
+    keep = std::make_shared<ov::op::v0::Unsqueeze>(keep, ov::op::v0::Constant::create(ov::element::i64, ov::Shape{1}, {-1}));
+    ids = std::make_shared<ov::op::v1::Maximum>(ids, zero);
+
+    auto gather_axis = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{}, {0});
+    ov::Output<ov::Node> selected_bias = std::make_shared<ov::op::v8::Gather>(bias, ids, gather_axis);
+    selected_bias = std::make_shared<ov::op::v1::Multiply>(selected_bias, keep);
+    selected_bias = std::make_shared<ov::op::v1::Reshape>(
+        selected_bias, std::make_shared<ov::op::v3::ShapeOf>(input, ov::element::i64), false);
+
+    if (selected_bias.get_element_type() != input.get_element_type()) {
+        selected_bias = std::make_shared<ov::op::v0::Convert>(selected_bias, input.get_element_type());
+    }
+
+    ov::Output<ov::Node> res = std::make_shared<ov::op::v1::Add>(input, selected_bias);
+    const auto output_type = context.get_output_type();
+    if (res.get_element_type() != output_type) {
+        res = std::make_shared<ov::op::v0::Convert>(res, output_type);
+    }
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/argsort.cpp

@@ -0,0 +1,51 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+#include "ggml.h"
+
+#include <openvino/frontend/exception.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/squeeze.hpp>
+#include <openvino/op/topk.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_argsort(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+
+    auto input = process_view_input_new(context, 0);
+
+    const int32_t order = context.get_output_op_params()[0];
+
+    ov::op::v11::TopK::Mode mode;
+    switch (order) {
+    case GGML_SORT_ORDER_ASC:
+        mode = ov::op::v11::TopK::Mode::MIN;
+        break;
+    case GGML_SORT_ORDER_DESC:
+        mode = ov::op::v11::TopK::Mode::MAX;
+        break;
+    default:
+        FRONT_END_OP_CONVERSION_CHECK(false, "Unsupported GGML_OP_ARGSORT order: ", order);
+    }
+
+    // Stateful models drop the leading size-1 batch dim, so the expert axis is 2 there
+    // instead of 3 (same rank-3-vs-rank-4 split as get_rows.cpp / process_view_input).
+    const int axis = (context.is_stateful() && input.get_partial_shape().rank() == 3) ? 2 : 3;
+
+    auto k = std::make_shared<ov::op::v0::Squeeze>(get_dimensions(input.get_node_shared_ptr(), {axis}),
+                                                   ov::op::v0::Constant::create(ov::element::i64, {1}, {0}));
+
+    auto topk = std::make_shared<ov::op::v11::TopK>(input, k, axis, mode, ov::op::v11::TopK::SortType::SORT_VALUES,
+                                                    context.get_output_type(), false);
+
+    return rename_outputs_with_suffix({topk->output(1)}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/clamp.cpp

@@ -0,0 +1,33 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <cstring>
+#include <openvino/op/clamp.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_clamp(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+
+    auto input = process_view_input_new(context, 0);
+
+    const int32_t * op_params = context.get_output_op_params();
+    FRONT_END_CHECK_IMPLEMENTED(op_params != nullptr, "CLAMP requires output op params");
+
+    float min;
+    float max;
+    std::memcpy(&min, reinterpret_cast<const float *>(op_params) + 0, sizeof(float));
+    std::memcpy(&max, reinterpret_cast<const float *>(op_params) + 1, sizeof(float));
+
+    auto res = std::make_shared<ov::op::v0::Clamp>(input, min, max);
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/concat.cpp

@@ -0,0 +1,48 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <memory>
+#include <openvino/frontend/exception.hpp>
+#include <openvino/op/concat.hpp>
+#include <openvino/op/convert.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_concat(const NodeContext & context) {
+    num_inputs_check(context, 2, 2);
+
+    const int32_t * op_params = context.get_output_op_params();
+    FRONT_END_CHECK_IMPLEMENTED(op_params != nullptr, "CONCAT requires output op params");
+
+    const auto output_shape = context.get_output_shape();
+    FRONT_END_CHECK_IMPLEMENTED(output_shape.rank().is_static(), "CONCAT requires static output rank");
+
+    const auto rank = output_shape.rank().get_length();
+    const int32_t ggml_dim = op_params[0];
+    FRONT_END_CHECK_IMPLEMENTED(ggml_dim >= 0 && ggml_dim < rank, "CONCAT axis is out of range");
+
+    auto input_0 = process_view_input_new(context, 0);
+    auto input_1 = process_view_input_new(context, 1);
+    const auto output_type = context.get_output_type();
+
+    if (input_0.get_element_type() != output_type) {
+        input_0 = std::make_shared<ov::op::v0::Convert>(input_0, output_type);
+    }
+    if (input_1.get_element_type() != output_type) {
+        input_1 = std::make_shared<ov::op::v0::Convert>(input_1, output_type);
+    }
+
+    const auto axis = static_cast<int64_t>(rank - 1 - ggml_dim);
+    auto res = std::make_shared<ov::op::v0::Concat>(OutputVector{input_0, input_1}, axis);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/cont.cpp

@@ -0,0 +1,37 @@
+
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <memory>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/slice.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_cont(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+
+    auto src_shape = context.get_input_shape(0).to_shape();
+    auto dst_shape = context.get_output_shape().to_shape();
+
+    if (context.get_op_dynamic_dim() != -1) {
+        dst_shape[3 - context.get_op_dynamic_dim()] = -1;
+    }
+
+    auto input = process_view_input_new(context, 0);
+
+    ov::Output<Node> res;
+    res = std::make_shared<ov::op::v1::Reshape>(
+        input, ov::op::v0::Constant::create(ov::element::i64, {dst_shape.size()}, dst_shape), false);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/conv.cpp

@@ -0,0 +1,233 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <openvino/op/constant.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/convolution.hpp>
+#include <openvino/op/group_conv.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/unsqueeze.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_conv_2d(const NodeContext & context) {
+    num_inputs_check(context, 2, 2);
+
+    ov::Output<Node> kernel = process_view_input_new(context, 0);
+    ov::Output<Node> input  = process_view_input_new(context, 1);
+
+    if (kernel.get_element_type() != input.get_element_type()) {
+        kernel = std::make_shared<ov::op::v0::Convert>(kernel, input.get_element_type());
+    }
+
+    const int32_t * params = context.get_output_op_params();
+    const int s0 = params[0];
+    const int s1 = params[1];
+    const int p0 = params[2];
+    const int p1 = params[3];
+    const int d0 = params[4];
+    const int d1 = params[5];
+
+    ov::Strides strides{static_cast<size_t>(s1), static_cast<size_t>(s0)};
+    ov::CoordinateDiff pads_begin{static_cast<ptrdiff_t>(p1), static_cast<ptrdiff_t>(p0)};
+    ov::CoordinateDiff pads_end{static_cast<ptrdiff_t>(p1), static_cast<ptrdiff_t>(p0)};
+    ov::Strides dilations{static_cast<size_t>(d1), static_cast<size_t>(d0)};
+
+    ov::Output<Node> res = std::make_shared<ov::op::v1::Convolution>(
+        input, kernel, strides, pads_begin, pads_end, dilations, ov::op::PadType::EXPLICIT);
+
+    const auto output_type = context.get_output_type();
+    if (res.get_element_type() != output_type) {
+        res = std::make_shared<ov::op::v0::Convert>(res, output_type);
+    }
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+OutputVector translate_conv_2d_dw(const NodeContext & context) {
+    num_inputs_check(context, 2, 2);
+
+    ov::Output<Node> kernel = process_view_input_new(context, 0);
+    ov::Output<Node> input  = process_view_input_new(context, 1);
+
+    if (kernel.get_element_type() != input.get_element_type()) {
+        kernel = std::make_shared<ov::op::v0::Convert>(kernel, input.get_element_type());
+    }
+
+    const int32_t * params = context.get_output_op_params();
+    const int s0 = params[0];
+    const int s1 = params[1];
+    const int p0 = params[2];
+    const int p1 = params[3];
+    const int d0 = params[4];
+    const int d1 = params[5];
+
+    // Reshape kernel from [C, 1, KH, KW] to [C, 1, 1, KH, KW] for 2D GroupConvolution
+    auto unsqueeze_axis = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{1}, {1});
+    auto kernel_5d = std::make_shared<ov::op::v0::Unsqueeze>(kernel, unsqueeze_axis);
+
+    ov::Strides strides{static_cast<size_t>(s1), static_cast<size_t>(s0)};
+    ov::CoordinateDiff pads_begin{static_cast<ptrdiff_t>(p1), static_cast<ptrdiff_t>(p0)};
+    ov::CoordinateDiff pads_end{static_cast<ptrdiff_t>(p1), static_cast<ptrdiff_t>(p0)};
+    ov::Strides dilations{static_cast<size_t>(d1), static_cast<size_t>(d0)};
+
+    ov::Output<Node> res = std::make_shared<ov::op::v1::GroupConvolution>(
+        input, kernel_5d, strides, pads_begin, pads_end, dilations, ov::op::PadType::EXPLICIT);
+
+    const auto output_type = context.get_output_type();
+    if (res.get_element_type() != output_type) {
+        res = std::make_shared<ov::op::v0::Convert>(res, output_type);
+    }
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+OutputVector translate_conv_transpose_1d(const NodeContext & context) {
+    num_inputs_check(context, 2, 2);
+
+    ov::Output<Node> kernel = process_view_input_new(context, 0);
+    ov::Output<Node> input  = process_view_input_new(context, 1);
+
+    if (kernel.get_element_type() != input.get_element_type()) {
+        kernel = std::make_shared<ov::op::v0::Convert>(kernel, input.get_element_type());
+    }
+
+    const int32_t * params = context.get_output_op_params();
+    const int s0 = params[0];
+    const int p0 = params[1];
+    const int d0 = params[2];
+
+    const auto kernel_shape = context.get_input_shape(0).to_shape();  // [1, Cin, Cout, K]
+    const int64_t Cin = kernel_shape[1];
+    const int64_t Cout = kernel_shape[2];
+    const int64_t K = kernel_shape[3];
+
+    const auto input_shape = context.get_input_shape(1).to_shape();   // [1, N, Cin, L]
+    const int64_t N = input_shape[0] * input_shape[1];
+    const int64_t L = input_shape[3];
+
+    auto kernel_3d = std::make_shared<ov::op::v1::Reshape>(
+        kernel, ov::op::v0::Constant::create(ov::element::i64, {3}, {Cin, Cout, K}), false);
+    auto input_3d = std::make_shared<ov::op::v1::Reshape>(
+        input, ov::op::v0::Constant::create(ov::element::i64, {3}, {N, Cin, L}), false);
+
+    ov::Strides strides{static_cast<size_t>(s0)};
+    ov::CoordinateDiff pads_begin{static_cast<ptrdiff_t>(p0)};
+    ov::CoordinateDiff pads_end{static_cast<ptrdiff_t>(p0)};
+    ov::Strides dilations{static_cast<size_t>(d0)};
+
+    auto conv_tr = std::make_shared<ov::op::v1::ConvolutionBackpropData>(
+        input_3d, kernel_3d, strides, pads_begin, pads_end, dilations);
+
+    const auto out_shape = context.get_output_shape().to_shape();
+    auto out_shape_const = ov::op::v0::Constant::create(
+        ov::element::i64, {4}, {static_cast<int64_t>(out_shape[0]), static_cast<int64_t>(out_shape[1]),
+                                static_cast<int64_t>(out_shape[2]), static_cast<int64_t>(out_shape[3])});
+    ov::Output<Node> res = std::make_shared<ov::op::v1::Reshape>(conv_tr, out_shape_const, false);
+
+    const auto output_type = context.get_output_type();
+    if (res.get_element_type() != output_type) {
+        res = std::make_shared<ov::op::v0::Convert>(res, output_type);
+    }
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+OutputVector translate_conv_transpose_2d(const NodeContext & context) {
+    num_inputs_check(context, 2, 2);
+
+    ov::Output<Node> kernel = process_view_input_new(context, 0);
+    ov::Output<Node> input  = process_view_input_new(context, 1);
+
+    if (kernel.get_element_type() != input.get_element_type()) {
+        kernel = std::make_shared<ov::op::v0::Convert>(kernel, input.get_element_type());
+    }
+
+    const int32_t * params = context.get_output_op_params();
+    const int stride = params[0];
+
+    ov::Strides strides{static_cast<size_t>(stride), static_cast<size_t>(stride)};
+    ov::CoordinateDiff pads_begin{0, 0};
+    ov::CoordinateDiff pads_end{0, 0};
+    ov::Strides dilations{1, 1};
+
+    ov::Output<Node> res = std::make_shared<ov::op::v1::ConvolutionBackpropData>(
+        input, kernel, strides, pads_begin, pads_end, dilations);
+
+    const auto output_type = context.get_output_type();
+    if (res.get_element_type() != output_type) {
+        res = std::make_shared<ov::op::v0::Convert>(res, output_type);
+    }
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+OutputVector translate_conv_3d(const NodeContext & context) {
+    num_inputs_check(context, 2, 2);
+
+    ov::Output<Node> kernel = process_view_input_new(context, 0);
+    ov::Output<Node> input  = process_view_input_new(context, 1);
+
+    if (kernel.get_element_type() != input.get_element_type()) {
+        kernel = std::make_shared<ov::op::v0::Convert>(kernel, input.get_element_type());
+    }
+
+    const int32_t * params = context.get_output_op_params();
+    const int s0 = params[0];
+    const int s1 = params[1];
+    const int s2 = params[2];
+    const int p0 = params[3];
+    const int p1 = params[4];
+    const int p2 = params[5];
+    const int d0 = params[6];
+    const int d1 = params[7];
+    const int d2 = params[8];
+    const int c  = params[9];
+    const int n  = params[10];
+    const int oc = params[11];
+
+    const auto kshape = context.get_input_shape(0).to_shape();  // [c*oc, KD, KH, KW]
+    const int64_t KD = kshape[1];
+    const int64_t KH = kshape[2];
+    const int64_t KW = kshape[3];
+
+    const auto inshape = context.get_input_shape(1).to_shape(); // [c*n, ID, IH, IW]
+    const int64_t ID = inshape[1];
+    const int64_t IH = inshape[2];
+    const int64_t IW = inshape[3];
+
+    auto kernel_5d = std::make_shared<ov::op::v1::Reshape>(
+        kernel, ov::op::v0::Constant::create(ov::element::i64, {5}, {static_cast<int64_t>(oc), static_cast<int64_t>(c), KD, KH, KW}), false);
+    auto input_5d = std::make_shared<ov::op::v1::Reshape>(
+        input, ov::op::v0::Constant::create(ov::element::i64, {5}, {static_cast<int64_t>(n), static_cast<int64_t>(c), ID, IH, IW}), false);
+
+    ov::Strides strides{static_cast<size_t>(s2), static_cast<size_t>(s1), static_cast<size_t>(s0)};
+    ov::CoordinateDiff pads_begin{static_cast<ptrdiff_t>(p2), static_cast<ptrdiff_t>(p1), static_cast<ptrdiff_t>(p0)};
+    ov::CoordinateDiff pads_end{static_cast<ptrdiff_t>(p2), static_cast<ptrdiff_t>(p1), static_cast<ptrdiff_t>(p0)};
+    ov::Strides dilations{static_cast<size_t>(d2), static_cast<size_t>(d1), static_cast<size_t>(d0)};
+
+    auto conv = std::make_shared<ov::op::v1::Convolution>(
+        input_5d, kernel_5d, strides, pads_begin, pads_end, dilations, ov::op::PadType::EXPLICIT);
+
+    const auto out_shape = context.get_output_shape().to_shape(); // [oc*n, OD, OH, OW]
+    auto out_shape_const = ov::op::v0::Constant::create(
+        ov::element::i64, {4}, {static_cast<int64_t>(out_shape[0]), static_cast<int64_t>(out_shape[1]),
+                                static_cast<int64_t>(out_shape[2]), static_cast<int64_t>(out_shape[3])});
+    ov::Output<Node> res = std::make_shared<ov::op::v1::Reshape>(conv, out_shape_const, false);
+
+    const auto output_type = context.get_output_type();
+    if (res.get_element_type() != output_type) {
+        res = std::make_shared<ov::op::v0::Convert>(res, output_type);
+    }
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/cpy.cpp

@@ -0,0 +1,314 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <climits>
+#include <cstdint>
+#include <cstdio>
+#include <memory>
+#include <numeric>
+#include <openvino/frontend/exception.hpp>
+#include <openvino/op/add.hpp>
+#include <openvino/op/concat.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/gather.hpp>
+#include <openvino/op/multiply.hpp>
+#include <openvino/op/negative.hpp>
+#include <openvino/op/range.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/scatter_update.hpp>
+#include <openvino/op/shape_of.hpp>
+#include <openvino/op/slice.hpp>
+#include <openvino/op/squeeze.hpp>
+#include <openvino/op/subtract.hpp>
+#include <vector>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_cpy(const NodeContext & context) {
+    auto op_case = context.get_op_case();
+    auto input_shape = context.get_input_shape(0);
+    auto output_shape = context.get_input_shape(1);
+
+    if (op_case == 4) {
+        auto src = process_view_input_new(context, 0);
+        auto base = context.get_input(1);
+
+        int64_t n_elems = 1;
+        for (const auto & dim : context.get_output_shape().to_shape()) {
+            n_elems *= static_cast<int64_t>(dim);
+        }
+
+        const auto output_stride = context.get_output_stride();
+        const size_t elem_size = output_stride.empty() ? context.get_output_type().size() : output_stride.back();
+        FRONT_END_OP_CONVERSION_CHECK(elem_size > 0, "CPY conv state view update has invalid element size");
+
+        const int64_t begin_val = static_cast<int64_t>(context.get_output_op_offset() / elem_size);
+        const int64_t end_val = begin_val + n_elems;
+
+        auto flat_shape = ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{1, 1, 1, -1});
+        src = std::make_shared<ov::op::v1::Reshape>(src, flat_shape, false);
+        if (src.get_element_type() != context.get_output_type()) {
+            src = std::make_shared<ov::op::v0::Convert>(src, context.get_output_type());
+        }
+
+        auto zero = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
+        auto begin = ov::op::v0::Constant::create(ov::element::i64, {1}, {begin_val});
+        auto end = ov::op::v0::Constant::create(ov::element::i64, {1}, {end_val});
+        auto int_max = ov::op::v0::Constant::create(ov::element::i64, {1}, {INT_MAX});
+        auto one = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+        auto axis = ov::op::v0::Constant::create(ov::element::i64, {1}, {3});
+
+        auto head_part = std::make_shared<ov::op::v8::Slice>(base, zero, begin, one, axis);
+        auto tail_part = std::make_shared<ov::op::v8::Slice>(base, end, int_max, one, axis);
+        auto res = std::make_shared<ov::op::v0::Concat>(ov::OutputVector{head_part, src, tail_part}, 3);
+        return rename_outputs_with_suffix({res}, context.get_name());
+    }
+
+    // Recurrent state cache writeback into a slot block of the cache. Where the block starts is a
+    // runtime input, so the cached model works for any kv head and active sequence count. The
+    // result is the full updated cache.
+    // op_case 1: gated-delta-net state, op_case 2: conv state, op_case 3: defrag remainder.
+    if (op_case == 3) {
+        // With -np 1 (and generally whenever there is no defrag remainder) this GET_ROWS gathers
+        // zero rows: nothing to write back, and the cache is unchanged. NPU rejects zero-size
+        // tensors, so short-circuit instead of building a degenerate Slice/Concat chain.
+        bool is_empty = false;
+        if (input_shape.rank().is_static()) {
+            for (const auto & d : input_shape) {
+                if (d.is_static() && d.get_length() == 0) {
+                    is_empty = true;
+                    break;
+                }
+            }
+        }
+        if (is_empty) {
+            return {context.get_input(1)};
+        }
+    }
+    // op_case 7/8/9 are the single-slot variants; op_case 10 writes native GDN state into a
+    // multi-slot cache without rollback snapshots.
+    const bool single_slot_assign = op_case >= 7 && op_case <= 9;
+    const bool direct_gdn_state = op_case == 7 || op_case == 10;
+    int writeback_case = op_case;
+    if (op_case == 10) {
+        writeback_case = 1;
+    } else if (single_slot_assign) {
+        writeback_case = op_case - 6;
+    }
+    const std::string slot_begin_name = "rs_slot_begin_" + context.get_name();
+    const bool slice_assign = writeback_case >= 1 && writeback_case <= 3 &&
+                              (single_slot_assign || context.has_input(slot_begin_name));
+    if (slice_assign) {
+        if (single_slot_assign && writeback_case == 3) {
+            return {context.get_input(1)};
+        }
+
+        const int64_t slot_axis = 2;
+        auto zero = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
+        auto one = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+        auto int_max = ov::op::v0::Constant::create(ov::element::i64, {1}, {INT_MAX});
+        auto axis = ov::op::v0::Constant::create(ov::element::i64, {1}, {slot_axis});
+        auto feature = ov::op::v0::Constant::create(ov::element::i64, {4},
+                                                    std::vector<int64_t>{1, 1, -1, output_shape[3].get_length()});
+
+        ov::Output<ov::Node> src;
+        ov::Output<ov::Node> begin;
+        if (!single_slot_assign) {
+            begin = context.get_input(slot_begin_name);
+        }
+        auto base = context.get_input(1);
+        if (writeback_case == 1) {
+            if (direct_gdn_state) {
+                // Non-rollback GDN publishes state directly as [active_slots, heads, value_dim,
+                // key_dim]. Flatten each active slot before replacing or updating the cache.
+                src = std::make_shared<ov::op::v1::Reshape>(context.get_input(0), feature, false);
+            } else {
+                // Multi-slot rollback still consumes GGML's packed [attention | state snapshots]
+                // layout. Slice the state block using the runtime source offset.
+                auto src_begin = context.get_input("rs_src_begin_" + context.get_name());
+                auto state_part =
+                    std::make_shared<ov::op::v8::Slice>(context.get_input(0), src_begin, int_max, one, axis);
+                src = std::make_shared<ov::op::v1::Reshape>(state_part, feature, false);
+            }
+        } else if (writeback_case == 2) {
+            // conv_input is [previous conv state | new tokens]; the snapshot is the conv_kernel_size - 1
+            // columns ending at the last *valid* token. Gather (rather than Slice) keeps the output
+            // shape static even though the window start is a runtime value.
+            auto window_size = (int64_t) input_shape[3].get_length();
+            ov::Output<ov::Node> window;
+            auto col_axis = ov::op::v0::Constant::create(ov::element::i64, {1}, {3});
+            const std::string src_begin_name = "rs_src_begin_" + context.get_name();
+            if (context.has_input(src_begin_name)) {
+                auto src_begin = context.get_input(src_begin_name);
+                auto src_end = std::make_shared<ov::op::v1::Add>(
+                    src_begin, ov::op::v0::Constant::create(ov::element::i64, {1}, {window_size}));
+                window = std::make_shared<ov::op::v8::Slice>(context.get_input(0), src_begin, src_end, one, col_axis);
+            } else if (context.has_input("chunk_valid_len")) {
+                std::vector<int64_t> offsets(window_size);
+                std::iota(offsets.begin(), offsets.end(), 0);
+                auto indices = std::make_shared<ov::op::v1::Add>(
+                    ov::op::v0::Constant::create(ov::element::i64, {(size_t) window_size}, offsets),
+                    context.get_input("chunk_valid_len"));
+                window = std::make_shared<ov::op::v8::Gather>(context.get_input(0), indices, col_axis);
+            } else {
+                auto window_begin = ov::op::v0::Constant::create(ov::element::i64, {1}, {-window_size});
+                window =
+                    std::make_shared<ov::op::v8::Slice>(context.get_input(0), window_begin, int_max, one, col_axis);
+            }
+            const auto base_shape = base.get_partial_shape();
+            FRONT_END_OP_CONVERSION_CHECK(base_shape.rank().is_static() && base_shape.rank().get_length() == 4,
+                                          "CPY conv state cache update requires rank-4 base cache");
+            FRONT_END_OP_CONVERSION_CHECK(base_shape[3].is_static(),
+                                          "CPY conv state cache update requires static feature size");
+            FRONT_END_OP_CONVERSION_CHECK(input_shape.rank().is_static() && input_shape.rank().get_length() == 4 &&
+                                              input_shape[2].is_static() && input_shape[3].is_static(),
+                                          "CPY conv state cache update requires static source feature view");
+
+            const int64_t full_feature_size = base_shape[3].get_length();
+            const int64_t update_feature_size = input_shape[2].get_length() * input_shape[3].get_length();
+            const auto output_stride = context.get_output_stride();
+            const size_t elem_size = output_stride.empty() ? context.get_output_type().size() : output_stride.back();
+            FRONT_END_OP_CONVERSION_CHECK(elem_size > 0,
+                                          "CPY conv state cache update has invalid element size");
+            const int64_t feature_begin = static_cast<int64_t>(context.get_output_op_offset() / elem_size) %
+                                          full_feature_size;
+            const int64_t feature_end = feature_begin + update_feature_size;
+            FRONT_END_OP_CONVERSION_CHECK(feature_begin >= 0 && feature_end <= full_feature_size,
+                                          "CPY conv state cache update feature range is out of bounds");
+
+            auto partial_feature = ov::op::v0::Constant::create(
+                ov::element::i64, {4}, std::vector<int64_t>{1, 1, -1, update_feature_size});
+            src = std::make_shared<ov::op::v1::Reshape>(window, partial_feature, false);
+            if (src.get_element_type() != context.get_output_type()) {
+                src = std::make_shared<ov::op::v0::Convert>(src, context.get_output_type());
+            }
+
+            if (single_slot_assign) {
+                return rename_outputs_with_suffix({src}, context.get_name());
+            }
+
+            auto src_len = std::make_shared<ov::op::v8::Gather>(
+                std::make_shared<ov::op::v3::ShapeOf>(src, ov::element::i64), axis,
+                ov::op::v0::Constant::create(ov::element::i64, {}, {0}));
+            auto slot_end = std::make_shared<ov::op::v1::Add>(begin, src_len);
+            auto active_slots = std::make_shared<ov::op::v8::Slice>(base, begin, slot_end, one, axis);
+
+            auto feature_axis = ov::op::v0::Constant::create(ov::element::i64, {1}, {3});
+            auto feature_begin_node = ov::op::v0::Constant::create(ov::element::i64, {1}, {feature_begin});
+            auto feature_end_node = ov::op::v0::Constant::create(ov::element::i64, {1}, {feature_end});
+            auto feature_head = std::make_shared<ov::op::v8::Slice>(active_slots, zero, feature_begin_node, one,
+                                                                    feature_axis);
+            auto feature_tail = std::make_shared<ov::op::v8::Slice>(active_slots, feature_end_node, int_max, one,
+                                                                    feature_axis);
+            src = std::make_shared<ov::op::v0::Concat>(ov::OutputVector{feature_head, src, feature_tail}, 3);
+        } else {
+            // op_case 3: gathered remainder rows already have the cache slot layout [1, 1, extra, feature]
+            src = context.get_input(0);
+        }
+
+        if (src.get_element_type() != context.get_output_type()) {
+            src = std::make_shared<ov::op::v0::Convert>(src, context.get_output_type());
+        }
+
+        if (single_slot_assign) {
+            return rename_outputs_with_suffix({src}, context.get_name());
+        }
+
+        auto src_len =
+            std::make_shared<ov::op::v8::Gather>(std::make_shared<ov::op::v3::ShapeOf>(src, ov::element::i64), axis,
+                                                 ov::op::v0::Constant::create(ov::element::i64, {}, {0}));
+        auto end = std::make_shared<ov::op::v1::Add>(begin, src_len);
+        auto head_part = std::make_shared<ov::op::v8::Slice>(base, zero, begin, one, axis);
+        auto tail_part = std::make_shared<ov::op::v8::Slice>(base, end, int_max, one, axis);
+        auto res = std::make_shared<ov::op::v0::Concat>(ov::OutputVector{head_part, src, tail_part}, slot_axis);
+        return rename_outputs_with_suffix({res}, context.get_name());
+    }
+
+    auto input = process_view_input_new(context, 0);
+
+    if (op_case == 5 || op_case == 6) {
+        auto input_shape = context.get_input_shape(0);
+        auto output_shape = context.get_output_shape();
+        auto dst_ggml_shape = context.get_view_input_ggml_shape(1, 0);
+        auto dst_stride = context.get_view_input_stride(1, 0);
+        size_t offset_bytes = context.get_view_input_offset(1, 0);
+        auto n_state = (int64_t) context.get_input_shape(0)[3].get_length();
+        auto n_state_c = ov::op::v0::Constant::create(ov::element::i64, {1}, {n_state});
+        auto kv_buf = context.get_input(1);  // shape {1,1,1,N}
+
+        Output<Node> token_len_per_seq;
+        Output<Node> n_write_dyn;
+        if (context.has_input("token_len_per_seq")) {
+            token_len_per_seq = context.get_input("token_len_per_seq");
+            n_write_dyn = std::make_shared<ov::op::v1::Multiply>(token_len_per_seq, n_state_c);
+        } else {
+            n_write_dyn = ov::op::v0::Constant::create(ov::element::i64, {1}, {(int64_t) dst_ggml_shape[3]});
+        }
+        size_t elem_size = dst_stride[3];
+        FRONT_END_OP_CONVERSION_CHECK(elem_size > 0, "CPY KV cache view update has invalid element size");
+        int64_t start_elem = (int64_t) (offset_bytes / elem_size);
+        // op_case 5: decoder self-attention – write offset advances each step.
+        // op_case 6: encoder self-attn or cross-attn – offset fixed at compile time.
+        const bool is_decoder_self_attn = (op_case == 5);
+        auto ones_c = ov::op::v0::Constant::create(ov::element::i64, {3}, std::vector<int64_t>{1, 1, 1});
+        auto new_shape = std::make_shared<ov::op::v0::Concat>(ov::OutputVector{ones_c, n_write_dyn}, 0);
+
+        auto reshaped = std::make_shared<ov::op::v1::Reshape>(input, new_shape, false);
+        auto data = std::make_shared<ov::op::v0::Convert>(reshaped, context.get_output_type());
+        // Indices [start_elem .. start_elem + n_write) on axis 3 of {1,1,1,N}
+        // For decoder self-attention the write offset advances each step, so compute it
+        // dynamically from the model inputs: start = (attention_size - token_len_per_seq) * n_state.
+        // For encoder self-attn and cross-attn the offset is fixed at graph-compile time.
+        ov::Output<ov::Node> start;
+        if (is_decoder_self_attn && context.has_input("attention_size") && context.has_input("token_len_per_seq")) {
+            auto attention_size_in = context.get_input("attention_size");
+            auto token_len_in = context.get_input("token_len_per_seq");
+            auto past_tokens = std::make_shared<ov::op::v1::Subtract>(attention_size_in, token_len_in);
+            auto new_start = std::make_shared<ov::op::v1::Multiply>(past_tokens, n_state_c);
+            start = std::make_shared<ov::op::v1::Add>(
+                new_start, ov::op::v0::Constant::create(ov::element::i64, {1}, {start_elem}));
+        } else {
+            start = ov::op::v0::Constant::create(ov::element::i64, {1}, {start_elem});
+        }
+        auto start_squeezed = std::make_shared<ov::op::v0::Squeeze>(start);
+        auto end = std::make_shared<ov::op::v1::Add>(start_squeezed, n_write_dyn);
+        auto end_squeezed = std::make_shared<ov::op::v0::Squeeze>(end);
+        auto step = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+        auto step_squeezed = std::make_shared<ov::op::v0::Squeeze>(step);
+        auto indices =
+            std::make_shared<ov::op::v4::Range>(start_squeezed, end_squeezed, step_squeezed, ov::element::i64);
+        auto axis = ov::op::v0::Constant::create(ov::element::i64, {1}, {3});
+
+        auto kv_updated = std::make_shared<ov::op::v3::ScatterUpdate>(kv_buf, indices, data, axis);
+        return rename_outputs_with_suffix({kv_updated}, context.get_name());
+    }
+
+    if (input_shape != output_shape) {
+        auto new_shape = ov::op::v0::Constant::create(
+            ov::element::i64, {static_cast<size_t>(output_shape.rank().get_length())}, output_shape.to_shape());
+        input = std::make_shared<ov::op::v1::Reshape>(input, new_shape, false);
+    }
+
+    ov::Output<Node> res;
+    if (context.get_input_type(0) != context.get_output_type()) {
+        res = std::make_shared<ov::op::v0::Convert>(input, context.get_output_type());
+    } else {
+        res = input;
+    }
+
+    if (res.get_node_shared_ptr() == context.get_input(0).get_node_shared_ptr()) {
+        return {res};
+    }
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/cumsum.cpp

@@ -0,0 +1,29 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <openvino/op/constant.hpp>
+#include <openvino/op/cum_sum.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+// GGML cumsum computes prefix sum along dim 0 (the innermost/fastest dimension).
+// In OV layout the dims are reversed: ggml [ne0, ne1, ne2, ne3] → OV [ne3, ne2, ne1, ne0],
+// so ggml dim 0 maps to OV axis 3 (last axis).
+OutputVector translate_cumsum(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+
+    auto x    = context.get_input(0);
+    auto axis = ov::op::v0::Constant::create(ov::element::i64, {}, {3});
+    auto res  = std::make_shared<ov::op::v0::CumSum>(x, axis);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/diag.cpp

@@ -0,0 +1,37 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <openvino/op/constant.hpp>
+#include <openvino/op/eye.hpp>
+#include <openvino/op/multiply.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+// GGML DIAG takes a 1D vector (ne0, 1, ne2, ne3) and produces a diagonal matrix
+// of shape (ne0, ne0, ne2, ne3).
+// In OV layout (ggml [ne0, ne1, ne2, ne3] → OV [ne3, ne2, ne1, ne0]):
+//   input:  [ne3, ne2, 1, ne0]
+//   output: [ne3, ne2, ne0, ne0]
+// The diagonal: output[..., i, j] = input[..., 0, j] if i == j, else 0.
+OutputVector translate_diag(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+
+    auto x = process_view_input_new(context, 0);  // OV shape: [ne3, ne2, 1, ne0]
+
+    auto n = get_dimensions(x.get_node_shared_ptr(), {3});
+    auto zero_diag = ov::op::v0::Constant::create(ov::element::i64, {}, {0});
+
+    auto eye = std::make_shared<ov::op::v9::Eye>(n, n, zero_diag, x.get_element_type());
+    auto res = std::make_shared<ov::op::v1::Multiply>(x, eye);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/div.cpp

@@ -0,0 +1,138 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+#include "ggml.h"
+
+#include <memory>
+#include <openvino/core/type.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/divide.hpp>
+#include <openvino/op/multiply.hpp>
+#include <openvino/op/shape_of.hpp>
+#include <openvino/op/sigmoid.hpp>
+#include <openvino/op/swish.hpp>
+#include <openvino/op/tile.hpp>
+#include <openvino/op/util/precision_sensitive_attribute.hpp>
+#include <vector>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+namespace {
+
+bool is_silu_div_pattern(const ov::Output<ov::Node> & numerator,
+                         const ov::Output<ov::Node> & denominator,
+                         const NodeContext & context) {
+    if (context.get_input_size() != 2) {
+        return false;
+    }
+
+    const auto * unary_op = reinterpret_cast<const ggml_unary_op *>(context.get_input_op_params(0));
+    if (unary_op == nullptr || *unary_op != GGML_UNARY_OP_SILU) {
+        return false;
+    }
+
+    const auto denom_node = denominator.get_node_shared_ptr();
+
+    if (auto swish = ov::as_type_ptr<ov::op::v4::Swish>(numerator.get_node_shared_ptr())) {
+        return swish->input_value(0).get_node_shared_ptr() == denom_node;
+    }
+    return false;
+}
+
+ov::Output<ov::Node> repeat_input_to_match(const NodeContext & context,
+                                           const ov::Output<ov::Node> & input,
+                                           const ov::Output<ov::Node> & target,
+                                           size_t input_index) {
+    const auto input_shape = context.get_input_shape(input_index);
+    const auto target_shape = context.get_input_shape(0);
+
+    if (input_shape == target_shape) {
+        return input;
+    }
+
+    if (input_shape.rank().is_static() && target_shape.rank().is_static()) {
+        const auto rank = static_cast<size_t>(input_shape.rank().get_length());
+        std::vector<int64_t> repeats(rank, 1);
+        bool needs_repeat = false;
+
+        for (size_t axis = 0; axis < rank; ++axis) {
+            FRONT_END_OP_CONVERSION_CHECK(input_shape[axis].is_static() && target_shape[axis].is_static(),
+                                          "DIV repeat requires static dimensions on both inputs");
+
+            const int64_t input_dim = input_shape[axis].get_length();
+            const int64_t target_dim = target_shape[axis].get_length();
+
+            FRONT_END_OP_CONVERSION_CHECK(input_dim > 0 && target_dim > 0 && target_dim % input_dim == 0,
+                                          "DIV input shape ", input_shape, " cannot repeat to match ", target_shape);
+
+            repeats[axis] = target_dim / input_dim;
+            needs_repeat = needs_repeat || repeats[axis] != 1;
+        }
+
+        if (!needs_repeat) {
+            return input;
+        }
+
+        auto repeats_node = ov::op::v0::Constant::create(ov::element::i64, {repeats.size()}, repeats);
+        return std::make_shared<ov::op::v0::Tile>(input, repeats_node);
+    }
+
+    auto input_shape_node = std::make_shared<ov::op::v3::ShapeOf>(input, ov::element::i64);
+    auto target_shape_node = std::make_shared<ov::op::v3::ShapeOf>(target, ov::element::i64);
+    auto repeats_node = std::make_shared<ov::op::v1::Divide>(target_shape_node, input_shape_node);
+    return std::make_shared<ov::op::v0::Tile>(input, repeats_node);
+}
+
+}  // namespace
+
+OutputVector translate_div(const NodeContext & context) {
+    num_inputs_check(context, 2, 2);
+
+    auto input_0 = process_view_input_new(context, 0);
+    auto input_1 = process_view_input_new(context, 1);
+
+    if (is_silu_div_pattern(input_0, input_1, context)) {
+        ov::Output<ov::Node> res = std::make_shared<ov::op::v0::Sigmoid>(input_1);
+        if (res.get_element_type() != context.get_output_type()) {
+            res = std::make_shared<ov::op::v0::Convert>(res, context.get_output_type());
+        }
+        return rename_outputs_with_suffix({res}, context.get_name());
+    }
+
+    input_1 = repeat_input_to_match(context, input_1, input_0, 1);
+
+    const auto output_type = context.get_output_type();
+    const bool use_f32_compute = input_0.get_element_type() != ov::element::f32 ||
+                                 input_1.get_element_type() != ov::element::f32 || output_type != ov::element::f32;
+
+    if (use_f32_compute) {
+        input_0 = std::make_shared<ov::op::v0::Convert>(input_0, ov::element::f32);
+        input_1 = std::make_shared<ov::op::v0::Convert>(input_1, ov::element::f32);
+    }
+
+    ov::Output<ov::Node> res = std::make_shared<ov::op::v1::Divide>(input_0, input_1);
+    if (use_f32_compute) {
+        // Keep the reciprocal/divide path in FP32. Without this hint, the GPU
+        // plugin can still compress the subgraph back to FP16 and overflow on
+        // small shexp gate values (e.g. silu(x) / x in qwen2moe).
+        ov::mark_as_precision_sensitive(res.get_node_shared_ptr()->input(0));
+        ov::mark_as_precision_sensitive(res.get_node_shared_ptr()->input(1));
+    }
+    if (res.get_element_type() != output_type) {
+        auto output_convert = std::make_shared<ov::op::v0::Convert>(res, output_type);
+        if (use_f32_compute) {
+            ov::mark_as_precision_sensitive(output_convert->input(0));
+        }
+        res = output_convert;
+    }
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/fill.cpp

@@ -0,0 +1,34 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <openvino/op/broadcast.hpp>
+#include <openvino/op/constant.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+// GGML FILL sets all elements of a tensor to a constant value.
+// The constant is stored as a float in op_params[0].
+OutputVector translate_fill(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+
+    float c;
+    memcpy(&c, context.get_output_op_params(), sizeof(float));
+
+    auto shape = context.get_input_shape(0).to_shape();
+
+    auto val = ov::op::v0::Constant::create(context.get_output_type(), {}, {c});
+    auto target_shape = ov::op::v0::Constant::create(ov::element::i64, {shape.size()},
+        std::vector<int64_t>(shape.begin(), shape.end()));
+    auto res = std::make_shared<ov::op::v3::Broadcast>(val, target_shape);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/flash_attn_ext.cpp

@@ -0,0 +1,242 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+#include "ggml-openvino/ggml-openvino-extra.h"
+
+#include <cstddef>
+#include <cstdint>
+#include <memory>
+#include <openvino/op/add.hpp>
+#include <openvino/op/broadcast.hpp>
+#include <openvino/op/concat.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/matmul.hpp>
+#include <openvino/op/multiply.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/scaled_dot_product_attention.hpp>
+#include <openvino/op/slice.hpp>
+#include <openvino/op/softmax.hpp>
+#include <openvino/op/transpose.hpp>
+#include <openvino/op/unsqueeze.hpp>
+#include <string>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+static ov::Output<ov::Node> reshape_flat_kv(const ov::Output<ov::Node> & kv_flat,
+                                            size_t view_offset_bytes,
+                                            size_t nb1_bytes,
+                                            int64_t n_head,
+                                            int64_t head_size,
+                                            const ov::Output<ov::Node> & attention_size) {
+    int64_t n_state = n_head * head_size;
+    int64_t layer_start_elem = (int64_t) (view_offset_bytes / (nb1_bytes / n_state));
+    // Dynamic slice: [layer_start_elem, layer_start_elem + n_kv * n_state)
+    auto start_c = ov::op::v0::Constant::create(ov::element::i64, {1}, {layer_start_elem});
+    auto n_state_c = ov::op::v0::Constant::create(ov::element::i64, {1}, {n_state});
+    // end = start + attention_size * n_state  (both static + dynamic)
+    auto kv_len_elems = std::make_shared<ov::op::v1::Multiply>(attention_size, n_state_c);
+    auto end_c = std::make_shared<ov::op::v1::Add>(start_c, kv_len_elems);
+    auto step_c = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+    auto axis_c = ov::op::v0::Constant::create(ov::element::i64, {1}, {3});
+    auto sliced = std::make_shared<ov::op::v8::Slice>(kv_flat, start_c, end_c, step_c, axis_c);
+
+    // KV cache is laid out as {n_kv, n_head, head_size} in memory
+    // Reshape to {1, n_kv, n_head, head_size}, then transpose to {1, n_head, n_kv, head_size}
+    // as required by SDPA.
+    auto one_c = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+    auto n_head_c = ov::op::v0::Constant::create(ov::element::i64, {1}, {n_head});
+    auto head_size_c = ov::op::v0::Constant::create(ov::element::i64, {1}, {head_size});
+    // reshape: {n_kv*n_state} -> {1, n_kv, n_head, head_size}
+    auto new_shape =
+        std::make_shared<ov::op::v0::Concat>(ov::OutputVector{one_c, attention_size, n_head_c, head_size_c}, 0);
+    auto reshaped = std::make_shared<ov::op::v1::Reshape>(sliced, new_shape, false);
+    // transpose: {1, n_kv, n_head, head_size} -> {1, n_head, n_kv, head_size}
+    auto perm = ov::op::v0::Constant::create(ov::element::i64, {4}, {0, 2, 1, 3});
+    auto ret = std::make_shared<ov::op::v1::Transpose>(reshaped, perm);
+    return ret;
+}
+
+OutputVector translate_flash_attn_ext(const NodeContext & context) {
+    num_inputs_check(context, 3, 4);
+    const bool has_mask = context.get_input_size() == 4;
+    auto q_f32 = context.get_input(0);
+    auto k = context.get_input(1);
+    auto v = context.get_input(2);
+    const int op_case = context.get_op_case();
+
+    if (op_case == 1 || op_case == 2) {
+        int64_t n_state_head = (int64_t) context.get_view_input_ggml_shape(1, 0)[3];
+        int64_t n_head = (int64_t) context.get_view_input_ggml_shape(1, 0)[1];
+        size_t nb1 = context.get_view_input_stride(1, 0)[2];
+        size_t offset = context.get_view_input_offset(1, 0);
+        ov::Output<ov::Node> attention_size;
+        if (op_case == 1) {
+            attention_size = context.get_input("attention_size");
+        } else {
+            attention_size = context.get_input("attention_size_static");
+        }
+        k = reshape_flat_kv(k, offset, nb1, n_head, n_state_head, attention_size);
+        v = reshape_flat_kv(v, offset, nb1, n_head, n_state_head, attention_size);
+    }
+
+    float * params = reinterpret_cast<float *>(context.get_output_op_params());
+    float scale = params[0];
+    // float max_bias      = params[1];
+    // float logit_softcap = params[2];
+
+    auto q = std::make_shared<ov::op::v0::Convert>(q_f32, ov::element::f16);
+    auto scale_node = std::make_shared<ov::op::v0::Constant>(ov::element::f16, ov::Shape{}, std::vector<float>{scale});
+
+    ov::Output<ov::Node> res;
+
+    // For stateful
+    ov::Output<ov::Node> mask;
+    if (has_mask) {
+        mask = context.get_input(3);
+        std::string mask_name = "KQ_mask_sliced";
+        if (context.get_input_names()[3].find("swa") != std::string::npos) {
+            mask_name = "KQ_mask_swa_sliced";
+        }
+        if (context.has_input(mask_name)) {
+            mask = context.get_input(mask_name);
+        }
+        if (mask.get_element_type() != ov::element::f16) {
+            mask = std::make_shared<ov::op::v0::Convert>(mask, ov::element::f16);
+        }
+    }
+
+    //auto tile_kv = [&](int64_t num_heads, int64_t num_heads_kv, int64_t head_size, ov::Output<Node> kv) {
+    //    int64_t factor = num_heads / num_heads_kv;
+    //    if (factor > 1 && num_heads_kv > 1) {
+    auto q_shape = context.get_input_shape(0).to_shape();
+    auto k_shape = context.get_input_shape(1).to_shape();
+    const int64_t num_heads = q_shape[1];
+    const int64_t num_heads_kv = k_shape[1];
+    const int64_t head_size = q_shape[3];
+    const int64_t factor = num_heads / num_heads_kv;
+
+    // Manual GQA attention: enabled by default on GPU in stateless mode.
+    // Set GGML_OPENVINO_MANUAL_GQA_ATTN to a positive value (e.g. 1) to force-enable,
+    // or to 0 to force-disable. Unset falls back to the device-based default.
+    static const bool manual_gqa_enabled = []() {
+        const char * env = ggml_openvino_getenv_str("GGML_OPENVINO_MANUAL_GQA_ATTN");
+        if (env != nullptr) {
+            return ggml_openvino_getenv_int("GGML_OPENVINO_MANUAL_GQA_ATTN") > 0;
+        }
+        return ggml_openvino_is_gpu();
+    }();
+    const bool use_manual_gqa_attention =
+        manual_gqa_enabled && factor > 1 && num_heads_kv > 1 && !context.is_stateful();
+
+    if (use_manual_gqa_attention) {
+        // Q, K, V arrive as [B, n_heads(_kv), S, head_size], where B is the active
+        // batch (n_seq_active) and may be > 1 (llama-perplexity, llama-server -np > 1)
+        // or dynamic. Reshape to
+        //   K_r: [B, num_heads_kv, 1, S, head_size]
+        //   Q_r: [B, num_heads_kv, factor, S_q, head_size]
+        // and let MatMul broadcast across the factor dim without materialising
+        // an expanded K/V. The leading 0 + special_zero=true copies B at runtime,
+        // so this is correct for B == 1, B > 1, and dynamic B alike. Only the head
+        // dims and head_size are baked in as literals; the sequence dim stays -1.
+        auto k_5d_shape = ov::op::v0::Constant::create(ov::element::i64, {5},
+                                                       std::vector<int64_t>{0, num_heads_kv, 1, -1, head_size});
+        auto v_5d_shape = ov::op::v0::Constant::create(ov::element::i64, {5},
+                                                       std::vector<int64_t>{0, num_heads_kv, 1, -1, head_size});
+        auto q_5d_shape = ov::op::v0::Constant::create(ov::element::i64, {5},
+                                                       std::vector<int64_t>{0, num_heads_kv, factor, -1, head_size});
+
+        auto k_r = std::make_shared<ov::op::v1::Reshape>(k, k_5d_shape, true);
+        auto v_r = std::make_shared<ov::op::v1::Reshape>(v, v_5d_shape, true);
+        auto q_r = std::make_shared<ov::op::v1::Reshape>(q, q_5d_shape, true);
+
+        // QK^T → [B, num_heads_kv, factor, S_q, S_k]
+        auto qk = std::make_shared<ov::op::v0::MatMul>(q_r, k_r, /*tA=*/false, /*tB=*/true);
+        auto qk_scaled = std::make_shared<ov::op::v1::Multiply>(qk, scale_node);
+
+        // Mask arrives as [B, 1, S_q, S_k]. Unsqueeze a factor axis at position 2 to
+        // get [B, 1, 1, S_q, S_k], which NUMPY-broadcasts cleanly against the
+        // [B, num_heads_kv, factor, S_q, S_k] scores: B==B, then 1→num_heads_kv and
+        // 1→factor on the head dims.
+        ov::Output<ov::Node> qk_masked;
+        if (has_mask) {
+            auto mask_unsq1 =
+                std::make_shared<ov::op::v0::Unsqueeze>(mask, ov::op::v0::Constant::create(ov::element::i64, {1}, {2}));
+            qk_masked = std::make_shared<ov::op::v1::Add>(qk_scaled, mask_unsq1);
+        } else {
+            qk_masked = qk_scaled;
+        }
+
+        auto softmax = std::make_shared<ov::op::v8::Softmax>(qk_masked, /*axis=*/-1);
+
+        // softmax @ V → [B, num_heads_kv, factor, S_q, head_size]
+        auto attn = std::make_shared<ov::op::v0::MatMul>(softmax, v_r);
+
+        // Reshape back to [B, num_heads, S_q, head_size] (combine num_heads_kv * factor).
+        // Leading 0 + special_zero=true copies B at runtime.
+        auto out_4d_shape =
+            ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{0, num_heads, -1, head_size});
+        auto out_4d = std::make_shared<ov::op::v1::Reshape>(attn, out_4d_shape, true);
+
+        // The standard SDPA path's downstream is Transpose(0,2,1,3) → Convert(f32).
+        // Replicate it here so callers see the same output layout/dtype.
+        res = std::make_shared<ov::op::v1::Transpose>(
+            out_4d, ov::op::v0::Constant::create(ov::element::i64, {4}, {0, 2, 1, 3}));
+        res = std::make_shared<ov::op::v0::Convert>(res, ov::element::f32);
+        return rename_outputs_with_suffix({res}, context.get_name());
+    }
+
+    // Default path: explicit Broadcast → SDPA. Kept as the fallback because
+    // (a) it goes through the GPU plugin's micro-SDPA fast path (FlashAttention
+    // tiles via DPAS), and (b) the manual path above is still being validated.
+    auto tile_kv = [&](int64_t n_heads, int64_t n_heads_kv, int64_t hs, ov::Output<Node> kv) {
+        int64_t f = n_heads / n_heads_kv;
+        if (f > 1 && n_heads_kv > 1) {
+            ov::Output<ov::Node> kv_broadcast_shape;
+            ov::Output<ov::Node> kv_unsqueezed;
+            ov::Output<ov::Node> new_kv_shape;
+            auto unsqueeze_axes = ov::op::v0::Constant::create(ov::element::i64, Shape{}, {2});
+            kv_unsqueezed = std::make_shared<ov::op::v0::Unsqueeze>(kv, unsqueeze_axes);
+
+            kv_broadcast_shape = ov::op::v0::Constant::create(ov::element::i64, {5},
+                                                              {(int64_t) 1, (int64_t) 1, f, (int64_t) 1, (int64_t) 1});
+            new_kv_shape =
+                ov::op::v0::Constant::create(ov::element::i64, {4}, {(int64_t) 0, n_heads, (int64_t) -1, hs});
+            //    ov::element::i64, {5}, {(int64_t) 1, (int64_t) 1, factor, (int64_t) 1, (int64_t) 1});
+            //new_kv_shape =
+            //    ov::op::v0::Constant::create(ov::element::i64, {4}, {(int64_t) 0, num_heads, (int64_t) -1, head_size});
+
+            kv = std::make_shared<ov::op::v3::Broadcast>(kv_unsqueezed, kv_broadcast_shape,
+                                                         ov::op::BroadcastType::BIDIRECTIONAL);
+            kv = std::make_shared<ov::op::v1::Reshape>(kv, new_kv_shape, true);
+        }
+        return kv;
+    };
+
+    //auto q_shape = context.get_input_shape(0).to_shape();
+    //auto k_shape = context.get_input_shape(1).to_shape();
+    //k = tile_kv(q_shape[1], k_shape[1], q_shape[3], k);
+    //v = tile_kv(q_shape[1], k_shape[1], q_shape[3], v);
+    k = tile_kv(num_heads, num_heads_kv, head_size, k);
+    v = tile_kv(num_heads, num_heads_kv, head_size, v);
+
+    constexpr auto causal = false;
+    if (has_mask) {
+        auto sdpa = std::make_shared<ov::op::v13::ScaledDotProductAttention>(q, k, v, mask, scale_node, causal);
+        res = std::make_shared<ov::op::v1::Transpose>(
+            sdpa, ov::op::v0::Constant::create(ov::element::i64, {4}, {0, 2, 1, 3}));
+    } else {
+        auto sdpa = std::make_shared<ov::op::v13::ScaledDotProductAttention>(q, k, v, scale_node, causal);
+        res = std::make_shared<ov::op::v1::Transpose>(
+            sdpa, ov::op::v0::Constant::create(ov::element::i64, {4}, {0, 2, 1, 3}));
+    }
+    res = std::make_shared<ov::op::v0::Convert>(res, ov::element::f32);
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/gated_delta_net.cpp

@@ -0,0 +1,407 @@
+#include "gated_delta_net.hpp"
+
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+#include "ggml-openvino/ggml-openvino-extra.h"
+
+#include <cmath>
+#include <cstdint>
+#include <memory>
+#include <numeric>
+#include <openvino/op/add.hpp>
+#include <openvino/op/broadcast.hpp>
+#include <openvino/op/concat.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/divide.hpp>
+#include <openvino/op/exp.hpp>
+#include <openvino/op/gather.hpp>
+#include <openvino/op/less.hpp>
+#include <openvino/op/loop.hpp>
+#include <openvino/op/matmul.hpp>
+#include <openvino/op/multiply.hpp>
+#include <openvino/op/reduce_mean.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/squeeze.hpp>
+#include <openvino/op/sqrt.hpp>
+#include <openvino/op/subtract.hpp>
+#include <openvino/op/tile.hpp>
+#include <openvino/op/transpose.hpp>
+#include <openvino/op/unsqueeze.hpp>
+#include <vector>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+static OutputVector translate_gated_delta_net_ref(const NodeContext & context);
+
+static bool match_gdn_l2_norm(const Output<Node> & normalized, Output<Node> & input, float & eps) {
+    // Match the RMSNorm decomposition emitted by translate_rms_norm, followed by GGML SCALE.
+    const auto scale = ov::as_type_ptr<ov::op::v1::Multiply>(normalized.get_node_shared_ptr());
+    if (!scale) {
+        return false;
+    }
+    const auto factor = ov::as_type_ptr<ov::op::v0::Constant>(scale->get_input_node_shared_ptr(1));
+    const auto rms = ov::as_type_ptr<ov::op::v1::Multiply>(scale->get_input_node_shared_ptr(0));
+    if (!factor || ov::shape_size(factor->get_shape()) != 1 || !rms) {
+        return false;
+    }
+    const auto x = rms->input_value(0);
+    const auto & shape = x.get_partial_shape();
+    if (x.get_element_type() != ov::element::f32 || shape.rank() != 4 || shape[3].is_dynamic() ||
+        shape[3].get_length() <= 0 || normalized.get_partial_shape() != shape) {
+        return false;
+    }
+    const float dim = static_cast<float>(shape[3].get_length());
+    if (factor->cast_vector<float>()[0] != 1.0f / std::sqrt(dim)) {
+        return false;
+    }
+    const auto reciprocal = ov::as_type_ptr<ov::op::v1::Divide>(rms->get_input_node_shared_ptr(1));
+    if (!reciprocal) {
+        return false;
+    }
+    const auto one = ov::as_type_ptr<ov::op::v0::Constant>(reciprocal->get_input_node_shared_ptr(0));
+    const auto root = ov::as_type_ptr<ov::op::v0::Sqrt>(reciprocal->get_input_node_shared_ptr(1));
+    if (!one || ov::shape_size(one->get_shape()) != 1 || one->cast_vector<float>()[0] != 1.0f || !root) {
+        return false;
+    }
+    const auto add = ov::as_type_ptr<ov::op::v1::Add>(root->get_input_node_shared_ptr(0));
+    if (!add) {
+        return false;
+    }
+    const auto mean = ov::as_type_ptr<ov::op::v1::ReduceMean>(add->get_input_node_shared_ptr(0));
+    const auto rms_eps = ov::as_type_ptr<ov::op::v0::Constant>(add->get_input_node_shared_ptr(1));
+    if (!mean || !mean->get_keep_dims() || !rms_eps || ov::shape_size(rms_eps->get_shape()) != 1) {
+        return false;
+    }
+    const auto axes = ov::as_type_ptr<ov::op::v0::Constant>(mean->get_input_node_shared_ptr(1));
+    const auto square = ov::as_type_ptr<ov::op::v1::Multiply>(mean->get_input_node_shared_ptr(0));
+    if (!axes || axes->cast_vector<int64_t>() != std::vector<int64_t>{-1} || !square ||
+        square->input_value(0) != x || square->input_value(1) != x) {
+        return false;
+    }
+    // RMSNorm(x, rms_eps) / sqrt(D) = x / sqrt(sum(x*x) + D*rms_eps).
+    eps = dim * rms_eps->cast_vector<float>()[0];
+    if (!std::isfinite(eps) || eps <= 0.0f) {
+        return false;
+    }
+    input = x;
+    return true;
+}
+
+OutputVector translate_gated_delta_net(const NodeContext & context) {
+    auto v_shape = context.get_input_shape(2).to_shape();  // [B, T, H_v, S_v]
+    auto q_shape = context.get_input_shape(0).to_shape();  // [B, T, H_k, S_k]
+
+    // Fused GatedDeltaNet op only supports scalar gate (kda=0).
+    // Fall back to reference implementation for per-key-dimension gating.
+    // if (kda) {
+    //     return translate_gated_delta_net_ref(context);
+    // }
+
+    // const int64_t B = v_shape[0];
+    // const int64_t T = v_shape[1];
+    const int64_t H_v = v_shape[2];
+    const int64_t S_v = v_shape[3];
+    const int64_t H_k = q_shape[2];
+    // const int64_t S_k = q_shape[3];
+
+    auto q = context.get_input(0);
+    auto k = context.get_input(1);
+    auto v = process_view_input(context, 2, H_v * S_v, 3);
+    auto g = context.get_input(3);
+    auto beta = context.get_input(4);
+    auto state = context.get_input(5);
+
+    Output<Node> raw_q, raw_k;
+    float q_eps = 1e-6f, k_eps = 1e-6f;
+    const bool fuse_qk_l2norm = ggml_openvino_is_gpu() &&
+        match_gdn_l2_norm(q, raw_q, q_eps) && match_gdn_l2_norm(k, raw_k, k_eps);
+    if (fuse_qk_l2norm) {
+        // Keep head tiling below; GDN applies normalization per head and keeps its attention scale.
+        q = raw_q;
+        k = raw_k;
+    }
+
+    // ggml maps GQA heads in tiled order, while OV GDN maps repeated heads in grouped order.
+    if (H_v != H_k) {
+        const int64_t repeat = H_v / H_k;
+        auto repeats = ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{1, 1, repeat, 1});
+        q = std::make_shared<ov::op::v0::Tile>(q, repeats);
+        k = std::make_shared<ov::op::v0::Tile>(k, repeats);
+    }
+
+    if (context.get_view_input_size(2)) {
+        // Same as l2_norm case 1
+        v = std::make_shared<ov::op::v0::Squeeze>(v, ov::op::v0::Constant::create(ov::element::i64, {1}, {0}));
+        auto v_shape = context.get_input_shape(2).to_shape();
+        std::vector<int64_t> reshape_pattern = {0, 0, (int64_t) v_shape[2], (int64_t) v_shape[3]};
+        v = std::make_shared<ov::op::v1::Reshape>(
+            v, ov::op::v0::Constant::create(ov::element::i64, {4}, reshape_pattern), true);
+    }
+
+    // ggml state layout (OV notation): [B, H_v, value_dim, key_dim]
+    // GatedDeltaNet op expects: [B, H_v, key_dim, value_dim]
+    auto state_perm = ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{0, 1, 3, 2});
+    state = std::make_shared<ov::op::v1::Transpose>(state, state_perm);
+
+    g = std::make_shared<ov::op::v0::Squeeze>(g, ov::op::v0::Constant::create(ov::element::i64, {1}, {3}));
+    beta = std::make_shared<ov::op::v0::Squeeze>(beta, ov::op::v0::Constant::create(ov::element::i64, {1}, {3}));
+
+    if (context.has_input("chunk_valid_len")) {
+        // The last prefill chunk is padded with fabricated tokens. The recurrence is
+        //   S_t = S_{t-1} * exp(g_t) + k_t (x) ((v_t - S_{t-1}^T k_t) * beta_t)
+        // so forcing g = 0 and beta = 0 makes a padded step an exact identity and keeps the final
+        // state equal to the state after the last real token. Attention output at those positions
+        // is garbage but never read.
+        const auto & g_ps = g.get_partial_shape();
+        FRONT_END_OP_CONVERSION_CHECK(g_ps.rank().is_static() && g_ps.rank().get_length() == 3 && g_ps[1].is_static(),
+                                      "GATED_DELTA_NET pad masking requires a static token dimension");
+        const int64_t n_tokens = g_ps[1].get_length();
+        std::vector<int64_t> positions(n_tokens);
+        std::iota(positions.begin(), positions.end(), 0);
+        auto valid = std::make_shared<ov::op::v1::Less>(
+            ov::op::v0::Constant::create(ov::element::i64, {(size_t) n_tokens}, positions),
+            context.get_input("chunk_valid_len"));
+        auto mask = std::make_shared<ov::op::v0::Unsqueeze>(
+            std::make_shared<ov::op::v0::Convert>(valid, g.get_element_type()),
+            ov::op::v0::Constant::create(ov::element::i64, {2}, std::vector<int64_t>{0, 2}));
+        g = std::make_shared<ov::op::v1::Multiply>(g, mask);
+        beta = std::make_shared<ov::op::v1::Multiply>(beta, mask);
+    }
+
+    // std::cout << "GatedDeltaNet input shapes: q=" << q.get_partial_shape() << ", k=" << k.get_partial_shape()
+    //           << ", v=" << v.get_partial_shape() << ", g=" << g.get_partial_shape()
+    //           << ", beta=" << beta.get_partial_shape() << ", state=" << state.get_partial_shape() << std::endl;
+
+    auto gdn = std::make_shared<ov::op::internal::GatedDeltaNet>(q, k, v, state, g, beta, fuse_qk_l2norm, q_eps, k_eps);
+    auto attn_4d = gdn->output(0);
+    auto state_4d = gdn->output(1);  // [B, H_v, key_dim, value_dim]
+
+    // std::cout << "GatedDeltaNet output shapes: attn=" << gdn->output(0).get_partial_shape()
+    //           << ", new_state=" << gdn->output(1).get_partial_shape() << std::endl;
+
+    // Transpose output state back to ggml layout [B, H_v, value_dim, key_dim]
+    auto state_transposed = std::make_shared<ov::op::v1::Transpose>(state_4d, state_perm);
+    if (context.get_output_names().size() == 2) {
+        // The canonical graph consumes the packed GGML result only through separate attention
+        // and state VIEWs. Publish the native outputs under those VIEW names to avoid
+        // flatten -> concat -> reshape -> slice -> reshape chains. This also works for B > 1.
+        return rename_outputs_with_suffix({attn_4d, state_transposed}, context.get_name());
+    }
+
+    auto flat_shape_1d = ov::op::v0::Constant::create(ov::element::i64, {1}, {-1});
+    auto attn = std::make_shared<ov::op::v1::Reshape>(attn_4d, flat_shape_1d, false);
+    auto new_state = std::make_shared<ov::op::v1::Reshape>(state_transposed, flat_shape_1d, false);
+    auto packed = std::make_shared<ov::op::v0::Concat>(ov::OutputVector{attn, new_state}, 0);
+    auto out_shape = ov::op::v0::Constant::create(ov::element::i64, {4},
+                                                  std::vector<int64_t>{1, 1, -1 /*T * B + S_v * B*/, S_v * H_v});
+    auto res = std::make_shared<ov::op::v1::Reshape>(packed, out_shape, false);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+static OutputVector translate_gated_delta_net_ref(const NodeContext & context) {
+    num_inputs_check(context, 6, 6);
+
+    // Inputs (OV shapes are reversed from ggml):
+    // ggml: q[S_k, H_k, T, B], k[S_k, H_k, T, B], v[S_v, H_v, T, B]
+    // OV:   q[B, T, H_k, S_k], k[B, T, H_k, S_k], v[B, T, H_v, S_v]
+    // ggml: g[1 or S_v, H_v, T, B], beta[1, H_v, T, B]
+    // OV:   g[B, T, H_v, 1 or S_v], beta[B, T, H_v, 1]
+    // ggml: state[S_v, S_v, H_v, B]
+    // OV:   state[B, H_v, S_v, S_v]
+    auto q = process_view_input_new(context, 0);
+    auto k = process_view_input_new(context, 1);
+    auto v = process_view_input_new(context, 2);
+    auto g = process_view_input_new(context, 3);
+    auto beta = process_view_input_new(context, 4);
+    auto state = process_view_input_new(context, 5);
+
+    auto v_shape = context.get_input_shape(2).to_shape();  // [B, T, H_v, S_v]
+    auto q_shape = context.get_input_shape(0).to_shape();  // [B, T, H_k, S_k]
+    auto g_shape = context.get_input_shape(3).to_shape();  // [B, T, H_v, 1 or S_v]
+
+    const int64_t B = v_shape[0];
+    const int64_t T = v_shape[1];
+    const int64_t H_v = v_shape[2];
+    const int64_t S_v = v_shape[3];
+    const int64_t H_k = q_shape[2];
+    const bool kda = (g_shape[3] == (size_t) S_v);
+
+    const int64_t rq1 = H_v / H_k;  // head repeat factor
+    const float scale = 1.0f / std::sqrt((float) S_v);
+
+    auto axis_1 = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+    auto axis_2 = ov::op::v0::Constant::create(ov::element::i64, {1}, {2});
+
+    // Transpose inputs from [B, T, H, S] to [B, H, T, S] for easier per-head processing
+    auto perm_0213 = ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{0, 2, 1, 3});
+    auto q_t = std::make_shared<ov::op::v1::Transpose>(q, perm_0213);        // [B, H_k, T, S_k]
+    auto k_t = std::make_shared<ov::op::v1::Transpose>(k, perm_0213);        // [B, H_k, T, S_k]
+    auto v_t = std::make_shared<ov::op::v1::Transpose>(v, perm_0213);        // [B, H_v, T, S_v]
+    auto g_t = std::make_shared<ov::op::v1::Transpose>(g, perm_0213);        // [B, H_v, T, 1 or S_v]
+    auto beta_t = std::make_shared<ov::op::v1::Transpose>(beta, perm_0213);  // [B, H_v, T, 1]
+
+    // Broadcast Q, K heads to match V heads if GQA is used (H_v > H_k)
+    ov::Output<ov::Node> q_bh = q_t;
+    ov::Output<ov::Node> k_bh = k_t;
+    if (rq1 > 1) {
+        auto q_unsq = std::make_shared<ov::op::v0::Unsqueeze>(q_t, axis_2);  // [B, H_k, 1, T, S]
+        auto k_unsq = std::make_shared<ov::op::v0::Unsqueeze>(k_t, axis_2);  // [B, H_k, 1, T, S]
+
+        auto bcast_shape = ov::op::v0::Constant::create(ov::element::i64, {5}, std::vector<int64_t>{1, 1, rq1, 1, 1});
+        auto q_bcast =
+            std::make_shared<ov::op::v3::Broadcast>(q_unsq, bcast_shape, ov::op::BroadcastType::BIDIRECTIONAL);
+        auto k_bcast =
+            std::make_shared<ov::op::v3::Broadcast>(k_unsq, bcast_shape, ov::op::BroadcastType::BIDIRECTIONAL);
+
+        // Transpose [B, H_k, rq1, T, S] -> [B, rq1, H_k, T, S] so that reshape merges
+        // as [rq1, H_k] giving repeat-blocks pattern matching CPU: iq1 = iv1 % H_k
+        auto perm_5d = ov::op::v0::Constant::create(ov::element::i64, {5}, std::vector<int64_t>{0, 2, 1, 3, 4});
+        auto q_transposed = std::make_shared<ov::op::v1::Transpose>(q_bcast, perm_5d);
+        auto k_transposed = std::make_shared<ov::op::v1::Transpose>(k_bcast, perm_5d);
+
+        auto new_shape = ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{B, H_v, T, S_v});
+        q_bh = std::make_shared<ov::op::v1::Reshape>(q_transposed, new_shape, false);
+        k_bh = std::make_shared<ov::op::v1::Reshape>(k_transposed, new_shape, false);
+    }
+
+    // Merge batch and head dims: [B*H_v, T, S_v]
+    auto merge_bh = [&](const ov::Output<ov::Node> & x, int64_t last_dim) {
+        auto shape = ov::op::v0::Constant::create(ov::element::i64, {3}, std::vector<int64_t>{B * H_v, T, last_dim});
+        return std::make_shared<ov::op::v1::Reshape>(x, shape, false);
+    };
+
+    auto q_m = merge_bh(q_bh, S_v);           // [B*H_v, T, S_v]
+    auto k_m = merge_bh(k_bh, S_v);           // [B*H_v, T, S_v]
+    auto v_m = merge_bh(v_t, S_v);            // [B*H_v, T, S_v]
+    auto g_m = merge_bh(g_t, kda ? S_v : 1);  // [B*H_v, T, 1 or S_v]
+    auto beta_m = merge_bh(beta_t, 1);        // [B*H_v, T, 1]
+
+    // State: [B, H_v, S_v, S_v] -> [B*H_v, S_v, S_v]
+    auto state_shape = ov::op::v0::Constant::create(ov::element::i64, {3}, std::vector<int64_t>{B * H_v, S_v, S_v});
+    auto state_m = std::make_shared<ov::op::v1::Reshape>(state, state_shape, false);
+
+    auto scale_const = ov::op::v0::Constant::create(ov::element::f32, {}, std::vector<float>{scale});
+
+    // --- Build Loop body ---
+    // Body parameters (no iteration counter needed, use -1 in special ports)
+    auto body_state = std::make_shared<ov::op::v0::Parameter>(ov::element::f32, ov::PartialShape::dynamic());
+    auto body_q = std::make_shared<ov::op::v0::Parameter>(ov::element::f32, ov::PartialShape::dynamic());
+    auto body_k = std::make_shared<ov::op::v0::Parameter>(ov::element::f32, ov::PartialShape::dynamic());
+    auto body_v = std::make_shared<ov::op::v0::Parameter>(ov::element::f32, ov::PartialShape::dynamic());
+    auto body_g = std::make_shared<ov::op::v0::Parameter>(ov::element::f32, ov::PartialShape::dynamic());
+    auto body_beta = std::make_shared<ov::op::v0::Parameter>(ov::element::f32, ov::PartialShape::dynamic());
+    auto body_iter = std::make_shared<ov::op::v0::Parameter>(ov::element::i64, ov::Shape{1});
+
+    // Condition output (always true - we rely on trip_count for termination)
+    auto body_cond_out = ov::op::v0::Constant::create(ov::element::boolean, ov::Shape{1}, std::vector<bool>{true});
+
+    // Gather current token from invariant inputs using iteration counter
+    auto q_t_cur = std::make_shared<ov::op::v8::Gather>(body_q, body_iter, axis_1);     // [B*H_v, 1, S_v]
+    auto k_t_cur = std::make_shared<ov::op::v8::Gather>(body_k, body_iter, axis_1);     // [B*H_v, 1, S_v]
+    auto v_t_cur = std::make_shared<ov::op::v8::Gather>(body_v, body_iter, axis_1);     // [B*H_v, 1, S_v]
+    auto g_t_cur = std::make_shared<ov::op::v8::Gather>(body_g, body_iter, axis_1);     // [B*H_v, 1, 1 or S_v]
+    auto b_t_cur = std::make_shared<ov::op::v8::Gather>(body_beta, body_iter, axis_1);  // [B*H_v, 1, 1]
+
+    // Squeeze token dim
+    auto q_cur = std::make_shared<ov::op::v0::Squeeze>(q_t_cur, axis_1);  // [B*H_v, S_v]
+    auto k_cur = std::make_shared<ov::op::v0::Squeeze>(k_t_cur, axis_1);  // [B*H_v, S_v]
+    auto v_cur = std::make_shared<ov::op::v0::Squeeze>(v_t_cur, axis_1);  // [B*H_v, S_v]
+    auto g_cur = std::make_shared<ov::op::v0::Squeeze>(g_t_cur, axis_1);  // [B*H_v, 1 or S_v]
+    auto b_cur = std::make_shared<ov::op::v0::Squeeze>(b_t_cur, axis_1);  // [B*H_v, 1]
+
+    // Step 1: Apply decay gate to state
+    auto exp_g = std::make_shared<ov::op::v0::Exp>(g_cur);                                // [B*H_v, 1 or S_v]
+    auto exp_g_unsq = std::make_shared<ov::op::v0::Unsqueeze>(exp_g, axis_1);             // [B*H_v, 1, 1 or S_v]
+    auto state_decayed = std::make_shared<ov::op::v1::Multiply>(body_state, exp_g_unsq);  // [B*H_v, S_v, S_v]
+
+    // Step 2: delta = (v - S @ k) * beta
+    auto k_col = std::make_shared<ov::op::v0::Unsqueeze>(k_cur, axis_2);                 // [B*H_v, S_v, 1]
+    auto sk = std::make_shared<ov::op::v0::MatMul>(state_decayed, k_col, false, false);  // [B*H_v, S_v, 1]
+    auto sk_sq = std::make_shared<ov::op::v0::Squeeze>(sk, axis_2);                      // [B*H_v, S_v]
+    auto v_minus_sk = std::make_shared<ov::op::v1::Subtract>(v_cur, sk_sq);              // [B*H_v, S_v]
+    auto delta = std::make_shared<ov::op::v1::Multiply>(v_minus_sk, b_cur);              // [B*H_v, S_v]
+
+    // Step 3: state += outer(delta, k)
+    auto delta_col = std::make_shared<ov::op::v0::Unsqueeze>(delta, axis_2);                 // [B*H_v, S_v, 1]
+    auto k_row = std::make_shared<ov::op::v0::Unsqueeze>(k_cur, axis_1);                     // [B*H_v, 1, S_v]
+    auto outer_prod = std::make_shared<ov::op::v0::MatMul>(delta_col, k_row, false, false);  // [B*H_v, S_v, S_v]
+    auto state_updated = std::make_shared<ov::op::v1::Add>(state_decayed, outer_prod);       // [B*H_v, S_v, S_v]
+
+    // Step 4: attn_out = S @ q * scale
+    auto q_col = std::make_shared<ov::op::v0::Unsqueeze>(q_cur, axis_2);                 // [B*H_v, S_v, 1]
+    auto sq = std::make_shared<ov::op::v0::MatMul>(state_updated, q_col, false, false);  // [B*H_v, S_v, 1]
+    auto sq_squeezed = std::make_shared<ov::op::v0::Squeeze>(sq, axis_2);                // [B*H_v, S_v]
+    auto attn_out = std::make_shared<ov::op::v1::Multiply>(sq_squeezed, scale_const);    // [B*H_v, S_v]
+
+    // Unsqueeze attn_out to [B*H_v, 1, S_v] for scan output concatenation
+    auto attn_out_unsq = std::make_shared<ov::op::v0::Unsqueeze>(attn_out, axis_1);  // [B*H_v, 1, S_v]
+
+    // --- Assemble Loop ---
+    // Body: results = [condition, state_updated, attn_out_unsq]
+    auto body = std::make_shared<ov::Model>(
+        ov::OutputVector{body_cond_out, state_updated, attn_out_unsq},
+        ov::ParameterVector{body_iter, body_state, body_q, body_k, body_v, body_g, body_beta});
+
+    auto trip_count = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{1}, std::vector<int64_t>{T});
+    auto exec_cond = ov::op::v0::Constant::create(ov::element::boolean, ov::Shape{1}, std::vector<bool>{true});
+
+    auto loop = std::make_shared<ov::op::v5::Loop>(trip_count, exec_cond);
+    loop->set_function(body);
+    loop->set_special_body_ports(ov::op::v5::Loop::SpecialBodyPorts{0, 0});
+
+    // Carried state: feeds back from body output 1 to body_state param
+    loop->set_merged_input(body_state, state_m, state_updated);
+    // Invariant inputs: passed through unchanged each iteration
+    loop->set_invariant_input(body_q, q_m);
+    loop->set_invariant_input(body_k, k_m);
+    loop->set_invariant_input(body_v, v_m);
+    loop->set_invariant_input(body_g, g_m);
+    loop->set_invariant_input(body_beta, beta_m);
+
+    // Loop outputs:
+    // 1) Final state (last iteration value of state_updated)
+    auto final_state_out = loop->get_iter_value(state_updated, -1);  // [B*H_v, S_v, S_v]
+    // 2) Concatenated attention outputs across all iterations along axis 1
+    auto attn_concat_out = loop->get_concatenated_slices(attn_out_unsq, 0, 1, 1, -1, 1);  // [B*H_v, T, S_v]
+
+    // --- Pack outputs to match ggml layout ---
+    // ggml output ne = {S_v*H, T*B + S_v*B, 1, 1} -> OV [1, 1, T*B+S_v*B, S_v*H_v]
+    // attn: [B, T, H_v, S_v] row-major, state: [B, H_v, S_v, S_v] row-major
+
+    // attn: [B*H_v, T, S_v] -> [B, H_v, T, S_v] -> transpose to [B, T, H_v, S_v] -> flatten
+    auto attn_4d_shape = ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{B, H_v, T, S_v});
+    auto attn_4d = std::make_shared<ov::op::v1::Reshape>(attn_concat_out, attn_4d_shape, false);
+    auto attn_perm = std::make_shared<ov::op::v1::Transpose>(attn_4d, perm_0213);  // [B, T, H_v, S_v]
+
+    auto flat_shape_1d = ov::op::v0::Constant::create(ov::element::i64, {1}, std::vector<int64_t>{-1});
+    auto attn_1d = std::make_shared<ov::op::v1::Reshape>(attn_perm, flat_shape_1d, false);
+
+    // state: [B*H_v, S_v, S_v] -> [B, H_v, S_v, S_v] -> flatten
+    auto state_4d_shape = ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{B, H_v, S_v, S_v});
+    auto state_4d = std::make_shared<ov::op::v1::Reshape>(final_state_out, state_4d_shape, false);
+    if (context.get_output_names().size() == 2) {
+        // Match the fused translator's direct attention/state contract.
+        return rename_outputs_with_suffix({attn_perm, state_4d}, context.get_name());
+    }
+
+    auto state_1d = std::make_shared<ov::op::v1::Reshape>(state_4d, flat_shape_1d, false);
+
+    // Concat [attn | state] and reshape to final output
+    auto packed = std::make_shared<ov::op::v0::Concat>(ov::OutputVector{attn_1d, state_1d}, 0);
+    auto out_shape =
+        ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{1, 1, T * B + S_v * B, S_v * H_v});
+    auto res = std::make_shared<ov::op::v1::Reshape>(packed, out_shape, false);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/gated_delta_net.hpp

@@ -0,0 +1,65 @@
+#pragma once
+
+#include "openvino/op/op.hpp"
+
+namespace ov::op::internal {
+/// \note GatedDeltaNet op class is under development and subject to change
+///
+/// \brief Operator performing Gated Delta Net computation
+/// \ingroup ov_ops_cpp_api
+class OPENVINO_API GatedDeltaNet : public ov::op::Op {
+public:
+    OPENVINO_OP("GatedDeltaNet")
+
+    GatedDeltaNet() = default;
+    /// \brief Constructs a GatedDeltaNet operation.
+    ///
+    /// \param query Query tensor input.
+    /// \param key Key tensor input.
+    /// \param value Value tensor input.
+    /// \param recurrent_state Initial recurrent state tensor.
+    /// \param gate Gate tensor controlling state decay/update.
+    /// \param beta Beta tensor scaling the delta update.
+    /// \param fuse_qk_l2norm Enables fusing q/k L2-normalization into this op.
+    /// \param q_l2_norm_eps Epsilon used for query L2-normalization when fusion is enabled.
+    /// \param k_l2_norm_eps Epsilon used for key L2-normalization when fusion is enabled.
+    GatedDeltaNet(const Output<Node>& query,
+                  const Output<Node>& key,
+                  const Output<Node>& value,
+                  const Output<Node>& recurrent_state,
+                  const Output<Node>& gate,
+                  const Output<Node>& beta,
+                  const bool fuse_qk_l2norm = false,
+                  const float q_l2_norm_eps = 1e-6F,
+                  const float k_l2_norm_eps = 1e-6F);
+
+    /// \brief Constructs a GatedDeltaNet operation from input vector.
+    ///
+    /// \param args Input tensor vector in order: query, key, value, recurrent_state, gate, beta.
+    /// \param fuse_qk_l2norm Enables fusing q/k L2-normalization into this op.
+    /// \param q_l2_norm_eps Epsilon used for query L2-normalization when fusion is enabled.
+    /// \param k_l2_norm_eps Epsilon used for key L2-normalization when fusion is enabled.
+    GatedDeltaNet(const ov::OutputVector& args,
+                  const bool fuse_qk_l2norm = false,
+                  const float q_l2_norm_eps = 1e-6F,
+                  const float k_l2_norm_eps = 1e-6F);
+    void validate_and_infer_types() override;
+    bool visit_attributes(AttributeVisitor& visitor) override;
+    std::shared_ptr<ov::Node> clone_with_new_inputs(const ov::OutputVector& new_args) const override;
+    bool get_fuse_qk_l2norm() const {
+        return m_fuse_qk_l2norm;
+    }
+    float get_q_l2_norm_eps() const {
+        return m_q_l2_norm_eps;
+    }
+    float get_k_l2_norm_eps() const {
+        return m_k_l2_norm_eps;
+    }
+
+private:
+    bool m_fuse_qk_l2norm = false;
+    float m_q_l2_norm_eps = 1e-6F;
+    float m_k_l2_norm_eps = 1e-6F;
+};
+
+}  // namespace ov::op::internal

llama.cpp/ggml/src/ggml-openvino/openvino/op/gather_matmul.hpp

@@ -0,0 +1,43 @@
+// Copyright (C) 2018-2026 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+// Local mirror of OpenVINO's internal ov::op::internal::GatherMatmul op.
+//
+// The op class body (validate_and_infer_types / clone_with_new_inputs) is
+// provided by the linked libopenvino.so; only the declaration is needed here so
+// the backend can construct the node directly (same approach as GatedDeltaNet).
+// The class layout must stay in sync with
+//   openvino/src/common/transformations/include/ov_ops/gather_matmul.hpp
+//
+// \note GatherMatmul op class is under development and subject to change.
+
+#pragma once
+
+#include "openvino/op/op.hpp"
+
+namespace ov::op::internal {
+
+class OPENVINO_API GatherMatmul : public ov::op::Op {
+public:
+    OPENVINO_OP("GatherMatmul")
+
+    GatherMatmul() = default;
+
+    GatherMatmul(const ov::Output<Node>& A,
+                 const ov::Output<Node>& B,
+                 const ov::Output<Node>& indices,
+                 const ov::Output<Node>& bias);
+
+    GatherMatmul(const ov::Output<Node>& A, const ov::Output<Node>& B, const ov::Output<Node>& indices);
+
+    std::shared_ptr<Node> clone_with_new_inputs(const ov::OutputVector& new_args) const override;
+
+    void validate_and_infer_types() override;
+
+private:
+    // the weights matrix B is expected to have the transposed form [group, N, K]
+    static constexpr bool transp_a = false;
+    static constexpr bool transp_b = true;
+};
+
+}  // namespace ov::op::internal

llama.cpp/ggml/src/ggml-openvino/openvino/op/get_rows.cpp

@@ -0,0 +1,178 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <climits>
+#include <openvino/core/node.hpp>
+#include <openvino/core/node_output.hpp>
+#include <openvino/op/add.hpp>
+#include <openvino/op/broadcast.hpp>
+#include <openvino/op/concat.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/gather.hpp>
+#include <openvino/op/shape_of.hpp>
+#include <openvino/op/slice.hpp>
+#include <openvino/op/squeeze.hpp>
+#include <openvino/op/unsqueeze.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_get_rows(const NodeContext & context) {
+    num_inputs_check(context, 2, 2);
+
+    Output<Node> res;
+    Output<Node> data;
+    int64_t row_offset = 0;
+    if (context.get_view_input_size(0) > 0 && context.get_input(0).get_partial_shape().rank() == 2) {
+        // A row range view over a 2D weight Constant folds into the gather indices,
+        // which keeps the dequantization subgraph intact for the plugins.
+        const auto view_shape = context.get_view_input_ggml_shape(0, 0);
+        const auto view_stride = context.get_view_input_stride(0, 0);
+        const size_t view_offset = context.get_view_input_offset(0, 0);
+        const size_t row_bytes = view_stride[2];
+        data = context.get_input(0);
+        FRONT_END_OP_CONVERSION_CHECK(row_bytes > 0 && view_offset % row_bytes == 0 &&
+                                          data.get_partial_shape()[1].compatible(view_shape[3]),
+                                      "GET_ROWS: view over a weight must be a row range");
+        row_offset = static_cast<int64_t>(view_offset / row_bytes);
+    } else {
+        data = process_view_input_new(context, 0);
+    }
+
+    auto op_case = context.get_op_case();
+    if (op_case == 3 || op_case == 4) {
+        return {data};
+    }
+
+    ov::Output<ov::Node> indices;
+    if ((op_case == 1 || op_case == 2) && context.has_input("s_copy_active_slot_len")) {
+        // Recurrent state reorder (inp->s_copy): slice the active (op_case 1) or extra (op_case 2)
+        // segment from the s_copy index list at runtime, instead of baking the static view offset,
+        // so the cached IR works for any number of active sequences.
+        auto s_copy = context.get_input(1);
+        auto len = context.get_input("s_copy_active_slot_len");
+        auto step = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+        auto axis = ov::op::v0::Constant::create(ov::element::i64, {1}, {3});
+        if (op_case == 1) {
+            auto begin = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
+            indices = std::make_shared<ov::op::v8::Slice>(s_copy, begin, len, step, axis);
+        } else {
+            auto end = ov::op::v0::Constant::create(ov::element::i64, {1}, {INT_MAX});
+            indices = std::make_shared<ov::op::v8::Slice>(s_copy, len, end, step, axis);
+        }
+    } else {
+        indices = process_view_input_new(context, 1);
+    }
+
+    // data[1,b,x,y] ind[1,1,b,x'] test-backend-ops case
+    // data[x,y] ind[1,1,1,x'] normal case
+    // Squeeze the leading dims down to [b,x']. Stateful models drop one rank, so a hardcoded
+    // {0,1} would also strip the batch dim whenever b == 1 (every decode step).
+    const auto indices_rank = indices.get_partial_shape().rank();
+    FRONT_END_OP_CONVERSION_CHECK(indices_rank.is_static(), "Expected static rank for GET_ROWS indices");
+    std::vector<int64_t> indices_squeeze_axes;
+    for (int64_t i = 0; i + 2 < indices_rank.get_length(); ++i) {
+        indices_squeeze_axes.push_back(i);
+    }
+    if (!indices_squeeze_axes.empty()) {
+        indices = std::make_shared<ov::op::v0::Squeeze>(
+            indices, ov::op::v0::Constant::create(ov::element::i64, {indices_squeeze_axes.size()},
+                                                  indices_squeeze_axes));
+    }
+    if (row_offset != 0) {
+        indices = std::make_shared<ov::op::v1::Add>(
+            indices, ov::op::v0::Constant::create(indices.get_element_type(), {}, {row_offset}));
+    }
+    if (data.get_partial_shape().rank() == 4) {
+        if (!(data.get_partial_shape()[1].is_dynamic()) && data.get_partial_shape()[1].get_length() == 1) {
+            // Work-around for a bug in ov cpu plugin for test-backend-ops
+            data = std::make_shared<ov::op::v0::Squeeze>(data,
+                                                         ov::op::v0::Constant::create(ov::element::i64, {2}, {0, 1}));
+            auto axis = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{}, {0});
+            res = std::make_shared<ov::op::v8::Gather>(data, indices, axis);
+        } else {
+            auto axis = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{}, {1});
+            data =
+                std::make_shared<ov::op::v0::Squeeze>(data, ov::op::v0::Constant::create(ov::element::i64, {1}, {0}));
+            // data: [batch, rows, ...], indices: [batch, n] - this is a batched gather
+            // (batch_dims=1) along the rows axis. The data and indices batch dims are
+            // logically equal (both == n_tokens) but reach this node through independent
+            // reshapes, so the GPU plugin's gather shape inference cannot prove
+            // data.shape[0] == indices.shape[0] and rejects the node. We must tie both
+            // batch dims to the SAME value, and crucially that value must stay DYNAMIC.
+            const auto data_ps = data.get_partial_shape();
+            const auto idx_ps = indices.get_partial_shape();
+            const bool data_batch_static = data_ps.rank().is_static() && data_ps[0].is_static();
+            const bool idx_batch_dynamic = idx_ps.rank().is_dynamic() || idx_ps[0].is_dynamic();
+
+            if (data_batch_static && idx_batch_dynamic) {
+                // MoE per-expert-scale path: `data` is a statically-tiled REPEAT
+                // (ggml_repeat_4d(scale, 1, n_expert, n_tokens, 1)) whose batch dim is a
+                // compile-time-constant n_tokens, and every batch slice is IDENTICAL (it was
+                // tiled from a single [1, n_expert, 1] scale). `indices` (selected_experts)
+                // carries the genuinely dynamic token dim. Broadcasting indices up to the
+                // static data batch (the naive fix) would freeze the token dim to the
+                // captured prefill length, and that static value then flows through the
+                // gather into the residual stream, making every following decoder layer
+                // static -> triggers the GPU in-place-concat KV-cache corruption (only
+                // layer 0 stays dynamic). A static->dynamic Broadcast cannot expand, so
+                // instead collapse the redundant data batch to 1 and broadcast 1->dynamic to
+                // match the indices batch. Mathematically identical (the slices are equal),
+                // and the whole graph stays dynamic.
+                auto zero = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
+                auto one = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+                auto axis0 = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
+                auto data_b1 = std::make_shared<ov::op::v8::Slice>(data, zero, one, one, axis0);  // [1, rows, ...]
+
+                auto idx_shape = std::make_shared<ov::op::v3::ShapeOf>(indices, ov::element::i64);
+                auto idx_batch = get_dimensions(idx_shape, {0});  // [batch] (dynamic)
+                auto data_b1_shape = std::make_shared<ov::op::v3::ShapeOf>(data_b1, ov::element::i64);
+                const auto rank = data_ps.rank().get_length();
+                std::vector<int> rest_axes;
+                for (int a = 1; a < rank; ++a) {
+                    rest_axes.push_back(a);
+                }
+                auto data_rest = get_dimensions(data_b1_shape, rest_axes);  // [rows, ...]
+                auto data_target = std::make_shared<ov::op::v0::Concat>(ov::OutputVector{idx_batch, data_rest}, 0);
+                data =
+                    std::make_shared<ov::op::v3::Broadcast>(data_b1, data_target, ov::op::BroadcastType::BIDIRECTIONAL);
+                res = std::make_shared<ov::op::v8::Gather>(data, indices, axis, 1);
+            } else {
+                // General case: tie the indices batch to the data batch (the data batch is
+                // already dynamic, e.g. the routing-weights gather whose data comes from the
+                // activations). Broadcast indices to [data_batch, indices_n].
+                auto data_shape = std::make_shared<ov::op::v3::ShapeOf>(data, ov::element::i64);
+                auto data_batch = get_dimensions(data_shape, {0});  // [batch]
+                auto idx_shape = std::make_shared<ov::op::v3::ShapeOf>(indices, ov::element::i64);
+                auto idx_n = get_dimensions(idx_shape, {1});  // [n]
+                auto idx_target = std::make_shared<ov::op::v0::Concat>(ov::OutputVector{data_batch, idx_n}, 0);
+                indices = std::make_shared<ov::op::v3::Broadcast>(indices, idx_target,
+                                                                  ov::op::BroadcastType::BIDIRECTIONAL);
+                res = std::make_shared<ov::op::v8::Gather>(data, indices, axis, 1);
+            }
+        }
+    } else if (context.is_stateful() && data.get_partial_shape().rank() == 3) {
+        auto axis = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{}, {1});
+        res = std::make_shared<ov::op::v8::Gather>(data, indices, axis, 1);
+    } else {
+        auto axis = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{}, {0});
+        res = std::make_shared<ov::op::v8::Gather>(data, indices, axis);
+    }
+
+    if (res.get_element_type() != context.get_output_type()) {
+        res = std::make_shared<ov::op::v0::Convert>(res, context.get_output_type());
+    }
+    if (!(context.is_stateful())) {
+        res = std::make_shared<ov::op::v0::Unsqueeze>(res, ov::op::v0::Constant::create(ov::element::i64, {1}, {0}));
+    }
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/glu_geglu.cpp

@@ -0,0 +1,75 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <memory>
+#include <openvino/core/node_output.hpp>
+#include <openvino/op/clamp.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/gelu.hpp>
+#include <openvino/op/multiply.hpp>
+#include <openvino/op/sigmoid.hpp>
+#include <openvino/op/slice.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_glu_geglu(const NodeContext & context) {
+    num_inputs_check(context, 1, 2);
+
+    ov::Output<ov::Node> src0;
+    ov::Output<ov::Node> src1;
+    if (context.get_input_size() == 2) {
+        // Inputs may be VIEW slices of a combined gate_up tensor (MoE experts):
+        // resolve them so each half has its real sliced shape, not the base tensor.
+        src0 = process_view_input_new(context, 0);
+        src1 = process_view_input_new(context, 1);
+    } else {
+        // GGML splits along ne[0] (OV last axis) using floor division: nc = ne[0] / 2.
+        // Both halves are nc elements; if the dimension is odd, the last element is dropped.
+        // Use Slice instead of Split to handle odd dimensions correctly.
+        // Resolve a VIEW input (e.g. non-contiguous slice) to its real shape first.
+        auto combined = process_view_input_new(context, 0);
+        auto combined_shape = combined.get_partial_shape();
+        int64_t last_dim_val = combined_shape[combined_shape.rank().get_length() - 1].get_length();
+        int64_t nc = last_dim_val / 2;
+
+        auto axis = ov::op::v0::Constant::create(ov::element::i64, {1}, {-1});
+        auto step = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+        auto start0 = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
+        auto stop0 = ov::op::v0::Constant::create(ov::element::i64, {1}, {nc});
+        auto start1 = ov::op::v0::Constant::create(ov::element::i64, {1}, {nc});
+        auto stop1 = ov::op::v0::Constant::create(ov::element::i64, {1}, {2 * nc});
+
+        src0 = std::make_shared<ov::op::v8::Slice>(combined, start0, stop0, step, axis);
+        src1 = std::make_shared<ov::op::v8::Slice>(combined, start1, stop1, step, axis);
+    }
+
+    int32_t * params = context.get_output_op_params();
+    const int32_t swapped = params[1];
+    if (swapped) {
+        std::swap(src0, src1);
+    }
+
+    if (context.is_static()) {
+        // TODO: Temporary solution for NPU accuracy issue due to fp16 overflow
+        // To be removed once permanent solution is implemented
+        // Justification:
+        // For |x| > 5, GELU(x) ≈ max(x, 0)  (behaves like ReLU)
+        // So Clamp(-10, 10) only affects values where GELU would return ≈ x anyway.
+        // The only loss: values > 10 get mapped to 10 instead of x.
+        // In practice, FFN intermediates rarely exceed 10 after GEGLU gating.
+        src0 = std::make_shared<ov::op::v0::Clamp>(src0, -10.0, 10.0);
+    }
+    auto gelu = std::make_shared<ov::op::v7::Gelu>(src0);
+    auto res = std::make_shared<ov::op::v1::Multiply>(gelu, src1);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/glu_geglu_quick.cpp

@@ -0,0 +1,62 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <memory>
+#include <openvino/core/node_output.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/multiply.hpp>
+#include <openvino/op/slice.hpp>
+#include <openvino/op/swish.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_glu_geglu_quick(const NodeContext & context) {
+    num_inputs_check(context, 1, 2);
+
+    ov::Output<ov::Node> src0;
+    ov::Output<ov::Node> src1;
+    if (context.get_input_size() == 2) {
+        src0 = process_view_input_new(context, 0);
+        src1 = process_view_input_new(context, 1);
+    } else {
+        // split along last axis, nc = ne[0] / 2
+        auto combined = process_view_input_new(context, 0);
+        auto combined_shape = combined.get_partial_shape();
+        int64_t last_dim_val = combined_shape[combined_shape.rank().get_length() - 1].get_length();
+        int64_t nc = last_dim_val / 2;
+
+        auto axis  = ov::op::v0::Constant::create(ov::element::i64, {1}, {-1});
+        auto step  = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+        auto start0 = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
+        auto stop0  = ov::op::v0::Constant::create(ov::element::i64, {1}, {nc});
+        auto start1 = ov::op::v0::Constant::create(ov::element::i64, {1}, {nc});
+        auto stop1  = ov::op::v0::Constant::create(ov::element::i64, {1}, {2 * nc});
+
+        src0 = std::make_shared<ov::op::v8::Slice>(combined, start0, stop0, step, axis);
+        src1 = std::make_shared<ov::op::v8::Slice>(combined, start1, stop1, step, axis);
+    }
+
+    int32_t * params = context.get_output_op_params();
+    const int32_t swapped = params[1];
+    if (swapped) {
+        std::swap(src0, src1);
+    }
+
+    // GELU_QUICK(x) = x * sigmoid(1.702 * x)
+    // Create the constant in the same type as src0 to avoid f16/f32 mismatch.
+    auto input_type = src0.get_element_type();
+    auto coef = ov::op::v0::Constant::create(input_type, ov::Shape{}, {1.702f});
+    auto gated   = std::make_shared<ov::op::v4::Swish>(src0, coef);
+    auto res     = std::make_shared<ov::op::v1::Multiply>(gated, src1);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/glu_swiglu.cpp

@@ -0,0 +1,131 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <cstdint>
+#include <limits>
+#include <memory>
+#include <openvino/core/node_output.hpp>
+#include <openvino/op/add.hpp>
+#include <openvino/op/clamp.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/multiply.hpp>
+#include <openvino/op/slice.hpp>
+#include <openvino/op/swish.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+static std::pair<ov::Output<ov::Node>, ov::Output<ov::Node>> get_glu_inputs(const NodeContext & context) {
+    num_inputs_check(context, 1, 2);
+
+    ov::Output<ov::Node> src0;
+    ov::Output<ov::Node> src1;
+    if (context.get_input_size() == 2) {
+        // Inputs may be VIEW slices of a combined gate_up tensor (MoE experts):
+        // resolve them so each half has its real sliced shape, not the base tensor.
+        src0 = process_view_input_new(context, 0);
+        src1 = process_view_input_new(context, 1);
+    } else {
+        // GGML splits along ne[0] (OV last axis) using floor division: nc = ne[0] / 2.
+        // Both halves are nc elements; if the dimension is odd, the last element is dropped.
+        // Use Slice instead of Split to handle odd dimensions correctly.
+        // Resolve a VIEW input (e.g. non-contiguous slice) to its real shape first.
+        auto combined = process_view_input_new(context, 0);
+        auto combined_shape = combined.get_partial_shape();
+        int64_t last_dim_val = combined_shape[combined_shape.rank().get_length() - 1].get_length();
+        int64_t nc = last_dim_val / 2;
+
+        auto axis = ov::op::v0::Constant::create(ov::element::i64, {1}, {-1});
+        auto step = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+        auto start0 = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
+        auto stop0 = ov::op::v0::Constant::create(ov::element::i64, {1}, {nc});
+        auto start1 = ov::op::v0::Constant::create(ov::element::i64, {1}, {nc});
+        auto stop1 = ov::op::v0::Constant::create(ov::element::i64, {1}, {2 * nc});
+
+        src0 = std::make_shared<ov::op::v8::Slice>(combined, start0, stop0, step, axis);
+        src1 = std::make_shared<ov::op::v8::Slice>(combined, start1, stop1, step, axis);
+    }
+
+    int32_t * params = context.get_output_op_params();
+    const int32_t swapped = params[1];
+    if (swapped) {
+        std::swap(src0, src1);
+    }
+
+    return {src0, src1};
+}
+
+OutputVector translate_glu_swiglu(const NodeContext & context) {
+    auto [src0, src1] = get_glu_inputs(context);
+
+    auto silu = std::make_shared<ov::op::v4::Swish>(src0);
+    auto res = std::make_shared<ov::op::v1::Multiply>(silu, src1);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+OutputVector translate_glu_swiglu_oai(const NodeContext & context) {
+    auto [src0, src1] = get_glu_inputs(context);
+
+    const int32_t * params = context.get_output_op_params();
+    const float alpha = reinterpret_cast<const float *>(params)[2];
+    const float limit = reinterpret_cast<const float *>(params)[3];
+
+    // Compute in f32: alpha_const/one are f32, and Swish/Add need matching input types.
+    auto output_type = context.get_output_type();
+    if (src0.get_element_type() != ov::element::f32) {
+        src0 = std::make_shared<ov::op::v0::Convert>(src0, ov::element::f32);
+    }
+    if (src1.get_element_type() != ov::element::f32) {
+        src1 = std::make_shared<ov::op::v0::Convert>(src1, ov::element::f32);
+    }
+
+    auto gate = std::make_shared<ov::op::v0::Clamp>(src0, -std::numeric_limits<float>::infinity(), limit);
+    auto alpha_const = ov::op::v0::Constant::create(ov::element::f32, {}, {alpha});
+    auto out_glu = std::make_shared<ov::op::v4::Swish>(gate, alpha_const);
+
+    auto up = std::make_shared<ov::op::v0::Clamp>(src1, -limit, limit);
+    auto one = ov::op::v0::Constant::create(ov::element::f32, {}, {1.0f});
+    auto up_plus_one = std::make_shared<ov::op::v1::Add>(up, one);
+    ov::Output<ov::Node> res = std::make_shared<ov::op::v1::Multiply>(out_glu, up_plus_one);
+    if (res.get_element_type() != output_type) {
+        res = std::make_shared<ov::op::v0::Convert>(res, output_type);
+    }
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+OutputVector translate_glu_swiglu_clamp(const NodeContext & context) {
+    auto [src0, src1] = get_glu_inputs(context);
+
+    const int32_t * params = context.get_output_op_params();
+    const float limit = reinterpret_cast<const float *>(params)[3];
+
+    // Compute in f32: f16 Swish/Clamp rounding drifts past the 1e-7 test tolerance.
+    auto output_type = context.get_output_type();
+    if (src0.get_element_type() != ov::element::f32) {
+        src0 = std::make_shared<ov::op::v0::Convert>(src0, ov::element::f32);
+    }
+    if (src1.get_element_type() != ov::element::f32) {
+        src1 = std::make_shared<ov::op::v0::Convert>(src1, ov::element::f32);
+    }
+
+    auto gate = std::make_shared<ov::op::v0::Clamp>(src0, -std::numeric_limits<float>::infinity(), limit);
+    auto silu = std::make_shared<ov::op::v4::Swish>(gate);
+    auto up = std::make_shared<ov::op::v0::Clamp>(src1, -limit, limit);
+    ov::Output<ov::Node> res = std::make_shared<ov::op::v1::Multiply>(silu, up);
+    if (res.get_element_type() != output_type) {
+        res = std::make_shared<ov::op::v0::Convert>(res, output_type);
+    }
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/im2col.cpp

@@ -0,0 +1,239 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <cstddef>
+#include <memory>
+#include <openvino/core/shape.hpp>
+#include <openvino/core/strides.hpp>
+#include <openvino/op/concat.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/extractimagepatches.hpp>
+#include <openvino/op/pad.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/slice.hpp>
+#include <openvino/op/transpose.hpp>
+#include <openvino/op/util/attr_types.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_im2col(const NodeContext & context) {
+    num_inputs_check(context, 2, 2);
+    const int32_t * params = context.get_output_op_params();
+    int32_t s0 = params[0];
+    int32_t s1 = params[1];
+    int32_t p0 = params[2];
+    int32_t p1 = params[3];
+    int32_t d0 = params[4];
+    int32_t d1 = params[5];
+    bool is_2D = params[6] == 1;
+    ov::Output<Node> res;
+
+    ov::Output<Node> image = context.get_input(1);
+    const ov::Shape kernel_shape = context.get_input(0).get_shape();
+
+    const size_t IC = is_2D ? kernel_shape[1] : kernel_shape[2];
+    const size_t KH = is_2D ? kernel_shape[2] : 1;
+    const size_t KW = kernel_shape[3];
+
+    int32_t stride_w = s0;
+    int32_t stride_h = is_2D ? s1 : 1;
+    int32_t pad_w = p0;
+    int32_t pad_h = is_2D ? p1 : 0;
+    int32_t dil_w = d0;
+    int32_t dil_h = is_2D ? d1 : 1;
+
+    if (!is_2D) {
+        // GGML input shape: [IW, IC, N, 1]
+        // OpenVINO input shape: [1, N, IC, IW]
+        // Reshape image to: [N, IC, 1, IW]
+        const ov::Shape image_shape = image.get_shape();
+        const size_t N = image_shape[1];
+        const size_t IW = image_shape[3];
+        auto image_reshape_shape = ov::op::v0::Constant::create(
+            ov::element::i64, ov::Shape{4},
+            std::vector<int64_t>{static_cast<int64_t>(N), static_cast<int64_t>(IC), 1, static_cast<int64_t>(IW)});
+        image = std::make_shared<ov::op::v1::Reshape>(image, image_reshape_shape, false);
+    }
+
+    const ov::Shape patch_sizes = {KH, KW};
+    const ov::Strides strides = {static_cast<size_t>(stride_h), static_cast<size_t>(stride_w)};
+    const ov::Shape rates = {static_cast<size_t>(dil_h), static_cast<size_t>(dil_w)};
+
+    auto pads_begin =
+        ov::op::v0::Constant::create(ov::element::i64, ov::Shape{4}, std::vector<int64_t>{0, 0, pad_h, pad_w});
+    auto pads_end =
+        ov::op::v0::Constant::create(ov::element::i64, ov::Shape{4}, std::vector<int64_t>{0, 0, pad_h, pad_w});
+
+    auto pad = std::make_shared<ov::op::v1::Pad>(image, pads_begin, pads_end, ov::op::PadMode::CONSTANT);
+    auto patches =
+        std::make_shared<ov::op::v3::ExtractImagePatches>(pad, patch_sizes, strides, rates, ov::op::PadType::VALID);
+
+    // [N, KH*KW*IC, OH, OW] → [N, OH, OW, KH*KW*IC]
+    auto perm1 = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{4}, std::vector<int64_t>{0, 2, 3, 1});
+    auto t1 = std::make_shared<ov::op::v1::Transpose>(patches, perm1);
+
+    // [N, OH, OW, KH*KW*IC] → [N, OH, OW, KH*KW, IC]
+    const ov::Shape out_shape = t1->get_output_shape(0);
+    const size_t N = out_shape[0];
+    const size_t OH = out_shape[1];
+    const size_t OW = out_shape[2];
+    auto reshape1_shape = ov::op::v0::Constant::create(
+        ov::element::i64, ov::Shape{5},
+        std::vector<int64_t>{static_cast<int64_t>(N), static_cast<int64_t>(OH), static_cast<int64_t>(OW),
+                             static_cast<int64_t>(KH * KW), static_cast<int64_t>(IC)});
+    auto r1 = std::make_shared<ov::op::v1::Reshape>(t1, reshape1_shape, false);
+
+    // [N, OH, OW, KH*KW, IC] → [N, OH, OW, IC, KH*KW]
+    auto perm2 = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{5}, std::vector<int64_t>{0, 1, 2, 4, 3});
+    auto t2 = std::make_shared<ov::op::v1::Transpose>(r1, perm2);
+
+    // flatten back to [N, OH, OW, IC*KH*KW]
+    auto r2_shape = ov::op::v0::Constant::create(
+        ov::element::i64, ov::Shape{4},
+        std::vector<int64_t>{static_cast<int64_t>(N), static_cast<int64_t>(OH), static_cast<int64_t>(OW),
+                             static_cast<int64_t>(IC * KH * KW)});
+    res = std::make_shared<ov::op::v1::Reshape>(t2, r2_shape, false);
+
+    if (!is_2D) {
+        // [N, 1, OW, IC * KW] -> [1, N, OW, IC * KW]
+        auto final_reshape_shape = ov::op::v0::Constant::create(
+            ov::element::i64, ov::Shape{4},
+            std::vector<int64_t>{1, static_cast<int64_t>(N), static_cast<int64_t>(OW), static_cast<int64_t>(IC * KW)});
+        res = std::make_shared<ov::op::v1::Reshape>(res, final_reshape_shape, false);
+    }
+
+    auto output_type = context.get_output_type();
+    if (res.get_element_type() != output_type) {
+        res = std::make_shared<ov::op::v0::Convert>(res, output_type);
+    }
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+OutputVector translate_im2col_3d(const NodeContext & context) {
+    num_inputs_check(context, 2, 2);
+    const int32_t * params = context.get_output_op_params();
+    int32_t s0 = params[0];
+    int32_t s1 = params[1];
+    int32_t s2 = params[2];
+    int32_t p0 = params[3];
+    int32_t p1 = params[4];
+    int32_t p2 = params[5];
+    int32_t d0 = params[6];
+    int32_t d1 = params[7];
+    int32_t d2 = params[8];
+    int32_t IC = params[9];
+
+    ov::Output<Node> image = process_view_input_new(context, 1);
+    const ov::Shape kernel_shape = context.get_input(0).get_shape();
+    const ov::Shape image_shape = image.get_shape();
+    const ov::Shape out_shape = context.get_output_shape().to_shape();
+
+    const size_t KD = kernel_shape[1];
+    const size_t KH = kernel_shape[2];
+    const size_t KW = kernel_shape[3];
+
+    const size_t N = image_shape[0] / static_cast<size_t>(IC);
+    const size_t ID = image_shape[1];
+    const size_t IH = image_shape[2];
+    const size_t IW = image_shape[3];
+
+    const size_t OD = (ID + 2 * p2 - d2 * (KD - 1) - 1) / s2 + 1;
+    const size_t OH = (IH + 2 * p1 - d1 * (KH - 1) - 1) / s1 + 1;
+    const size_t OW = (IW + 2 * p0 - d0 * (KW - 1) - 1) / s0 + 1;
+
+    if (N == 0 || OD == 0 || OH == 0 || OW == 0) {
+        auto output_type = context.get_output_type();
+        ov::Output<Node> res = ov::op::v0::Constant::create(
+            output_type, ov::Shape{N * OD, OH, OW, static_cast<size_t>(IC * KD * KH * KW)}, {});
+        return rename_outputs_with_suffix({res}, context.get_name());
+    }
+
+    const size_t IH_pad = IH + 2 * p1;
+    const size_t IW_pad = IW + 2 * p0;
+
+    auto image_5d_shape = ov::op::v0::Constant::create(
+        ov::element::i64, ov::Shape{5},
+        std::vector<int64_t>{static_cast<int64_t>(N), static_cast<int64_t>(IC), static_cast<int64_t>(ID),
+                             static_cast<int64_t>(IH), static_cast<int64_t>(IW)});
+    auto image_5d = std::make_shared<ov::op::v1::Reshape>(image, image_5d_shape, false);
+
+    auto pads_begin = ov::op::v0::Constant::create(
+        ov::element::i64, ov::Shape{5}, std::vector<int64_t>{0, 0, p2, p1, p0});
+    auto pads_end = ov::op::v0::Constant::create(
+        ov::element::i64, ov::Shape{5}, std::vector<int64_t>{0, 0, p2, p1, p0});
+    auto pad_3d = std::make_shared<ov::op::v1::Pad>(image_5d, pads_begin, pads_end, ov::op::PadMode::CONSTANT);
+
+    const ov::Shape patch_sizes = {KH, KW};
+    const ov::Strides strides = {static_cast<size_t>(s1), static_cast<size_t>(s0)};
+    const ov::Shape rates = {static_cast<size_t>(d1), static_cast<size_t>(d0)};
+
+    ov::OutputVector kd_slices;
+    kd_slices.reserve(KD);
+
+    auto perm_nod = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{5}, {0, 2, 1, 3, 4});
+    auto reshape_4d_shape = ov::op::v0::Constant::create(
+        ov::element::i64, ov::Shape{4},
+        std::vector<int64_t>{static_cast<int64_t>(N * OD), static_cast<int64_t>(IC),
+                             static_cast<int64_t>(IH_pad), static_cast<int64_t>(IW_pad)});
+    auto perm1 = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{4}, {0, 2, 3, 1});
+    auto r1_shape = ov::op::v0::Constant::create(
+        ov::element::i64, ov::Shape{5},
+        std::vector<int64_t>{static_cast<int64_t>(N * OD), static_cast<int64_t>(OH), static_cast<int64_t>(OW),
+                             static_cast<int64_t>(KH * KW), static_cast<int64_t>(IC)});
+    auto perm2 = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{5}, {0, 1, 2, 4, 3});
+    auto r2_shape = ov::op::v0::Constant::create(
+        ov::element::i64, ov::Shape{6},
+        std::vector<int64_t>{static_cast<int64_t>(N * OD), static_cast<int64_t>(OH), static_cast<int64_t>(OW),
+                             static_cast<int64_t>(IC), 1, static_cast<int64_t>(KH * KW)});
+    auto step_c = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{1}, {static_cast<int64_t>(s2)});
+    auto axes_c = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{1}, {2});
+
+    for (size_t ikd = 0; ikd < KD; ++ikd) {
+        auto start_c = ov::op::v0::Constant::create(
+            ov::element::i64, ov::Shape{1}, {static_cast<int64_t>(ikd * d2)});
+        auto stop_c = ov::op::v0::Constant::create(
+            ov::element::i64, ov::Shape{1}, {static_cast<int64_t>(ikd * d2 + OD * s2)});
+        auto depth_slice = std::make_shared<ov::op::v8::Slice>(pad_3d, start_c, stop_c, step_c, axes_c);
+        auto depth_slice_trans = std::make_shared<ov::op::v1::Transpose>(depth_slice, perm_nod);
+        auto depth_slice_4d = std::make_shared<ov::op::v1::Reshape>(depth_slice_trans, reshape_4d_shape, false);
+
+        auto patches = std::make_shared<ov::op::v3::ExtractImagePatches>(
+            depth_slice_4d, patch_sizes, strides, rates, ov::op::PadType::VALID);
+        auto t1 = std::make_shared<ov::op::v1::Transpose>(patches, perm1);
+        auto r1 = std::make_shared<ov::op::v1::Reshape>(t1, r1_shape, false);
+        auto t2 = std::make_shared<ov::op::v1::Transpose>(r1, perm2);
+        auto r2 = std::make_shared<ov::op::v1::Reshape>(t2, r2_shape, false);
+        kd_slices.push_back(r2);
+    }
+
+    ov::Output<Node> res;
+    if (KD == 1) {
+        res = kd_slices[0];
+    } else {
+        res = std::make_shared<ov::op::v0::Concat>(kd_slices, 4);
+    }
+
+    auto final_shape = ov::op::v0::Constant::create(
+        ov::element::i64, ov::Shape{4},
+        std::vector<int64_t>{static_cast<int64_t>(N * OD), static_cast<int64_t>(OH), static_cast<int64_t>(OW),
+                             static_cast<int64_t>(IC * KD * KH * KW)});
+    res = std::make_shared<ov::op::v1::Reshape>(res, final_shape, false);
+
+    auto output_type = context.get_output_type();
+    if (res.get_element_type() != output_type) {
+        res = std::make_shared<ov::op::v0::Convert>(res, output_type);
+    }
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/l2_norm.cpp

@@ -0,0 +1,61 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <memory>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/divide.hpp>
+#include <openvino/op/maximum.hpp>
+#include <openvino/op/multiply.hpp>
+#include <openvino/op/reduce_sum.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/sqrt.hpp>
+#include <openvino/op/squeeze.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_l2_norm(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+
+    auto input_node = process_view_input_new(context, 0);
+
+    if (context.get_op_case() == 1) {
+        // 92: [   128,    16,     1,     2] VIEW                 q_conv-1
+        //      [  6144,     1,     2,     1]            0: UNARY       conv_output_silu-1
+        // 93: [   128,    16,     1,     2] L2_NORM              q_conv_predelta-1
+        //      [   128,    16,     1,     2]            0: VIEW        q_conv-1
+        auto output_shape = context.get_output_shape().to_shape();
+        input_node = process_view_input(context, 0, output_shape[2] * output_shape[3], 3);
+        input_node =
+            std::make_shared<ov::op::v0::Squeeze>(input_node, ov::op::v0::Constant::create(ov::element::i64, {1}, {0}));
+
+        std::vector<int64_t> reshape_pattern = {0, 0, (int64_t) output_shape[2], (int64_t) output_shape[3]};
+        input_node = std::make_shared<ov::op::v1::Reshape>(
+            input_node, ov::op::v0::Constant::create(ov::element::i64, {4}, reshape_pattern), true);
+    }
+
+    auto squared = std::make_shared<ov::op::v1::Multiply>(input_node, input_node);
+
+    auto sum_squared = std::make_shared<ov::op::v1::ReduceSum>(
+        squared, ov::op::v0::Constant::create(ov::element::i64, ov::Shape{1}, {-1}), true);
+
+    auto l2_norm = std::make_shared<ov::op::v0::Sqrt>(sum_squared);
+
+    float eps;
+    memcpy(&eps, context.get_output_op_params(), sizeof(float));
+
+    auto eps_const = ov::op::v0::Constant::create(ov::element::f32, ov::Shape{1}, {eps});
+    auto clamped_norm = std::make_shared<ov::op::v1::Maximum>(l2_norm, eps_const);
+
+    auto res = std::make_shared<ov::op::v1::Divide>(input_node, clamped_norm);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/mean.cpp

@@ -0,0 +1,27 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <memory>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/reduce_mean.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_mean(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+
+    auto input = process_view_input_new(context, 0);
+    auto axis = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{1}, {-1});
+    auto res = std::make_shared<ov::op::v1::ReduceMean>(input, axis, true);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/moe_compressed.hpp

@@ -0,0 +1,90 @@
+// Copyright (C) 2018-2026 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+// Local mirror of OpenVINO's internal ov::op::internal::MOE and MOECompressed ops.
+//
+// The class bodies are provided by the linked libopenvino.so; only the declarations are
+// needed here so the backend can construct the node directly (same approach as
+// GatherMatmul and GatedDeltaNet). The class layout must stay in sync with
+//   openvino/src/core/dev_api/openvino/op/moe.hpp
+//   openvino/src/common/transformations/include/ov_ops/moe_compressed.hpp
+//
+// \note MOE op classes are under development and subject to change.
+
+#pragma once
+
+#include <optional>
+
+#include "openvino/core/type/element_type.hpp"
+#include "openvino/op/op.hpp"
+
+namespace ov::op::internal {
+
+class OPENVINO_API MOE : public ov::op::Op {
+public:
+    OPENVINO_OP("MOE")
+
+    MOE() = default;
+
+    MOE(const OutputVector & args) : Op(args) {}
+
+    enum class Expert_type { GEMM2_BIAS_SWIGLU_CLAMP, GEMM3_SWIGLU };
+
+    enum class Activation_type { SWIGLU, GEGLU_TANH, GEGLU_ERF };
+
+    struct Config {
+        Expert_type expert_type{ Expert_type::GEMM2_BIAS_SWIGLU_CLAMP };
+        float expert_alpha{ 0.0f };
+        float expert_beta{ 1.0f };
+        size_t gate_idx{ 0 };
+        Activation_type activation_type{ Activation_type::SWIGLU };
+    };
+
+    MOE(const OutputVector & args, const Config & config);
+
+    const Config & get_config() const;
+    void set_config(const Config & config);
+
+    bool visit_attributes(AttributeVisitor & visitor) override;
+    void validate_and_infer_types() override;
+    std::shared_ptr<Node> clone_with_new_inputs(const OutputVector & new_args) const override;
+
+private:
+    Config m_config;
+};
+
+class OPENVINO_API MOECompressed : public MOE {
+public:
+    OPENVINO_OP("MOECompressed", "", ov::op::internal::MOE)
+
+    MOECompressed() = default;
+
+    struct Config : public MOE::Config {
+        size_t hidden_size = 0;
+        size_t inter_size = 0;
+        size_t num_expert = 0;
+        size_t num_shared_expert = 0;
+        size_t top_k = 0;
+        // numeric_limits<size_t>::max() means per_channel compression (single group)
+        size_t group_size = 0;
+        bool has_batch_dim = false;
+        bool has_zp = false;
+        ov::element::Type out_type = ov::element::dynamic;
+        std::optional<float> scale_factor;
+    };
+
+    MOECompressed(const OutputVector & args, const Config & config);
+
+    const Config & get_config() const { return m_config; }
+
+    void set_scale_factor(float scale_factor) { m_config.scale_factor = scale_factor; }
+
+    bool visit_attributes(AttributeVisitor & visitor) override;
+    void validate_and_infer_types() override;
+    std::shared_ptr<Node> clone_with_new_inputs(const OutputVector & new_args) const override;
+
+protected:
+    Config m_config;
+};
+
+}  // namespace ov::op::internal

llama.cpp/ggml/src/ggml-openvino/openvino/op/mul_mat_id.cpp

@@ -0,0 +1,254 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+#include "gather_matmul.hpp"
+#include "ggml-openvino/ggml-openvino-extra.h"
+
+#include <cstdint>
+#include <cstring>
+#include <limits>
+#include <memory>
+#include <openvino/op/bitwise_and.hpp>
+#include <openvino/op/bitwise_right_shift.hpp>
+#include <openvino/op/broadcast.hpp>
+#include <openvino/op/concat.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/gather.hpp>
+#include <openvino/op/matmul.hpp>
+#include <openvino/op/multiply.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/shape_of.hpp>
+#include <openvino/op/slice.hpp>
+#include <openvino/op/transpose.hpp>
+#include <openvino/op/unsqueeze.hpp>
+#include <vector>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+namespace {
+
+std::shared_ptr<ov::op::v0::Constant> const_i64(const std::vector<int64_t> & values) {
+    return ov::op::v0::Constant::create(ov::element::i64, ov::Shape{values.size()}, values);
+}
+
+ov::Output<ov::Node> slice_axis(const ov::Output<ov::Node> & input, int64_t axis, int64_t begin, int64_t end) {
+    return std::make_shared<ov::op::v8::Slice>(input, const_i64({begin}), const_i64({end}), const_i64({1}),
+                                              const_i64({axis}));
+}
+
+// GGML tensors are rank 4, but stateful models drop the leading size-1 batch dim, so
+// activations and ids arrive one rank lower. Pick the trailing dims by actual rank.
+std::vector<int> trailing_dims(const ov::Output<ov::Node> & input, int count) {
+    const auto rank = input.get_partial_shape().rank();
+    FRONT_END_OP_CONVERSION_CHECK(rank.is_static(), "Expected static rank for MUL_MAT_ID input");
+    const int rank_len = static_cast<int>(rank.get_length());
+    FRONT_END_OP_CONVERSION_CHECK(rank_len >= count, "MUL_MAT_ID input rank is too low");
+
+    std::vector<int> dims;
+    for (int i = rank_len - count; i < rank_len; ++i) {
+        dims.push_back(i);
+    }
+    return dims;
+}
+
+ov::Output<ov::Node> static_shape_dims_or_shapeof(const ov::Output<ov::Node> & input,
+                                                  const std::vector<int> & dims) {
+    const auto & partial_shape = input.get_partial_shape();
+    if (partial_shape.is_static()) {
+        std::vector<int64_t> values;
+        values.reserve(dims.size());
+        for (const int64_t dim : dims) {
+            values.push_back(partial_shape[dim].get_length());
+        }
+        return const_i64(values);
+    }
+
+    auto shape = std::make_shared<ov::op::v3::ShapeOf>(input, ov::element::i64);
+    return get_dimensions(shape, dims);
+}
+
+ov::Output<ov::Node> translate_mul_mat_id_mxfp4_packed(const NodeContext & context,
+                                                       ov::Output<ov::Node> expert_weights,
+                                                       ov::Output<ov::Node> activations,
+                                                       ov::Output<ov::Node> ids) {
+    auto packed_shape = expert_weights.get_partial_shape().to_shape();
+    FRONT_END_OP_CONVERSION_CHECK(packed_shape.size() == 5 && packed_shape[4] == 17,
+                                  "Expected packed MXFP4 expert weights with shape [1, n_expert, m, k_blocks, 17]");
+
+    const int64_t n_expert = static_cast<int64_t>(packed_shape[1]);
+    const int64_t rows = static_cast<int64_t>(packed_shape[2]);
+    const int64_t k_blocks = static_cast<int64_t>(packed_shape[3]);
+    const int64_t qk = 32;
+    const int64_t cols = k_blocks * qk;
+
+    auto packed_shape_4d = const_i64({n_expert, rows, k_blocks, 17});
+    expert_weights = std::make_shared<ov::op::v1::Reshape>(expert_weights, packed_shape_4d, false);
+
+    auto activations_shape_4d = std::make_shared<ov::op::v3::ShapeOf>(activations, ov::element::i64);
+    auto ids_shape_4d = std::make_shared<ov::op::v3::ShapeOf>(ids, ov::element::i64);
+    auto activations_shape_3d = get_dimensions(activations_shape_4d, {1, 2, 3});
+    auto ids_shape_2d = get_dimensions(ids_shape_4d, {2, 3});
+
+    activations = std::make_shared<ov::op::v1::Reshape>(activations, activations_shape_3d, false);
+    ids = std::make_shared<ov::op::v1::Reshape>(ids, ids_shape_2d, false);
+    if (ids.get_element_type() != ov::element::i32 && ids.get_element_type() != ov::element::i64) {
+        ids = std::make_shared<ov::op::v0::Convert>(ids, ov::element::i32);
+    }
+
+    auto gather_axis = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{}, {0});
+
+    static const std::vector<float> f4e2m1_lut = {0.0f,  0.5f,  1.0f,  1.5f,  2.0f,  3.0f,  4.0f,  6.0f,
+                                                  -0.0f, -0.5f, -1.0f, -1.5f, -2.0f, -3.0f, -4.0f, -6.0f};
+    std::vector<float> e8m0_lut(256);
+    for (size_t i = 0; i < e8m0_lut.size(); ++i) {
+        uint32_t bits = static_cast<uint32_t>(i) << 23;
+        memcpy(&e8m0_lut[i], &bits, sizeof(float));
+    }
+    e8m0_lut[0] = std::numeric_limits<float>::min() / 2.0f;
+    e8m0_lut[255] = std::numeric_limits<float>::quiet_NaN();
+
+    auto f4_lut = ov::op::v0::Constant::create(ov::element::f32, ov::Shape{f4e2m1_lut.size()}, f4e2m1_lut);
+    auto scale_lut = ov::op::v0::Constant::create(ov::element::f32, ov::Shape{e8m0_lut.size()}, e8m0_lut);
+
+    auto selected_packed_weights = std::make_shared<ov::op::v8::Gather>(expert_weights, ids, gather_axis);
+    auto scale_byte = slice_axis(selected_packed_weights, 4, 0, 1);
+    auto qs = slice_axis(selected_packed_weights, 4, 1, 17);
+    auto low = std::make_shared<ov::op::v13::BitwiseAnd>(
+        qs, ov::op::v0::Constant::create(ov::element::u8, ov::Shape{}, {0x0F}), ov::op::AutoBroadcastType::NUMPY);
+    auto high_shift = std::make_shared<ov::op::v15::BitwiseRightShift>(
+        qs, ov::op::v0::Constant::create(ov::element::u8, ov::Shape{}, {4}), ov::op::AutoBroadcastType::NUMPY);
+    auto nibbles = std::make_shared<ov::op::v0::Concat>(ov::OutputVector{low, high_shift}, 4);
+    auto nibble_indices = std::make_shared<ov::op::v0::Convert>(nibbles, ov::element::i32);
+    auto weights_f32 = std::make_shared<ov::op::v8::Gather>(f4_lut, nibble_indices, gather_axis);
+
+    auto scale_indices = std::make_shared<ov::op::v0::Convert>(scale_byte, ov::element::i32);
+    auto scales_f32 = std::make_shared<ov::op::v8::Gather>(scale_lut, scale_indices, gather_axis);
+    ov::Output<ov::Node> selected_weights = std::make_shared<ov::op::v1::Multiply>(weights_f32, scales_f32,
+                                                                                  ov::op::AutoBroadcastType::NUMPY);
+
+    auto ids_shape = std::make_shared<ov::op::v3::ShapeOf>(ids, ov::element::i64);
+    auto selected_weights_target_dims = std::make_shared<ov::op::v0::Concat>(
+        ov::OutputVector{get_dimensions(ids_shape, {0, 1}), const_i64({rows, cols})}, 0);
+    selected_weights = std::make_shared<ov::op::v1::Reshape>(selected_weights, selected_weights_target_dims, false);
+
+    auto activations_shape = std::make_shared<ov::op::v3::ShapeOf>(activations, ov::element::i64);
+    ov::Output<ov::Node> acts_target_dims = std::make_shared<ov::op::v0::Concat>(
+        ov::OutputVector{
+            get_dimensions(activations_shape, {0}),
+            get_dimensions(ids_shape, {1}),
+            get_dimensions(activations_shape, {2}),
+        },
+        0);
+    ov::Output<ov::Node> acts_broadcasted =
+        std::make_shared<ov::op::v3::Broadcast>(activations, acts_target_dims, ov::op::BroadcastType::BIDIRECTIONAL);
+
+    auto activations_expanded = std::make_shared<ov::op::v0::Unsqueeze>(acts_broadcasted, const_i64({2}));
+    ov::Output<ov::Node> result =
+        std::make_shared<ov::op::v0::MatMul>(activations_expanded, selected_weights, false, true);
+
+    auto batch_dim = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+    auto row_dim = ov::op::v0::Constant::create(ov::element::i64, {1}, {rows});
+    auto result_target_dims = std::make_shared<ov::op::v0::Concat>(
+        ov::OutputVector{batch_dim, get_dimensions(ids_shape, {0, 1}), row_dim}, 0);
+    result = std::make_shared<ov::op::v1::Reshape>(result, result_target_dims, false);
+
+    const auto output_type = context.get_output_type();
+    if (result.get_element_type() != output_type) {
+        result = std::make_shared<ov::op::v0::Convert>(result, output_type);
+    }
+    return result;
+}
+
+}  // namespace
+
+OutputVector translate_mul_mat_id(const NodeContext & context) {
+    num_inputs_check(context, 3, 3);
+
+    auto expert_weights = process_view_input_new(context, 0);
+    auto activations = process_view_input_new(context, 1);
+    auto ids = process_view_input_new(context, 2);
+
+    if (activations.get_partial_shape().rank() == 3) {
+        activations = std::make_shared<ov::op::v0::Unsqueeze>(activations, const_i64({0}));
+    }
+    if (ids.get_partial_shape().rank() == 3) {
+        ids = std::make_shared<ov::op::v0::Unsqueeze>(ids, const_i64({0}));
+    }
+
+    if (expert_weights.get_element_type() == ov::element::u8 && expert_weights.get_partial_shape().rank().is_static() &&
+        expert_weights.get_partial_shape().rank().get_length() == 5) {
+        return rename_outputs_with_suffix({translate_mul_mat_id_mxfp4_packed(context, expert_weights, activations, ids)},
+                                          context.get_name());
+    }
+
+    // General (non-packed) path: dense F32/F16/BF16 weights, or the f16 dequantization chain for
+    // quantized MoE experts (see extract_quantized_weights / make_int4_weights / make_int8_weights in
+    // ggml-quants.cpp). Routed through ov::op::internal::GatherMatmul instead of a naive
+    // Gather+Broadcast+MatMul, so the selected expert's full weight matrix is never materialized per
+    // token. The CPU plugin's ConvertGatherMatmulToGatherMatmulCompressed pass (run during
+    // compile_model) fuses the dequantization chain feeding GatherMatmul's B input into a
+    // GatherMatmulCompressed node automatically, as long as MarkDequantization has marked the chain --
+    // see translate_session.cpp's apply_transformations for the MarkDequantization registration.
+    //
+    // OpenVINO sees GGML tensors in reversed dimension order:
+    //   activations: [1, n_tokens, n_used_or_1, k]
+    //   ids: [1, 1, n_tokens, n_used]
+    // expert_weights is either [1, n_expert, m, k] (4D, e.g. non-quantized weights without a
+    // pre-built extra) or already [n_expert, m, k] (3D, weights routed through
+    // process_weight_tensor) -- GatherMatmul's B input expects the latter.
+    auto expert_weights_rank = expert_weights.get_partial_shape().rank();
+    FRONT_END_OP_CONVERSION_CHECK(expert_weights_rank.is_static(),
+                                  "Expected static rank for MUL_MAT_ID expert weights");
+    if (expert_weights_rank.get_length() == 4) {
+        auto expert_weights_shape_3d = static_shape_dims_or_shapeof(expert_weights, {1, 2, 3});
+        expert_weights = std::make_shared<ov::op::v1::Reshape>(expert_weights, expert_weights_shape_3d, false);
+    }
+
+    auto activations_shape_3d = static_shape_dims_or_shapeof(activations, trailing_dims(activations, 3));
+    auto ids_shape_2d = static_shape_dims_or_shapeof(ids, trailing_dims(ids, 2));
+
+    activations = std::make_shared<ov::op::v1::Reshape>(activations, activations_shape_3d, false);
+    ids = std::make_shared<ov::op::v1::Reshape>(ids, ids_shape_2d, false);
+
+    if (ids.get_element_type() != ov::element::i32 && ids.get_element_type() != ov::element::i64) {
+        ids = std::make_shared<ov::op::v0::Convert>(ids, ov::element::i32);
+    }
+
+    const auto output_type = context.get_output_type();
+    const auto activations_type = ggml_openvino_is_gpu() ? ov::element::f16 : ov::element::f32;
+    if (activations.get_element_type() != activations_type) {
+        activations = std::make_shared<ov::op::v0::Convert>(activations, activations_type);
+    }
+
+    // GatherMatmul's A input is [n_used_or_1, n_tokens, k]; activations_3d is
+    // [n_tokens, n_used_or_1, k].
+    auto activations_transpose_order = const_i64({1, 0, 2});
+    ov::Output<ov::Node> activations_for_gather =
+        std::make_shared<ov::op::v1::Transpose>(activations, activations_transpose_order);
+
+    ov::Output<ov::Node> result = std::make_shared<ov::op::internal::GatherMatmul>(activations_for_gather, expert_weights, ids);
+
+    // result is [n_used, n_tokens, m]; GGML expects [1, n_tokens, n_used, m], except on the
+    // stateful path where the leading batch dim is dropped.
+    auto result_transpose_order = const_i64({1, 0, 2});
+    result = std::make_shared<ov::op::v1::Transpose>(result, result_transpose_order);
+    if (!context.is_stateful()) {
+        auto unsqueeze_axes = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
+        result = std::make_shared<ov::op::v0::Unsqueeze>(result, unsqueeze_axes);
+    }
+
+    if (result.get_element_type() != output_type) {
+        result = std::make_shared<ov::op::v0::Convert>(result, output_type);
+    }
+
+    return rename_outputs_with_suffix({result}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/mulmat.cpp

@@ -0,0 +1,92 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <climits>
+#include <cstdint>
+#include <memory>
+#include <openvino/core/node.hpp>
+#include <openvino/core/node_output.hpp>
+#include <openvino/op/broadcast.hpp>
+#include <openvino/op/concat.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/matmul.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/slice.hpp>
+#include <openvino/op/transpose.hpp>
+#include <openvino/op/unsqueeze.hpp>
+#include <openvino/op/util/op_types.hpp>
+#include <vector>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_mulmat(const NodeContext & context) {
+    num_inputs_check(context, 2, 2);
+
+    int op_case = context.get_op_case();
+
+    ov::Output<ov::Node> B = process_view_input_new(context, 0);
+    ov::Output<ov::Node> A = process_view_input_new(context, 1);
+
+    if (A.get_element_type() != B.get_element_type()) {
+        B = std::make_shared<ov::op::v0::Convert>(B, context.get_input_type(1));
+    }
+
+    auto B_shape = context.get_input_shape(0).to_shape();
+    auto A_shape = context.get_input_shape(1).to_shape();
+    int64_t A_batch = A_shape[1];
+    int64_t B_batch = B_shape[1];
+
+    auto A_batch_larger = A_batch > B_batch;
+    auto batch_large = A_batch_larger ? A_batch : B_batch;
+    auto batch_small = A_batch_larger ? B_batch : A_batch;
+
+    Output<Node> Z = A_batch_larger ? B : A;
+    auto Z_shape = A_batch_larger ? B_shape : A_shape;
+    int64_t factor = batch_large / batch_small;
+    if (factor > 1 && batch_small > 1) {
+        auto batch_large_node = ov::op::v0::Constant::create(ov::element::i64, {1}, std::vector<int64_t>{batch_large});
+        auto batch_small_node = ov::op::v0::Constant::create(ov::element::i64, {1}, std::vector<int64_t>{batch_small});
+        auto factor_node = ov::op::v0::Constant::create(ov::element::i64, {1}, std::vector<int64_t>{factor});
+
+        auto unsqueeze_axes = ov::op::v0::Constant::create(ov::element::i64, Shape{}, {2});
+        auto Z_unsqueezed = std::make_shared<ov::op::v0::Unsqueeze>(Z, unsqueeze_axes);
+
+        auto broadcast_shape = ov::op::v0::Constant::create(
+            ov::element::i64, {5}, {(int64_t) 1, (int64_t) 1, factor, (int64_t) 1, (int64_t) 1});
+        auto new_Z_shape = ov::op::v0::Constant::create(ov::element::i64, {4},
+                                                        {(int64_t) 0, batch_large, (int64_t) -1, (int64_t) Z_shape[3]});
+        if (op_case == 2) {
+            new_Z_shape = ov::op::v0::Constant::create(ov::element::i64, {4},
+                                                       {(int64_t) 0, batch_large, (int64_t) Z_shape[2], (int64_t) -1});
+        }
+
+        auto Z_broadcasted = std::make_shared<ov::op::v3::Broadcast>(Z_unsqueezed, broadcast_shape,
+                                                                     ov::op::BroadcastType::BIDIRECTIONAL);
+        Z = std::make_shared<ov::op::v1::Reshape>(Z_broadcasted, new_Z_shape, true);
+    }
+    if (A_batch_larger) {
+        B = Z;
+    } else {
+        A = Z;
+    }
+
+    bool transpose_b = true;
+    ov::Output<Node> res = std::make_shared<ov::op::v0::MatMul>(A, B, false, transpose_b);
+
+    const auto output_type = context.get_output_type();
+    if (res.get_element_type() != output_type) {
+        res = std::make_shared<ov::op::v0::Convert>(res, output_type);
+    }
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/norm.cpp

@@ -0,0 +1,31 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <cstring>
+#include <memory>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/mvn.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_norm(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+
+    auto input_node = process_view_input_new(context, 0);
+    float eps;
+    memcpy(&eps, context.get_output_op_params(), sizeof(float));
+
+    auto axes = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{1}, {-1});
+    auto res = std::make_shared<ov::op::v6::MVN>(input_node, axes, true, eps, ov::op::MVNEpsMode::INSIDE_SQRT);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/pad.cpp

@@ -0,0 +1,94 @@
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <array>
+#include <cstdint>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/gather.hpp>
+#include <openvino/op/pad.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/shape_of.hpp>
+#include <utility>
+#include <vector>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+namespace {
+
+ov::Output<ov::Node> translate_circular_pad(ov::Output<ov::Node> input,
+                                            const std::array<int32_t, 8> & pads,
+                                            const ov::Shape & input_shape) {
+    ov::Output<ov::Node> result = std::move(input);
+
+    const std::array<int32_t, 4> pads_begin = {pads[6], pads[4], pads[2], pads[0]};
+    const std::array<int32_t, 4> pads_end = {pads[7], pads[5], pads[3], pads[1]};
+
+    for (size_t axis = 0; axis < input_shape.size(); ++axis) {
+        const int64_t input_dim = static_cast<int64_t>(input_shape[axis]);
+        const int64_t pad_begin = pads_begin[axis];
+        const int64_t pad_end = pads_end[axis];
+
+        if (pad_begin == 0 && pad_end == 0) {
+            continue;
+        }
+
+        FRONT_END_CHECK_IMPLEMENTED(input_dim > 0, "Circular PAD requires static non-zero input dimensions");
+
+        std::vector<int64_t> indices(static_cast<size_t>(input_dim + pad_begin + pad_end));
+        for (int64_t index = 0; index < static_cast<int64_t>(indices.size()); ++index) {
+            int64_t wrapped = (index - pad_begin) % input_dim;
+            if (wrapped < 0) {
+                wrapped += input_dim;
+            }
+            indices[static_cast<size_t>(index)] = wrapped;
+        }
+
+        auto gather_indices = ov::op::v0::Constant::create(ov::element::i64, {indices.size()}, indices);
+        auto gather_axis = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{}, {axis});
+        result = std::make_shared<ov::op::v8::Gather>(result, gather_indices, gather_axis);
+    }
+
+    return result;
+}
+
+}  // namespace
+
+OutputVector translate_pad(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+
+    auto input = process_view_input_new(context, 0);
+    if (context.get_input_shape(0) == context.get_output_shape()) {
+        return rename_outputs_with_suffix({input}, context.get_name());
+    }
+
+    const int32_t * op_params = context.get_output_op_params();
+    FRONT_END_CHECK_IMPLEMENTED(op_params != nullptr, "PAD requires output op params");
+
+    const std::array<int32_t, 8> pads = {op_params[0], op_params[1], op_params[2], op_params[3],
+                                         op_params[4], op_params[5], op_params[6], op_params[7]};
+    const bool circular = op_params[8] != 0;
+
+    if (circular) {
+        auto res = translate_circular_pad(input, pads, context.get_input_shape(0).to_shape());
+        return rename_outputs_with_suffix({res}, context.get_name());
+    }
+
+    const std::vector<int64_t> pads_begin = {pads[6], pads[4], pads[2], pads[0]};
+    const std::vector<int64_t> pads_end = {pads[7], pads[5], pads[3], pads[1]};
+
+    auto pads_begin_node = ov::op::v0::Constant::create(ov::element::i64, {pads_begin.size()}, pads_begin);
+    auto pads_end_node = ov::op::v0::Constant::create(ov::element::i64, {pads_end.size()}, pads_end);
+    auto pad_value = ov::op::v0::Constant::create(context.get_input_type(0), ov::Shape{}, {0});
+    auto res =
+        std::make_shared<ov::op::v1::Pad>(input, pads_begin_node, pads_end_node, pad_value, ov::op::PadMode::CONSTANT);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/permute.cpp

@@ -0,0 +1,159 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <climits>
+#include <cstdint>
+#include <memory>
+#include <openvino/core/node.hpp>
+#include <openvino/op/add.hpp>
+#include <openvino/op/concat.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/slice.hpp>
+#include <openvino/op/transpose.hpp>
+#include <vector>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_permute(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+
+    int op_case = context.get_op_case();
+    FRONT_END_CHECK_IMPLEMENTED(op_case != 0, "Unsupported PERMUTE case");
+    // op_case 1 is trivial permute
+    // op_case 2 is to permute Q. It has a preceding VIEW that reshapes Q to restore the sequqence dimension
+    // op_case 3 4 it to permute KV cache in the default layout
+    // op_case 5 6 is to permute V cache when `-fa off`, where v_trans=true
+
+    ov::Output<Node> res;
+    ov::Output<Node> src;
+    if (op_case == 3 || op_case == 4 || op_case == 5 || op_case == 6) {
+        src = context.get_input(0);
+    } else {
+        src = process_view_input_new(context, 0);
+    }
+    std::vector<int64_t> perm_values{0, 2, 1, 3};
+    const int32_t * op_params = context.get_output_op_params();
+    if (op_params != nullptr) {
+        for (size_t input_axis = 0; input_axis < perm_values.size(); ++input_axis) {
+            const size_t output_axis = static_cast<size_t>(op_params[input_axis]);
+            perm_values[perm_values.size() - 1 - output_axis] =
+                static_cast<int64_t>(perm_values.size() - 1 - input_axis);
+        }
+    }
+    if (op_case == 1 || context.is_stateful()) {
+        // The stateful path carries hidden-state tensors in a rank-3 layout (the
+        // leading batch dim is dropped, e.g. Gemma4's per-layer-embedding path). The
+        // perm above is rank-4; when the actual input is rank-3, drop the batch axis
+        // (perm[0], which is always the identity 0 here) and shift the rest down by 1
+        // so the transpose order matches the input rank.
+        std::vector<int64_t> perm_used = perm_values;
+        const auto & src_ps = src.get_partial_shape();
+        if (src_ps.rank().is_static() && src_ps.rank().get_length() == 3 && perm_values.size() == 4 &&
+            perm_values[0] == 0) {
+            perm_used = {perm_values[1] - 1, perm_values[2] - 1, perm_values[3] - 1};
+        }
+        auto perm = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{perm_used.size()}, perm_used);
+        res = std::make_shared<ov::op::v1::Transpose>(src, perm);
+    } else if (op_case == 2) {
+        auto perm = ov::op::v0::Constant::create(ov::element::i64, {4}, perm_values);
+        auto output_shape = context.get_output_shape().to_shape();
+        auto n_heads = ov::op::v0::Constant::create(ov::element::i64, {1}, {output_shape[1]});
+        auto head_size = ov::op::v0::Constant::create(ov::element::i64, {1}, {output_shape[3]});
+        auto n_seq_active = context.has_input("n_seq_active") ?
+                                context.get_input("n_seq_active") :
+                                ov::op::v0::Constant::create(ov::element::i64, {1}, {output_shape[0]});
+        auto neg_one = ov::op::v0::Constant::create(ov::element::i64, {1}, {-1});
+
+        auto new_shape =
+            std::make_shared<ov::op::v0::Concat>(ov::OutputVector{n_seq_active, neg_one, n_heads, head_size}, 0);
+
+        // // Alternative
+        // auto zero = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
+        // auto new_shape = std::make_shared<ov::op::v0::Concat>(ov::OutputVector{n_seq_active, neg_one, zero, zero}, 0);
+
+        auto reshaped = std::make_shared<ov::op::v1::Reshape>(src, new_shape, true);
+        res = std::make_shared<ov::op::v1::Transpose>(reshaped, perm);
+    } else {
+        auto perm = ov::op::v0::Constant::create(ov::element::i64, {4}, perm_values);
+        auto cache_shape = src.get_partial_shape();
+        auto output_shape = context.get_output_shape().to_shape();
+        int64_t head_size = output_shape[3];
+        int64_t n_heads = output_shape[1];
+        if (op_case == 5 || op_case == 6) {
+            head_size = output_shape[2];
+            n_heads = output_shape[1];
+        }
+        int64_t ctx_per_seq = cache_shape[2].is_static() ? cache_shape[2].get_length() : -1;
+        int64_t n_seq = cache_shape[1].get_length();
+
+        Output<Node> attention_size;
+        if (!context.has_input("attention_size")) {
+            attention_size = ov::op::v0::Constant::create(ov::element::i64, {1}, {output_shape[2]});
+        } else if (op_case == 3 || op_case == 5) {
+            attention_size = context.get_input("attention_size");
+        } else {
+            attention_size = context.get_input("attention_size_swa");
+        }
+
+        Output<Node> seq_active_start;
+        Output<Node> seq_active_end;
+        if (context.has_input("seq_active_start")) {
+            seq_active_start = context.get_input("seq_active_start");
+            seq_active_end = context.get_input("seq_active_end");
+        } else {
+            int64_t n_seq_active = output_shape[0];
+            size_t offset = *((size_t *) context.get_input_op_params(0));
+            int64_t seq_active_start_val = offset / context.get_input_stride(0)[0];
+            int64_t seq_active_end_val = seq_active_start_val + n_seq_active;
+            seq_active_start = ov::op::v0::Constant::create(ov::element::i64, {1}, {seq_active_start_val});
+            seq_active_end = ov::op::v0::Constant::create(ov::element::i64, {1}, {seq_active_end_val});
+        }
+
+        // 1. reshape to [n_seq, ctx_per_seq, n_heads, head_size] (for `-fa off` [n_seq, n_heads, head_size, ctx_per_seq])
+        // 2. slice out the active sequences
+        // 3. slice out the attention part in each sequence
+        // 4. permute (skip for `-fa off`)
+        auto zero = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
+        auto one = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+
+        if (op_case == 3 || op_case == 4) {
+            auto src_reshaped = std::make_shared<ov::op::v1::Reshape>(
+                src, ov::op::v0::Constant::create(ov::element::i64, {4}, {n_seq, ctx_per_seq, n_heads, head_size}),
+                false);
+            ov::Output<ov::Node> after_seq_slice;
+            if (n_seq == 1) {
+                after_seq_slice = src_reshaped;
+            } else {
+                after_seq_slice =
+                    std::make_shared<ov::op::v8::Slice>(src_reshaped, seq_active_start, seq_active_end, one, zero);
+            }
+            auto slice2 = std::make_shared<ov::op::v8::Slice>(after_seq_slice, zero, attention_size, one, one);
+            res = std::make_shared<ov::op::v1::Transpose>(slice2, perm);
+        } else {
+            auto three = ov::op::v0::Constant::create(ov::element::i64, {1}, {3});
+            auto src_reshaped = std::make_shared<ov::op::v1::Reshape>(
+                src, ov::op::v0::Constant::create(ov::element::i64, {4}, {n_seq, n_heads, head_size, ctx_per_seq}),
+                false);
+            ov::Output<ov::Node> after_seq_slice;
+            if (n_seq == 1) {
+                after_seq_slice = src_reshaped;
+            } else {
+                after_seq_slice =
+                    std::make_shared<ov::op::v8::Slice>(src_reshaped, seq_active_start, seq_active_end, one, zero);
+            }
+            auto slice2 = std::make_shared<ov::op::v8::Slice>(after_seq_slice, zero, attention_size, one, three);
+            res = slice2;
+        }
+    }
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/pool_2d.cpp

@@ -0,0 +1,53 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <openvino/op/avg_pool.hpp>
+#include <openvino/op/max_pool.hpp>
+#include <openvino/op/convert.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_pool_2d(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+    const int32_t * params = context.get_output_op_params();
+
+    const int k0 = params[1];
+    const int k1 = params[2];
+    const int s0 = params[3];
+    const int s1 = params[4];
+    const int p0 = params[5];
+    const int p1 = params[6];
+
+    const int op_case = context.get_op_case();
+    ov::Output<Node> input = context.get_input(0);
+    ov::Strides strides{static_cast<size_t>(s1), static_cast<size_t>(s0)};
+    ov::Shape pads_begin{static_cast<size_t>(p1), static_cast<size_t>(p0)};
+    ov::Shape pads_end{static_cast<size_t>(p1), static_cast<size_t>(p0)};
+    ov::Shape kernel{static_cast<size_t>(k1), static_cast<size_t>(k0)};
+    ov::Output<Node> res;
+
+    switch (op_case) {
+    case 1:  // GGML_OP_POOL_MAX
+    {
+        res = std::make_shared<ov::op::v1::MaxPool>(input, strides, pads_begin, pads_end, kernel);
+        break;
+    }
+    case 2:  // GGML_OP_POOL_AVG
+    {
+        res = std::make_shared<ov::op::v1::AvgPool>(input, strides, pads_begin, pads_end, kernel, false);
+        break;
+    }
+    default:
+        break;
+    }
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/repeat.cpp

@@ -0,0 +1,47 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <memory>
+#include <openvino/op/broadcast.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/divide.hpp>
+#include <openvino/op/shape_of.hpp>
+#include <openvino/op/tile.hpp>
+#include <vector>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+// GGML_OP_REPEAT tiles src[0] to fill the destination shape. Every destination
+// dimension is an integer multiple of the corresponding source dimension.
+OutputVector translate_repeat(const NodeContext & context) {
+    num_inputs_check(context, 1, 2);
+
+    auto input = process_view_input_new(context, 0);
+
+    const auto input_shape = context.get_input_shape(0).to_shape();
+    const auto output_shape = context.get_output_shape().to_shape();
+
+    std::vector<int64_t> repeats(4, 1);
+    for (size_t axis = 0; axis < 4; ++axis) {
+        const int64_t input_dim = input_shape[axis];
+        const int64_t output_dim = output_shape[axis];
+
+        FRONT_END_OP_CONVERSION_CHECK(input_dim > 0 && output_dim > 0 && output_dim % input_dim == 0,
+                                      "REPEAT input shape ", input_shape, " cannot tile to match ", output_shape);
+
+        repeats[axis] = output_dim / input_dim;
+    }
+
+    auto repeats_node = ov::op::v0::Constant::create(ov::element::i64, {repeats.size()}, repeats);
+    ov::Output<ov::Node> res = std::make_shared<ov::op::v0::Tile>(input, repeats_node);
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/reshape.cpp

@@ -0,0 +1,86 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <cstdint>
+#include <memory>
+#include <openvino/core/node.hpp>
+#include <openvino/core/node_output.hpp>
+#include <openvino/frontend/exception.hpp>
+#include <openvino/op/concat.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/reshape.hpp>
+#include <vector>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_reshape(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+    if (context.get_input(0).get_partial_shape().is_static() &&
+        context.get_input_shape(0) == context.get_output_shape()) {
+        return {context.get_input(0)};
+    }
+
+    const int op_case = context.get_op_case();
+    const auto output_shape = context.get_output_shape().to_shape();
+    std::vector<int64_t> shape(output_shape.begin(), output_shape.end());
+    std::shared_ptr<ov::Node> new_shape_node;
+    switch (op_case) {
+    case 0:
+        break;
+    case 1:
+    case 9:
+        shape[1] = -1;
+        if (context.is_stateful() && op_case == 1) {
+            shape.erase(shape.begin());
+        }
+        break;
+    case 2:
+    case 3:
+        shape[2] = -1;
+        if (op_case == 3) {
+            shape[3] = 1;
+        }
+        break;
+    case 4:
+        return {context.get_input(0).get_node_shared_ptr()->input_value(0)};
+    case 5:
+    case 7:
+        shape = {1, 1, -1, shape[3]};
+        if (context.is_stateful() && op_case == 5) {
+            shape.erase(shape.begin());
+        }
+        break;
+    case 6:
+        // Recurrent inputs keep the active sequence count separate from the token count.
+        if (context.has_input("s_copy_active_slot_len")) {
+            auto n_slot_active_len = context.get_input("s_copy_active_slot_len");
+            auto emb_size = ov::op::v0::Constant::create(ov::element::i64, {1}, {shape[3]});
+            auto one = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+            auto neg_one = ov::op::v0::Constant::create(ov::element::i64, {1}, {-1});
+            new_shape_node =
+                std::make_shared<ov::op::v0::Concat>(ov::OutputVector{one, n_slot_active_len, neg_one, emb_size}, 0);
+        } else {
+            shape = {1, 1, -1, shape[3]};
+        }
+        break;
+    case 8:
+        shape[0] = -1;
+        break;
+    default:
+        FRONT_END_OP_CONVERSION_CHECK(false, "Unsupported RESHAPE case: ", op_case);
+    }
+    if (!new_shape_node) {
+        new_shape_node = ov::op::v0::Constant::create(ov::element::i64, {shape.size()}, shape);
+    }
+    auto res = std::make_shared<ov::op::v1::Reshape>(context.get_input(0), new_shape_node, false);
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/rms_norm.cpp

@@ -0,0 +1,89 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <memory>
+#include <openvino/op/add.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/divide.hpp>
+#include <openvino/op/multiply.hpp>
+#include <openvino/op/negative.hpp>
+#include <openvino/op/power.hpp>
+#include <openvino/op/reduce_mean.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/slice.hpp>
+#include <openvino/op/sqrt.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_rms_norm(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+
+    auto op_case = context.get_op_case();
+
+    ov::Output<ov::Node> input_node;
+    if (op_case == 3) {
+        // Flatten sequence and token dimensions to match the gate layout.
+        auto input_shape = context.get_input_shape(0).to_shape();
+        input_node = std::make_shared<ov::op::v1::Reshape>(
+            context.get_input(0),
+            ov::op::v0::Constant::create(
+                ov::element::i64, {4}, std::vector<int64_t>{1, -1, (int64_t) input_shape[2], (int64_t) input_shape[3]}),
+            false);
+    } else if (op_case == 1) {
+        input_node = process_view_input_new(context, 0);
+    } else if (op_case == 2) {
+        auto ssm_state_size = context.get_ssm_state_size();
+        // The GDN op packs [attn | new_state] along the row axis; the state occupies the last
+        // ssm_state_size * n_seqs rows. Slice it off (scaling by the active sequence count) to keep
+        // just the attention output.
+        ov::Output<ov::Node> state_end;
+        if (context.has_input("s_copy_active_slot_len")) {
+            auto len = context.get_input("s_copy_active_slot_len");
+            auto state_rows = std::make_shared<ov::op::v1::Multiply>(
+                ov::op::v0::Constant::create(ov::element::i64, {1}, {ssm_state_size}), len);
+            state_end = std::make_shared<ov::op::v0::Negative>(state_rows);
+        } else {
+            state_end = ov::op::v0::Constant::create(ov::element::i64, {1}, {-ssm_state_size});
+        }
+        auto gdn_attn_output = std::make_shared<ov::op::v8::Slice>(
+            context.get_input(0), ov::op::v0::Constant::create(ov::element::i64, {1}, {0}), state_end,
+            ov::op::v0::Constant::create(ov::element::i64, {1}, {1}),
+            ov::op::v0::Constant::create(ov::element::i64, {1}, {2}));
+
+        auto input_shape = context.get_input_shape(0).to_shape();
+        input_node = std::make_shared<ov::op::v1::Reshape>(
+            gdn_attn_output,
+            ov::op::v0::Constant::create(
+                ov::element::i64, {4}, std::vector<int64_t>{1, -1, (int64_t) input_shape[2], (int64_t) input_shape[3]}),
+            false);
+
+    } else {
+        input_node = process_view_input_new(context, 0);
+    }
+    auto square = std::make_shared<ov::op::v1::Multiply>(input_node, input_node);
+
+    auto mean = std::make_shared<ov::op::v1::ReduceMean>(
+        square, ov::op::v0::Constant::create(ov::element::i64, ov::Shape{1}, {-1}), true);
+
+    float eps;
+    memcpy(&eps, context.get_output_op_params(), sizeof(float));
+
+    auto rms = std::make_shared<ov::op::v0::Sqrt>(
+        std::make_shared<ov::op::v1::Add>(mean, ov::op::v0::Constant::create(ov::element::f32, ov::Shape{1}, {eps})));
+
+    auto reciprocal =
+        std::make_shared<ov::op::v1::Divide>(ov::op::v0::Constant::create(ov::element::f32, ov::Shape{1}, {1.0f}), rms);
+
+    auto res = std::make_shared<ov::op::v1::Multiply>(input_node, reciprocal);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/roll.cpp

@@ -0,0 +1,36 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <openvino/op/roll.hpp>
+#include <openvino/op/constant.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_roll(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+    const int32_t * params = context.get_output_op_params();
+
+    int64_t s0 = params[0];
+    int64_t s1 = params[1];
+    int64_t s2 = params[2];
+    int64_t s3 = params[3];
+
+    auto input = context.get_input(0);
+
+    auto shift = ov::op::v0::Constant::create(
+        ov::element::i64, ov::Shape{4}, std::vector<int64_t>{s3, s2, s1, s0});
+    auto axes  = ov::op::v0::Constant::create(
+        ov::element::i64, ov::Shape{4}, std::vector<int64_t>{0, 1, 2, 3});
+
+    auto roll = std::make_shared<ov::op::v7::Roll>(input, shift, axes);
+    return rename_outputs_with_suffix({roll}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/rope.cpp

@@ -0,0 +1,245 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <cstdint>
+#include <memory>
+#include <openvino/core/node.hpp>
+#include <openvino/core/node_output.hpp>
+#include <openvino/op/add.hpp>
+#include <openvino/op/broadcast.hpp>
+#include <openvino/op/concat.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/multiply.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/slice.hpp>
+#include <openvino/op/split.hpp>
+#include <openvino/op/subtract.hpp>
+#include <openvino/op/unsqueeze.hpp>
+#include <openvino/op/variadic_split.hpp>
+#include <vector>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_rope(const NodeContext & context) {
+    num_inputs_check(context, 2, 3);
+
+    int op_case = context.get_op_case();
+
+    ov::Output<Node> res;
+
+    auto data_node = process_view_input_new(context, 0).get_node_shared_ptr();
+    auto output_shape = context.get_output_shape().to_shape();
+    int32_t * op_params = context.get_output_op_params();
+    const int mode = op_case;
+    const int64_t head_dim = static_cast<int64_t>(output_shape[3]);
+    const int64_t configured_n_dims = static_cast<int64_t>(op_params[1]);
+    const int64_t n_dims = configured_n_dims == 0 ? head_dim : configured_n_dims;
+    const int64_t n_offs = static_cast<int64_t>(op_params[15]);
+
+    constexpr int TYPE_NORMAL = 0;
+    constexpr int TYPE_NEOX = 1;
+    constexpr int TYPE_IMROPE = 2;
+    constexpr int TYPE_VISION = 3;
+    constexpr int TYPE_MROPE = 4;
+
+    Output<Node> cos_theta_node;
+    Output<Node> sin_theta_node;
+    if (context.has_input("rope_cos")) {
+        cos_theta_node = context.get_input("rope_cos");
+        sin_theta_node = context.get_input("rope_sin");
+    } else {
+        std::string cache_key = "rope_sin_cos";
+        for (int i = 0; i < 15; i++) {
+            cache_key += "_" + std::to_string(op_params[i]);
+        }
+        if (context.get_input_size() == 3) {
+            cache_key += "_ff_" + context.get_input_names()[2];
+        }
+        if (context.has_input(cache_key + "_cos")) {
+            cos_theta_node = context.get_input(cache_key + "_cos");
+            sin_theta_node = context.get_input(cache_key + "_sin");
+        } else {
+            auto inp_pos = context.get_input(1).get_node_shared_ptr();
+            std::shared_ptr<ov::Node> rope_freqs_weight;
+            if (context.get_input_size() == 3) {
+                rope_freqs_weight = context.get_input(2).get_node_shared_ptr();
+            }
+            auto sin_cos = make_sin_cos(op_params, inp_pos, rope_freqs_weight, mode, false, head_dim);
+            sin_theta_node = sin_cos.first;
+            cos_theta_node = sin_cos.second;
+            context.put_shared(cache_key + "_cos", cos_theta_node);
+            context.put_shared(cache_key + "_sin", sin_theta_node);
+        }
+    }
+
+    auto output_type = context.get_output_type();
+    if (data_node->get_element_type() != ov::element::f32) {
+        data_node = std::make_shared<ov::op::v0::Convert>(data_node, ov::element::f32);
+    }
+
+    const int64_t total_rope_dims = (mode == TYPE_VISION) ? (2 * n_dims) : n_dims;
+    FRONT_END_OP_CONVERSION_CHECK(n_offs >= 0 && (n_offs % 2 == 0),
+                                  "ROPE expects non-negative even n_offs");
+    FRONT_END_OP_CONVERSION_CHECK(n_dims > 0 && total_rope_dims + n_offs <= head_dim && (n_dims % 2 == 0),
+                                  "ROPE expects even n_dims with total_rope_dims + n_offs <= head_dim");
+
+    // RoPEFusionFlux requires rank_equals(4) on x, t_cos and t_sin. The cos/sin
+    // tables are already built rank-4 ([1, S, 1, head_size/2]) for both modes. In
+    // stateful mode the data arrives rank-3 ([S, n_heads, head_size]), so lift it
+    // to rank-4 ([1, S, n_heads, head_size]) here. Stateful RoPE already produced
+    // rank-4 output, so downstream attention is unaffected.
+    if (context.is_stateful()) {
+        const int64_t batch = static_cast<int64_t>(output_shape[0]);
+        auto r4_shape = ov::op::v0::Constant::create(
+            ov::element::i64, {4},
+            std::vector<int64_t>{batch, -1, (int64_t) output_shape[2], (int64_t) output_shape[3]});
+        data_node = std::make_shared<ov::op::v1::Reshape>(data_node, r4_shape, false);
+    }
+    // For TYPE_NORMAL rope (both stateful and stateless) we emit the Flux-style
+    // interleaved pattern below so the GPU plugin's RoPEFusionFlux matcher folds it
+    // into ov::op::internal::RoPE.
+    if (mode == TYPE_NORMAL) {
+        auto axis_last = ov::op::v0::Constant::create(ov::element::i64, {1}, {-1});
+        auto step_one = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+
+        const int64_t batch = static_cast<int64_t>(output_shape[0]);
+        const int64_t n_heads = static_cast<int64_t>(output_shape[2]);
+        const int64_t half = n_dims / 2;
+        auto rot_start = ov::op::v0::Constant::create(ov::element::i64, {1}, {n_offs});
+        auto rot_end = ov::op::v0::Constant::create(ov::element::i64, {1}, {n_offs + n_dims});
+        auto rot_data = std::make_shared<ov::op::v8::Slice>(data_node, rot_start, rot_end, step_one, axis_last);
+
+        auto neg_one_f = ov::op::v0::Constant::create(data_node->get_element_type(), ov::Shape{}, {-1.0f});
+
+        auto paired_shape = ov::op::v0::Constant::create(
+            ov::element::i64, {5}, std::vector<int64_t>{batch, -1, n_heads, half, 2});
+        auto x_paired = std::make_shared<ov::op::v1::Reshape>(rot_data, paired_shape, false);
+
+        auto split_axis = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{}, {-1});
+        auto data_split = std::make_shared<ov::op::v1::Split>(x_paired, split_axis, 2);
+        Output<Node> x0 = data_split->outputs()[0];
+        Output<Node> x1 = data_split->outputs()[1];
+
+        auto x1_neg = std::make_shared<ov::op::v1::Multiply>(x1, neg_one_f);
+        auto x_rotated_paired = std::make_shared<ov::op::v0::Concat>(ov::OutputVector{x1_neg, x0}, -1);
+
+        auto flat_shape =
+            ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{batch, -1, n_heads, n_dims});
+        auto x_rotated =
+            std::make_shared<ov::op::v1::Reshape>(x_rotated_paired, flat_shape, false);
+
+        // Expand cos/sin from [..., n_dims/2] to [..., n_dims] by repeating each
+        // entry twice. Use special_zero on the final Reshape so the seq dim passes
+        // through dynamically. Final rank is 4 to satisfy the matcher's predicate.
+        auto expand_cos_sin = [&](const Output<Node>& cs) {
+            auto cs_unsq = std::make_shared<ov::op::v0::Unsqueeze>(
+                cs, ov::op::v0::Constant::create(ov::element::i64, {1}, {-1}));
+            auto bcast_target = ov::op::v0::Constant::create(
+                ov::element::i64, {5}, std::vector<int64_t>{1, 1, 1, half, 2});
+            auto bcast = std::make_shared<ov::op::v3::Broadcast>(
+                cs_unsq, bcast_target, ov::op::BroadcastType::BIDIRECTIONAL);
+            auto flat = ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{0, 0, 0, n_dims});
+            return std::make_shared<ov::op::v1::Reshape>(bcast, flat, true);
+        };
+        Output<Node> cos_full = expand_cos_sin(cos_theta_node);
+        Output<Node> sin_full = expand_cos_sin(sin_theta_node);
+
+        auto y1 = std::make_shared<ov::op::v1::Multiply>(rot_data, cos_full);
+        auto y2 = std::make_shared<ov::op::v1::Multiply>(x_rotated, sin_full);
+        auto rotated = std::make_shared<ov::op::v1::Add>(y1, y2);
+
+        ov::OutputVector concat_parts;
+        if (n_offs > 0) {
+            auto head_start = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
+            auto head_end = ov::op::v0::Constant::create(ov::element::i64, {1}, {n_offs});
+            auto head = std::make_shared<ov::op::v8::Slice>(data_node, head_start, head_end, step_one, axis_last);
+            concat_parts.push_back(head);
+        }
+        concat_parts.push_back(rotated);
+        if (n_offs + n_dims < head_dim) {
+            auto tail_start = ov::op::v0::Constant::create(ov::element::i64, {1}, {n_offs + n_dims});
+            auto tail_end = ov::op::v0::Constant::create(ov::element::i64, {1}, {head_dim});
+            auto tail = std::make_shared<ov::op::v8::Slice>(data_node, tail_start, tail_end, step_one, axis_last);
+            concat_parts.push_back(tail);
+        }
+        if (concat_parts.size() == 1) {
+            res = rotated;
+        } else {
+            res = std::make_shared<ov::op::v0::Concat>(concat_parts, -1);
+        }
+    } else if (mode == TYPE_NEOX || mode == TYPE_IMROPE || mode == TYPE_MROPE || mode == TYPE_VISION) {
+        const int64_t half = (mode == TYPE_VISION) ? n_dims : (n_dims / 2);
+        const int64_t rot_dims = 2 * half;
+
+        if (mode != TYPE_NEOX) {
+            auto cos_sin_shape = std::make_shared<ov::op::v0::Constant>(ov::element::i64, ov::Shape{4},
+                                                                        std::vector<int64_t>{1, -1, 1, half});
+            cos_theta_node = std::make_shared<ov::op::v1::Reshape>(cos_theta_node, cos_sin_shape, true);
+            sin_theta_node = std::make_shared<ov::op::v1::Reshape>(sin_theta_node, cos_sin_shape, true);
+        }
+
+        auto axis_last = ov::op::v0::Constant::create(ov::element::i64, {1}, {-1});
+        auto step_one = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+
+        Output<Node> rot_data = data_node;
+        if (n_offs > 0 || n_offs + rot_dims < head_dim) {
+            auto rot_start = ov::op::v0::Constant::create(ov::element::i64, {1}, {n_offs});
+            auto rot_end = ov::op::v0::Constant::create(ov::element::i64, {1}, {n_offs + rot_dims});
+            rot_data = std::make_shared<ov::op::v8::Slice>(data_node, rot_start, rot_end, step_one, axis_last);
+        }
+
+        auto neg_one_f = ov::op::v0::Constant::create(data_node->get_element_type(), ov::Shape{}, {-1.0f});
+
+        auto split_axis = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{}, {3});
+        auto split_lengths = ov::op::v0::Constant::create(ov::element::i64, {2}, {half, half});
+        auto data_split = std::make_shared<ov::op::v1::VariadicSplit>(rot_data, split_axis, split_lengths);
+        Output<Node> x1 = data_split->outputs()[0];
+        Output<Node> x2 = data_split->outputs()[1];
+
+        auto x2_neg = std::make_shared<ov::op::v1::Multiply>(x2, neg_one_f);
+        auto x_rotate_half = std::make_shared<ov::op::v0::Concat>(ov::OutputVector{x2_neg, x1}, -1);
+
+        auto cos_full = std::make_shared<ov::op::v0::Concat>(ov::OutputVector{cos_theta_node, cos_theta_node}, -1);
+        auto sin_full = std::make_shared<ov::op::v0::Concat>(ov::OutputVector{sin_theta_node, sin_theta_node}, -1);
+
+        auto y1 = std::make_shared<ov::op::v1::Multiply>(rot_data, cos_full);
+        auto y2 = std::make_shared<ov::op::v1::Multiply>(x_rotate_half, sin_full);
+        auto rotated = std::make_shared<ov::op::v1::Add>(y1, y2);
+
+        ov::OutputVector concat_parts;
+        if (n_offs > 0) {
+            auto head_start = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
+            auto head_end = ov::op::v0::Constant::create(ov::element::i64, {1}, {n_offs});
+            auto head = std::make_shared<ov::op::v8::Slice>(data_node, head_start, head_end, step_one, axis_last);
+            concat_parts.push_back(head);
+        }
+        concat_parts.push_back(rotated);
+        if (n_offs + rot_dims < head_dim) {
+            auto tail_start = ov::op::v0::Constant::create(ov::element::i64, {1}, {n_offs + rot_dims});
+            auto tail_end = ov::op::v0::Constant::create(ov::element::i64, {1}, {head_dim});
+            auto tail = std::make_shared<ov::op::v8::Slice>(data_node, tail_start, tail_end, step_one, axis_last);
+            concat_parts.push_back(tail);
+        }
+        if (concat_parts.size() == 1) {
+            res = rotated;
+        } else {
+            res = std::make_shared<ov::op::v0::Concat>(concat_parts, -1);
+        }
+    }
+
+    if (res.get_element_type() != output_type) {
+        res = std::make_shared<ov::op::v0::Convert>(res, output_type);
+    }
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/scale.cpp

@@ -0,0 +1,90 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <openvino/core/except.hpp>
+#include <openvino/op/add.hpp>
+#include <openvino/op/concat.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/equal.hpp>
+#include <openvino/op/gather.hpp>
+#include <openvino/op/greater_eq.hpp>
+#include <openvino/op/if.hpp>
+#include <openvino/op/less.hpp>
+#include <openvino/op/logical_or.hpp>
+#include <openvino/op/multiply.hpp>
+#include <openvino/op/range.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/shape_of.hpp>
+#include <openvino/op/slice.hpp>
+#include <openvino/op/squeeze.hpp>
+#include <openvino/op/unsqueeze.hpp>
+#include <vector>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_scale(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+
+    float scale;
+    float bias;
+    memcpy(&scale, (float *) context.get_output_op_params() + 0, sizeof(float));
+    memcpy(&bias, (float *) context.get_output_op_params() + 1, sizeof(float));
+
+    auto scale_node = std::make_shared<ov::op::v0::Constant>(ov::element::f32, ov::Shape{}, std::vector<float>{scale});
+
+    if (context.get_op_case() == 2) {
+        return {context.get_input(0)};
+    }
+
+    if (context.get_op_case() == 1 && context.has_input("cache_rs_reset_len")) {
+        auto cache_rs_reset_idx = context.get_input("cache_rs_reset_idx");
+        auto cache_rs_reset_len = context.get_input("cache_rs_reset_len");
+
+        auto cache_rs = context.get_input(0);
+
+        auto cache_shape = std::make_shared<ov::op::v3::ShapeOf>(cache_rs, ov::element::i64);
+        auto n_slots_1d = std::make_shared<ov::op::v8::Gather>(
+            cache_shape, ov::op::v0::Constant::create(ov::element::i64, ov::Shape{1}, {2}),
+            ov::op::v0::Constant::create(ov::element::i64, ov::Shape{}, {0}));
+        auto n_slots = std::make_shared<ov::op::v0::Squeeze>(n_slots_1d);
+
+        auto iota = std::make_shared<ov::op::v4::Range>(
+            ov::op::v0::Constant::create(ov::element::i64, ov::Shape{}, {0}), n_slots,
+            ov::op::v0::Constant::create(ov::element::i64, ov::Shape{}, {1}), ov::element::i64);
+
+        auto idx_plus_len = std::make_shared<ov::op::v1::Add>(cache_rs_reset_idx, cache_rs_reset_len);
+        auto less_than_idx = std::make_shared<ov::op::v1::Less>(iota, cache_rs_reset_idx);
+        auto greater_equal_idx_plus_len = std::make_shared<ov::op::v1::GreaterEqual>(iota, idx_plus_len);
+        auto keep_mask = std::make_shared<ov::op::v1::LogicalOr>(less_than_idx, greater_equal_idx_plus_len);
+
+        auto keep_mask_f32 = std::make_shared<ov::op::v0::Convert>(keep_mask, ov::element::f32);
+        auto keep_mask_reshape = std::make_shared<ov::op::v0::Unsqueeze>(
+            keep_mask_f32, ov::op::v0::Constant::create(ov::element::i64, ov::Shape{1}, {1}));
+
+        auto cleared_cache_rs = std::make_shared<ov::op::v1::Multiply>(cache_rs, keep_mask_reshape);
+        return rename_outputs_with_suffix({cleared_cache_rs}, context.get_name());
+    }
+
+    auto scaled = std::make_shared<ov::op::v1::Multiply>(context.get_input(0), scale_node);
+
+    std::shared_ptr<ov::Node> res;
+    if (bias != 0.0f) {
+        auto bias_node =
+            std::make_shared<ov::op::v0::Constant>(ov::element::f32, ov::Shape{}, std::vector<float>{bias});
+        res = std::make_shared<ov::op::v1::Add>(scaled, bias_node);
+    } else {
+        res = scaled;
+    }
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/set.cpp

@@ -0,0 +1,76 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <cstdint>
+#include <openvino/frontend/exception.hpp>
+#include <openvino/op/add.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/range.hpp>
+#include <openvino/op/reduce_prod.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/scatter_update.hpp>
+#include <openvino/op/shape_of.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+// GGML SET writes src1 into a view of src0 and returns the updated tensor.
+OutputVector translate_set(const NodeContext & context) {
+    num_inputs_check(context, 2, 2);
+
+    auto dst = process_view_input_new(context, 0);
+    auto src = process_view_input_new(context, 1);
+
+    src = std::make_shared<ov::op::v0::Convert>(src, context.get_output_type());
+
+    const auto dst_stride = context.get_input_stride(0);
+    FRONT_END_OP_CONVERSION_CHECK(dst_stride.size() >= 4, "SET requires 4D destination strides");
+
+    const auto * op_params = reinterpret_cast<const uint32_t *>(context.get_output_op_params());
+    const size_t offset = static_cast<size_t>(op_params[3]);
+
+    const size_t elem_size = dst_stride.back();
+    FRONT_END_OP_CONVERSION_CHECK(elem_size != 0 && offset % elem_size == 0,
+                                  "SET offset must be aligned to destination element size");
+
+    const int64_t offset_elems = static_cast<int64_t>(offset / elem_size);
+
+    auto dst_flat = std::make_shared<ov::op::v1::Reshape>(
+        dst,
+        ov::op::v0::Constant::create(ov::element::i64, {1}, {-1}),
+        false);
+
+    auto src_flat = std::make_shared<ov::op::v1::Reshape>(
+        src,
+        ov::op::v0::Constant::create(ov::element::i64, {1}, {-1}),
+        false);
+
+    auto src_shape = std::make_shared<ov::op::v3::ShapeOf>(src_flat, ov::element::i64);
+    auto src_len = std::make_shared<ov::op::v1::ReduceProd>(
+        src_shape,
+        ov::op::v0::Constant::create(ov::element::i64, {1}, {0}),
+        false);
+
+    auto start = ov::op::v0::Constant::create(ov::element::i64, {}, {offset_elems});
+    auto stop = std::make_shared<ov::op::v1::Add>(start, src_len);
+    auto step = ov::op::v0::Constant::create(ov::element::i64, {}, {1});
+
+    auto indices = std::make_shared<ov::op::v4::Range>(start, stop, step, ov::element::i64);
+    auto axis = ov::op::v0::Constant::create(ov::element::i64, {}, {0});
+
+    auto updated_flat = std::make_shared<ov::op::v3::ScatterUpdate>(dst_flat, indices, src_flat, axis);
+
+    auto dst_shape = std::make_shared<ov::op::v3::ShapeOf>(dst, ov::element::i64);
+    auto res = std::make_shared<ov::op::v1::Reshape>(updated_flat, dst_shape, false);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/set_rows.cpp

@@ -0,0 +1,96 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <cassert>
+#include <cstdint>
+#include <memory>
+#include <openvino/core/node.hpp>
+#include <openvino/core/node_output.hpp>
+#include <openvino/core/type.hpp>
+#include <openvino/frontend/exception.hpp>
+#include <openvino/op/broadcast.hpp>
+#include <openvino/op/concat.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/gather.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/scatter_elements_update.hpp>
+#include <openvino/op/scatter_update.hpp>
+#include <openvino/op/shape_of.hpp>
+#include <openvino/op/slice.hpp>
+#include <openvino/op/squeeze.hpp>
+#include <openvino/op/transpose.hpp>
+#include <vector>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_set_rows(const NodeContext & context) {
+    num_inputs_check(context, 3, 3);
+
+    auto data = process_view_input_new(context, 0);
+    auto indices = process_view_input_new(context, 1);
+    auto dst = process_view_input_new(context, 2);
+
+    data = std::make_shared<ov::op::v0::Convert>(data, context.get_output_type());
+
+    const auto indices_shape = context.get_input_shape(1);
+    const bool multidim_indices = indices_shape.rank().is_static() &&
+                                  indices_shape.rank().get_length() == 4 &&
+                                  ((indices_shape[1].is_static() && indices_shape[1].get_length() > 1) ||
+                                   (indices_shape[2].is_static() && indices_shape[2].get_length() > 1));
+
+    auto axes = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{}, {2});
+
+    Output<Node> res;
+    if (context.is_stateful()) {
+        int concat_axis = 1;
+        int64_t dim2 = dst.get_partial_shape()[2].get_length();
+        int64_t dim3 = dst.get_partial_shape()[3].get_length();
+        data = std::make_shared<ov::op::v1::Reshape>(
+            data, ov::op::v0::Constant::create(ov::element::i64, {4}, {(int64_t) 1, (int64_t) -1, dim2, dim3}), false);
+        res = std::make_shared<ov::op::v0::Concat>(OutputVector{dst, data}, concat_axis);
+    } else if (multidim_indices) {
+        auto updates_shape = std::make_shared<ov::op::v3::ShapeOf>(data, ov::element::i64);
+
+        auto indices_rank3 = std::make_shared<ov::op::v0::Squeeze>(
+            indices, ov::op::v0::Constant::create(ov::element::i64, {1}, {0}));
+        auto one = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+        auto indices_rank4_shape = std::make_shared<ov::op::v0::Concat>(OutputVector{get_dimensions(updates_shape, {0, 1, 2}), one}, 0);
+        auto indices_rank4 = std::make_shared<ov::op::v1::Reshape>(indices_rank3, indices_rank4_shape, false);
+        auto broadcasted_indices = std::make_shared<ov::op::v3::Broadcast>(indices_rank4, updates_shape);
+
+        res = std::make_shared<ov::op::v3::ScatterElementsUpdate>(dst, broadcasted_indices, data, axes);
+    } else {
+        auto row_size = context.get_input_shape(2)[3].get_length();
+        auto ind_squeezed = std::make_shared<ov::op::v0::Squeeze>(
+            indices, ov::op::v0::Constant::create(ov::element::i64, {3}, {0, 1, 2}));
+        auto data_reshaped = std::make_shared<ov::op::v1::Reshape>(
+            data,
+            ov::op::v0::Constant::create(ov::element::i64, {4},
+                                         {(int64_t) 1, (int64_t) 1, (int64_t) -1, (int64_t) row_size}),
+            false);
+        res = std::make_shared<ov::op::v3::ScatterUpdate>(dst, ind_squeezed, data_reshaped, axes);
+    }
+
+    auto dst_reshape = ov::as_type_ptr<ov::op::v1::Reshape>(dst.get_node_shared_ptr());
+    if (!multidim_indices && dst_reshape) {
+        // Fix the case of multiple sequences, reshape back to original shape [1, n_seq, ctx_per_seq, emb]
+        // ctx_per_seq is not fixed due to llama-bench compatibility
+        auto dst_shape_partial = dst_reshape->get_input_partial_shape(0);
+        std::vector<int64_t> dst_shape = {dst_shape_partial[0].get_length(), dst_shape_partial[1].get_length(),
+                                          dst_shape_partial[2].is_static() ? dst_shape_partial[2].get_length() : -1,
+                                          dst_shape_partial[3].get_length()};
+        res = std::make_shared<ov::op::v1::Reshape>(res, ov::op::v0::Constant::create(ov::element::i64, {4}, dst_shape),
+                                                    false);
+    }
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/softmax.cpp

@@ -0,0 +1,166 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <cmath>
+#include <cstdint>
+#include <cstring>
+#include <memory>
+#include <openvino/op/broadcast.hpp>
+#include <openvino/frontend/exception.hpp>
+#include <openvino/op/add.hpp>
+#include <openvino/op/concat.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/multiply.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/shape_of.hpp>
+#include <openvino/op/slice.hpp>
+#include <openvino/op/softmax.hpp>
+#include <vector>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+static bool is_static_one(const ov::Dimension & dim) {
+    return dim.is_static() && dim.get_length() == 1;
+}
+
+static bool same_static_dim(const ov::Dimension & lhs, const ov::Dimension & rhs) {
+    return lhs.is_static() && rhs.is_static() && lhs.get_length() == rhs.get_length();
+}
+
+static bool is_attention_sinks_input_shape(const ov::PartialShape & candidate, const ov::PartialShape & logits_shape) {
+    if (candidate.rank().is_dynamic() || logits_shape.rank().is_dynamic() || candidate.rank().get_length() != 4 ||
+        logits_shape.rank().get_length() != 4) {
+        return false;
+    }
+
+    return is_static_one(candidate[0]) && is_static_one(candidate[1]) && is_static_one(candidate[2]) &&
+           same_static_dim(candidate[3], logits_shape[1]);
+}
+
+// Reimplementation of GGML_OP_SOFT_MAX semantics for OpenVINO backend:
+// 1) logits = src0 * scale
+// 2) logits += mask (if provided)
+// 3) append attention sinks as hidden logits (if provided)
+// 4) softmax over the last dimension and remove the hidden sink column
+OutputVector translate_soft_max(const NodeContext & context) {
+    num_inputs_check(context, 1, 3);
+
+    float scale = 1.0f;
+    float max_bias = 0.0f;
+    memcpy(&scale, (float *) context.get_output_op_params() + 0, sizeof(float));
+    memcpy(&max_bias, (float *) context.get_output_op_params() + 1, sizeof(float));
+
+    ov::Output<ov::Node> logits = context.get_input(0);
+    const bool second_input_is_sinks =
+        context.get_input_size() == 2 && is_attention_sinks_input_shape(context.get_input_shape(1), context.get_output_shape());
+    const bool has_mask = context.get_input_size() > 1 && !second_input_is_sinks;
+    const bool has_sinks = second_input_is_sinks || context.get_input_size() > 2;
+    const size_t sinks_input_idx = second_input_is_sinks ? 1 : 2;
+
+    // Apply scale first: logits = src0 * scale
+    if (scale != 1.0f) {
+        auto scale_const =
+            std::make_shared<ov::op::v0::Constant>(ov::element::f32, ov::Shape{}, std::vector<float>{scale});
+        logits = std::make_shared<ov::op::v1::Multiply>(logits, scale_const);
+    }
+
+    FRONT_END_CHECK_IMPLEMENTED(!(max_bias > 0.0f && !has_mask),
+                                "OpenVINO softmax ALiBi path requires mask input");
+
+    // Optional mask add: logits += mask
+    // For max_bias > 0 (ALiBi), apply per-head slope to mask before adding.
+    if (has_mask) {
+        ov::Output<ov::Node> mask = context.get_input(1);
+
+        // For stateful
+        std::string mask_name = "KQ_mask_sliced";
+        if (context.get_input_names()[1].find("swa") != std::string::npos) {
+            mask_name = "KQ_mask_swa_sliced";
+        }
+        if (context.has_input(mask_name)) {
+            mask = context.get_input(mask_name);
+        }
+
+        if (mask.get_element_type() != logits.get_element_type()) {
+            mask = std::make_shared<ov::op::v0::Convert>(mask, logits.get_element_type());
+        }
+
+        if (max_bias > 0.0f) {
+            auto out_shape = context.get_output_shape().to_shape();
+            FRONT_END_CHECK_IMPLEMENTED(out_shape.size() == 4, "OpenVINO softmax ALiBi path expects rank-4 tensor");
+
+            const uint32_t n_head = static_cast<uint32_t>(out_shape[1]);
+            FRONT_END_CHECK_IMPLEMENTED(n_head > 0, "OpenVINO softmax ALiBi path expects n_head > 0");
+
+            const uint32_t n_head_log2 = 1u << static_cast<uint32_t>(std::floor(std::log2(static_cast<float>(n_head))));
+            const float m0 = std::pow(2.0f, -(max_bias) / static_cast<float>(n_head_log2));
+            const float m1 = std::pow(2.0f, -(max_bias / 2.0f) / static_cast<float>(n_head_log2));
+
+            std::vector<float> slopes(n_head);
+            for (uint32_t h = 0; h < n_head; ++h) {
+                slopes[h] = h < n_head_log2 ? std::pow(m0, static_cast<float>(h + 1)) :
+                                              std::pow(m1, static_cast<float>(2 * (h - n_head_log2) + 1));
+            }
+
+            ov::Output<ov::Node> slope_node =
+                std::make_shared<ov::op::v0::Constant>(ov::element::f32, ov::Shape{n_head}, slopes);
+            if (slope_node.get_element_type() != mask.get_element_type()) {
+                slope_node = std::make_shared<ov::op::v0::Convert>(slope_node, mask.get_element_type());
+            }
+
+            auto slope_shape = std::make_shared<ov::op::v0::Constant>(
+                ov::element::i64, ov::Shape{4}, std::vector<int64_t>{1, static_cast<int64_t>(n_head), 1, 1});
+            auto slope_4d = std::make_shared<ov::op::v1::Reshape>(slope_node, slope_shape, false);
+            mask = std::make_shared<ov::op::v1::Multiply>(mask, slope_4d);
+        }
+
+        logits = std::make_shared<ov::op::v1::Add>(logits, mask);
+    }
+
+    ov::Output<ov::Node> softmax_input = logits;
+    if (has_sinks) {
+        ov::Output<ov::Node> sinks = context.get_input(sinks_input_idx);
+        if (sinks.get_element_type() != logits.get_element_type()) {
+            sinks = std::make_shared<ov::op::v0::Convert>(sinks, logits.get_element_type());
+        }
+
+        auto sink_shape = ov::op::v0::Constant::create(ov::element::i64, {4}, {1, -1, 1, 1});
+        auto sinks_4d = std::make_shared<ov::op::v1::Reshape>(sinks, sink_shape, false);
+
+        auto logits_shape = std::make_shared<ov::op::v3::ShapeOf>(logits, ov::element::i64);
+        auto zero = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
+        auto one = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+        auto three = ov::op::v0::Constant::create(ov::element::i64, {1}, {3});
+        auto four = ov::op::v0::Constant::create(ov::element::i64, {1}, {4});
+        auto shape_axis = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
+
+        auto sink_prefix_shape = std::make_shared<ov::op::v8::Slice>(logits_shape, zero, three, one, shape_axis);
+        auto sink_last_dim = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+        auto sink_broadcast_shape = std::make_shared<ov::op::v0::Concat>(
+            ov::OutputVector{sink_prefix_shape, sink_last_dim}, 0);
+        auto sink_column = std::make_shared<ov::op::v3::Broadcast>(sinks_4d, sink_broadcast_shape,
+                                                                   ov::op::BroadcastType::BIDIRECTIONAL);
+        softmax_input = std::make_shared<ov::op::v0::Concat>(ov::OutputVector{logits, sink_column}, 3);
+
+        auto softmax_with_sink = std::make_shared<ov::op::v8::Softmax>(softmax_input, -1);
+        auto original_last_dim = std::make_shared<ov::op::v8::Slice>(logits_shape, three, four, one, shape_axis);
+        auto res = std::make_shared<ov::op::v8::Slice>(softmax_with_sink, zero, original_last_dim, one, three);
+
+        return rename_outputs_with_suffix({res}, context.get_name());
+    }
+
+    // Softmax along last dimension (equivalent to ggml softmax over ne[0]).
+    auto res = std::make_shared<ov::op::v8::Softmax>(softmax_input, -1);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/solve_tri.cpp

@@ -0,0 +1,108 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <openvino/op/broadcast.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/divide.hpp>
+#include <openvino/op/gather.hpp>
+#include <openvino/op/loop.hpp>
+#include <openvino/op/matmul.hpp>
+#include <openvino/op/scatter_update.hpp>
+#include <openvino/op/shape_of.hpp>
+#include <openvino/op/subtract.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+// GGML SOLVE_TRI: solve Ax = B for lower-triangular A via forward substitution.
+// Currently only lower, right, non-unitriangular variant is implemented.
+//
+// ggml layout: A [n, n, B1, B2], B [k, n, B1, B2] → X [k, n, B1, B2]
+// OV layout:   A [B2, B1, n, n], B [B2, B1, n, k] → X [B2, B1, n, k]
+//
+// Forward substitution row i:
+//   x[i] = (b[i] - sum_{t<i} A[i,t]*x[t]) / A[i,i]
+//
+// Implemented as an OV Loop op iterating n times with a carried X accumulator.
+// Key insight: A is lower-triangular and X starts as zeros, so the full matmul
+//   A_row_i @ X_partial = sum_{t<i} A[i,t]*x[t] exactly (upper triangle of A
+//   is zero; unfilled rows of X are zero).
+OutputVector translate_solve_tri(const NodeContext & context) {
+    num_inputs_check(context, 2, 2);
+
+    auto A = context.get_input(0);  // [B2, B1, n, n]
+    auto B = context.get_input(1);  // [B2, B1, n, k]
+
+    auto A_shape = context.get_input_shape(0).to_shape();
+    int64_t n = static_cast<int64_t>(A_shape[2]);
+
+    // Initial X: zeros with shape of B
+    auto B_shape_node = std::make_shared<ov::op::v3::ShapeOf>(B, ov::element::i64);
+    auto zero_f32  = ov::op::v0::Constant::create(ov::element::f32, {}, {0.0f});
+    auto X_init    = std::make_shared<ov::op::v3::Broadcast>(zero_f32, B_shape_node);
+
+    // --- Loop body parameters ---
+    // body_iter: iteration counter injected by the Loop op (i64, shape {1})
+    auto body_iter = std::make_shared<ov::op::v0::Parameter>(ov::element::i64, ov::Shape{1});
+    auto body_X    = std::make_shared<ov::op::v0::Parameter>(ov::element::f32, ov::PartialShape::dynamic(4));
+    auto body_A    = std::make_shared<ov::op::v0::Parameter>(ov::element::f32, ov::PartialShape::dynamic(4));
+    auto body_B_p  = std::make_shared<ov::op::v0::Parameter>(ov::element::f32, ov::PartialShape::dynamic(4));
+
+    auto c_axis2        = ov::op::v0::Constant::create(ov::element::i64, {1}, {int64_t(2)});
+    auto c_axis3        = ov::op::v0::Constant::create(ov::element::i64, {1}, {int64_t(3)});
+    auto c_axis2_scalar = ov::op::v0::Constant::create(ov::element::i64, {}, {int64_t(2)});
+
+    // b_i      = B[..., i, :]      [B2, B1, 1, k]
+    auto b_i = std::make_shared<ov::op::v8::Gather>(body_B_p, body_iter, c_axis2);
+
+    // A_row_i  = A[..., i, :]      [B2, B1, 1, n]
+    auto A_row_i = std::make_shared<ov::op::v8::Gather>(body_A, body_iter, c_axis2);
+
+    // sum_i    = A_row_i @ X       [B2, B1, 1, k]
+    // (lower-tri zeros + unfilled-X zeros make this equal to the partial sum)
+    auto sum_i = std::make_shared<ov::op::v0::MatMul>(A_row_i, body_X, false, false);
+
+    // diag_i   = A[..., i, i]      [B2, B1, 1, 1]
+    auto diag_i = std::make_shared<ov::op::v8::Gather>(A_row_i, body_iter, c_axis3);
+
+    // x_i      = (b_i - sum_i) / diag_i    [B2, B1, 1, k]
+    auto x_i = std::make_shared<ov::op::v1::Divide>(
+        std::make_shared<ov::op::v1::Subtract>(b_i, sum_i), diag_i);
+
+    // X_updated: scatter x_i into body_X at row i along axis 2
+    auto X_updated = std::make_shared<ov::op::v3::ScatterUpdate>(body_X, body_iter, x_i, c_axis2_scalar);
+
+    auto body_cond = ov::op::v0::Constant::create(ov::element::boolean, ov::Shape{1}, {true});
+
+    auto body = std::make_shared<ov::Model>(
+        ov::OutputVector{body_cond, X_updated},
+        ov::ParameterVector{body_iter, body_X, body_A, body_B_p});
+
+    // --- Assemble Loop ---
+    auto trip_count = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{1}, std::vector<int64_t>{n});
+    auto exec_cond  = ov::op::v0::Constant::create(ov::element::boolean, ov::Shape{1}, {true});
+
+    auto loop = std::make_shared<ov::op::v5::Loop>(trip_count, exec_cond);
+    loop->set_function(body);
+    // iter_counter_body_param_idx=0 (body_iter), exec_condition_body_result_idx=0 (body_cond)
+    loop->set_special_body_ports(ov::op::v5::Loop::SpecialBodyPorts{0, 0});
+
+    // Carried state: X feeds back from X_updated each iteration
+    loop->set_merged_input(body_X, X_init, X_updated);
+    // Invariant inputs passed through unchanged
+    loop->set_invariant_input(body_A,   A);
+    loop->set_invariant_input(body_B_p, B);
+
+    // Final output: value of X_updated after the last iteration
+    auto X_final = loop->get_iter_value(X_updated, -1);
+
+    return rename_outputs_with_suffix({X_final}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/sqr.cpp

@@ -0,0 +1,35 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <memory>
+#include <openvino/op/multiply.hpp>
+#include <openvino/op/sqrt.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_sqr(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+
+    auto input = process_view_input_new(context, 0);
+    auto res = std::make_shared<ov::op::v1::Multiply>(input, input);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+OutputVector translate_sqrt(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+
+    auto input = process_view_input_new(context, 0);
+    auto res = std::make_shared<ov::op::v0::Sqrt>(input);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/ssm_conv.cpp

@@ -0,0 +1,61 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <openvino/op/constant.hpp>
+#include <openvino/op/group_conv.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/squeeze.hpp>
+#include <openvino/op/transpose.hpp>
+#include <openvino/op/unsqueeze.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_ssm_conv(const NodeContext & context) {
+    num_inputs_check(context, 2, 2);
+
+    auto sx = context.get_input(0);                         // conv state + input: OV shape [1, n_s, d_inner, ncs]
+    auto c = context.get_input(1);                          // conv1d weight:      OV shape [1, 1, d_inner, d_conv]
+
+    auto sx_shape = context.get_input_shape(0).to_shape();  // [1, n_s, d_inner, ncs]
+    auto c_shape = context.get_input_shape(1).to_shape();   // [1, 1, d_inner, d_conv]
+
+    // int64_t n_s  = sx_shape[1];
+    int64_t d_inner = sx_shape[2];
+    // int64_t ncs  = sx_shape[3];  // d_conv - 1 + n_t
+    int64_t d_conv  = c_shape[3];
+    // int64_t n_t  = ncs - d_conv + 1;
+
+    // Reshape sx from [1, n_s, d_inner, ncs] to [n_s, d_inner, ncs] for 1D GroupConvolution
+    auto sx_reshaped =
+        std::make_shared<ov::op::v0::Squeeze>(sx, ov::op::v0::Constant::create(ov::element::i64, {1}, {0}));
+
+    // Reshape c from [1, 1, d_inner, d_conv] to [d_inner, 1, 1, d_conv]
+    // GroupConvolution filter: [groups, out_channels/groups, in_channels/groups, kernel_size]
+    auto c_new_shape = ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{d_inner, 1, 1, d_conv});
+    auto c_reshaped = std::make_shared<ov::op::v1::Reshape>(c, c_new_shape, false);
+
+    // Depthwise 1D convolution: groups=d_inner, stride=1, no padding, no dilation
+    // Input: [n_s, d_inner, ncs], Filter: [d_inner, 1, 1, d_conv]
+    // Output: [n_s, d_inner, n_t]
+    auto conv = std::make_shared<ov::op::v1::GroupConvolution>(
+        sx_reshaped, c_reshaped, ov::Strides{1}, ov::CoordinateDiff{0}, ov::CoordinateDiff{0}, ov::Strides{1});
+
+    // Transpose from [n_s, d_inner, n_t] to [n_s, n_t, d_inner]
+    auto perm = ov::op::v0::Constant::create(ov::element::i64, {3}, std::vector<int64_t>{0, 2, 1});
+    auto transposed = std::make_shared<ov::op::v1::Transpose>(conv, perm);
+
+    // Reshape to output shape [1, n_s, n_t, d_inner]
+    auto res =
+        std::make_shared<ov::op::v0::Unsqueeze>(transposed, ov::op::v0::Constant::create(ov::element::i64, {1}, {0}));
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/sum.cpp

@@ -0,0 +1,30 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <memory>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/range.hpp>
+#include <openvino/op/reduce_sum.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/shape_of.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_sum(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+
+    auto input = process_view_input_new(context, 0);
+    auto axes = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{4}, {0, 1, 2, 3});
+    auto res = std::make_shared<ov::op::v1::ReduceSum>(input, axes, true);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/sum_rows.cpp

@@ -0,0 +1,27 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <memory>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/reduce_sum.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_sum_rows(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+
+    auto input = process_view_input_new(context, 0);
+    auto res = std::make_shared<ov::op::v1::ReduceSum>(
+        input, ov::op::v0::Constant::create(ov::element::i64, ov::Shape{1}, {-1}), true);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/transpose.cpp

@@ -0,0 +1,52 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <openvino/op/transpose.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_transpose(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+
+    // Compute permute order from input/output shape and stride information
+    // so it adapts to different input and output layouts.
+    auto input_stride = context.get_input_stride(0);
+    auto output_stride = context.get_output_stride();
+
+    // Compute permute order by matching output and input stride rankings.
+    // Build <stride, dim_index> pairs.
+    std::vector<std::pair<size_t, int>> output_stride_dims;
+    std::vector<std::pair<size_t, int>> input_stride_dims;
+
+    for (int i = 0; i < 4; ++i) {
+        output_stride_dims.push_back({output_stride[i], i});
+        input_stride_dims.push_back({input_stride[i], i});
+    }
+
+    // Sort by stride in descending order.
+    std::sort(output_stride_dims.rbegin(), output_stride_dims.rend());
+    std::sort(input_stride_dims.rbegin(), input_stride_dims.rend());
+
+    // Build permute order.
+    std::vector<int64_t> permute_order(4);
+    for (int i = 0; i < 4; ++i) {
+        int output_dim = output_stride_dims[i].second;
+        int input_dim = input_stride_dims[i].second;
+        permute_order[output_dim] = input_dim;
+    }
+
+    auto input = process_view_input_new(context, 0);
+
+    auto res = std::make_shared<ov::op::v1::Transpose>(
+        input, ov::op::v0::Constant::create(ov::element::i64, {4}, permute_order));
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/tri.cpp

@@ -0,0 +1,82 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <openvino/op/constant.hpp>
+#include <openvino/op/greater.hpp>
+#include <openvino/op/greater_eq.hpp>
+#include <openvino/op/less.hpp>
+#include <openvino/op/less_eq.hpp>
+#include <openvino/op/range.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/select.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+// GGML TRI zeroes out elements outside a triangular region of a square matrix.
+// The type param (stored in op_params[0]) maps to ggml_tri_type:
+//   0 = UPPER_DIAG : keep where col >= row
+//   1 = UPPER      : keep where col >  row
+//   2 = LOWER_DIAG : keep where col <= row
+//   3 = LOWER      : keep where col <  row
+//
+// In OV layout (ggml [ne0, ne1, ne2, ne3] → OV [ne3, ne2, ne1, ne0]):
+//   ggml dim 0 (ne0, cols) → OV axis 3
+//   ggml dim 1 (ne1, rows) → OV axis 2
+// The matrix is square so ne0 == ne1.
+OutputVector translate_tri(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+
+    auto x = context.get_input(0);  // OV shape: [ne3, ne2, ne1, ne0]
+
+    int32_t tri_type = context.get_output_op_params()[0];
+
+    auto shape = context.get_input_shape(0).to_shape();
+    int64_t n = static_cast<int64_t>(shape[3]);  // ne0 == ne1
+
+    // Build index range [0, 1, ..., n-1]
+    auto start = ov::op::v0::Constant::create(ov::element::i64, {}, {int64_t(0)});
+    auto stop  = ov::op::v0::Constant::create(ov::element::i64, {}, {n});
+    auto step  = ov::op::v0::Constant::create(ov::element::i64, {}, {int64_t(1)});
+    auto range = std::make_shared<ov::op::v4::Range>(start, stop, step, ov::element::i64);
+
+    // col_idx shape [1, 1, 1, n] — broadcasts over batch and row dims
+    auto col_shape = ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{1, 1, 1, n});
+    auto col_idx   = std::make_shared<ov::op::v1::Reshape>(range, col_shape, false);
+
+    // row_idx shape [1, 1, n, 1] — broadcasts over batch and col dims
+    auto row_shape = ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{1, 1, n, 1});
+    auto row_idx   = std::make_shared<ov::op::v1::Reshape>(range, row_shape, false);
+
+    // Build boolean mask: true where element should be kept
+    std::shared_ptr<ov::Node> mask;
+    switch (tri_type) {
+        case 0:  // UPPER_DIAG: col >= row
+            mask = std::make_shared<ov::op::v1::GreaterEqual>(col_idx, row_idx);
+            break;
+        case 1:  // UPPER: col > row
+            mask = std::make_shared<ov::op::v1::Greater>(col_idx, row_idx);
+            break;
+        case 2:  // LOWER_DIAG: col <= row
+            mask = std::make_shared<ov::op::v1::LessEqual>(col_idx, row_idx);
+            break;
+        case 3:  // LOWER: col < row
+            mask = std::make_shared<ov::op::v1::Less>(col_idx, row_idx);
+            break;
+        default:
+            throw std::runtime_error("translate_tri: invalid tri_type " + std::to_string(tri_type));
+    }
+
+    auto zero = ov::op::v0::Constant::create(ov::element::f32, {}, {0.0f});
+    auto res  = std::make_shared<ov::op::v1::Select>(mask, x, zero);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/unary.cpp

@@ -0,0 +1,138 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+#include "ggml-openvino/ggml-openvino-extra.h"
+
+#include <memory>
+#include <openvino/op/abs.hpp>
+#include <openvino/op/add.hpp>
+#include <openvino/op/clamp.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/elu.hpp>
+#include <openvino/op/exp.hpp>
+#include <openvino/op/gelu.hpp>
+#include <openvino/op/greater.hpp>
+#include <openvino/op/hard_sigmoid.hpp>
+#include <openvino/op/log.hpp>
+#include <openvino/op/multiply.hpp>
+#include <openvino/op/negative.hpp>
+#include <openvino/op/relu.hpp>
+#include <openvino/op/round.hpp>
+#include <openvino/op/sigmoid.hpp>
+#include <openvino/op/softplus.hpp>
+#include <openvino/op/subtract.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_unary_gelu(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+    auto input = process_view_input_new(context, 0);
+    auto res = std::make_shared<ov::op::v7::Gelu>(input, ov::op::GeluApproximationMode::TANH);
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+OutputVector translate_unary_gelu_erf(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+    auto input = process_view_input_new(context, 0);
+    auto res = std::make_shared<ov::op::v7::Gelu>(input, ov::op::GeluApproximationMode::ERF);
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+OutputVector translate_unary_gelu_quick(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+    auto input = process_view_input_new(context, 0);
+    auto scale = ov::op::v0::Constant::create(input.get_element_type(), ov::Shape{}, {1.702f});
+    auto mul = std::make_shared<ov::op::v1::Multiply>(input, scale);
+    auto sig = std::make_shared<ov::op::v0::Sigmoid>(mul);
+    auto res = std::make_shared<ov::op::v1::Multiply>(input, sig);
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+OutputVector translate_unary_elu(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+    auto input = process_view_input_new(context, 0);
+    auto res = std::make_shared<ov::op::v0::Elu>(input, 1.0);
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+OutputVector translate_unary_hardsigmoid(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+    // compute in f32 like the ggml reference: 1/6 is not exact in f16/bf16 (NPU cannot take the f32 path)
+    auto input = process_view_input_new(context, 0);
+    const auto type = ggml_openvino_is_npu() ? input.get_element_type() : ov::element::f32;
+    ov::Output<ov::Node> x = input;
+    if (type != input.get_element_type()) {
+        x = std::make_shared<ov::op::v0::Convert>(input, type);
+    }
+    auto alpha = ov::op::v0::Constant::create(type, ov::Shape{}, {1.0f / 6.0f});
+    auto beta = ov::op::v0::Constant::create(type, ov::Shape{}, {0.5f});
+    ov::Output<ov::Node> res = std::make_shared<ov::op::v0::HardSigmoid>(x, alpha, beta);
+    if (type != input.get_element_type()) {
+        res = std::make_shared<ov::op::v0::Convert>(res, input.get_element_type());
+    }
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+OutputVector translate_unary_step(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+    auto input = process_view_input_new(context, 0);
+    auto zero = ov::op::v0::Constant::create(input.get_element_type(), ov::Shape{}, {0.0f});
+    auto cond = std::make_shared<ov::op::v1::Greater>(input, zero);
+    auto res = std::make_shared<ov::op::v0::Convert>(cond, input.get_element_type());
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+OutputVector translate_unary_round(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+    auto input = process_view_input_new(context, 0);
+    auto res = std::make_shared<ov::op::v5::Round>(input, ov::op::v5::Round::RoundMode::HALF_AWAY_FROM_ZERO);
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+OutputVector translate_unary_expm1(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+    // compute in f32 like the ggml reference: exp(x) - 1 in f16 loses the small-x digits (NPU cannot take the f32 path)
+    auto input = process_view_input_new(context, 0);
+    const auto type = ggml_openvino_is_npu() ? input.get_element_type() : ov::element::f32;
+    ov::Output<ov::Node> x = input;
+    if (type != input.get_element_type()) {
+        x = std::make_shared<ov::op::v0::Convert>(input, type);
+    }
+    auto exp = std::make_shared<ov::op::v0::Exp>(x);
+    auto one = ov::op::v0::Constant::create(type, ov::Shape{}, {1.0f});
+    ov::Output<ov::Node> res = std::make_shared<ov::op::v1::Subtract>(exp, one);
+    if (type != input.get_element_type()) {
+        res = std::make_shared<ov::op::v0::Convert>(res, input.get_element_type());
+    }
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+OutputVector translate_unary_softplus(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+
+    if (ggml_openvino_getenv_int("GGML_OPENVINO_NATIVE_SOFTPLUS") != 0) {
+        return translate_1to1_match_1_input<ov::op::v4::SoftPlus>(context);
+    }
+
+    auto input = process_view_input_new(context, 0);
+    const auto element_type = input.get_element_type();
+    auto one = ov::op::v0::Constant::create(element_type, ov::Shape{}, {1.0f});
+
+    auto positive = std::make_shared<ov::op::v0::Relu>(input);
+    auto abs = std::make_shared<ov::op::v0::Abs>(input);
+    auto neg_abs = std::make_shared<ov::op::v0::Negative>(abs);
+    auto exp_neg_abs = std::make_shared<ov::op::v0::Exp>(neg_abs);
+    auto log_term = std::make_shared<ov::op::v0::Log>(std::make_shared<ov::op::v1::Add>(one, exp_neg_abs));
+    auto res = std::make_shared<ov::op::v1::Add>(positive, log_term);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/upscale.cpp

@@ -0,0 +1,181 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+#include "ggml.h"
+
+#include <cmath>
+#include <cstddef>
+#include <memory>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/gather.hpp>
+#include <openvino/op/interpolate.hpp>
+#include <openvino/op/matmul.hpp>
+#include <vector>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_upscale(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+
+    using Interpolate = ov::op::v4::Interpolate;
+    auto input = process_view_input_new(context, 0);
+
+    const auto input_shape = context.get_input_shape(0).to_shape();
+    const auto output_shape = context.get_output_shape().to_shape();
+
+    if (input_shape == output_shape) {
+        return rename_outputs_with_suffix({input}, context.get_name());
+    }
+
+    ov::Output<ov::Node> res = input;
+
+    // Resample batch / channel dimensions (ne[3] and ne[2], corresponding to OV axes 0 and 1)
+    // using nearest-neighbor index mapping: i0_d = floor(i_d * in_d / out_d).
+    for (size_t axis = 0; axis < 2; ++axis) {
+        const size_t in_dim = input_shape[axis];
+        const size_t out_dim = output_shape[axis];
+        if (in_dim != out_dim) {
+            std::vector<int64_t> indices(out_dim);
+            for (size_t i = 0; i < out_dim; ++i) {
+                indices[i] = static_cast<int64_t>((i * in_dim) / out_dim);
+            }
+            auto indices_node = ov::op::v0::Constant::create(ov::element::i64, {indices.size()}, indices);
+            auto axis_node = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{}, {axis});
+            res = std::make_shared<ov::op::v8::Gather>(res, indices_node, axis_node);
+        }
+    }
+
+    // Spatial interpolation for ne[1] and ne[0] (OV axes 2 and 3)
+    if (input_shape[2] != output_shape[2] || input_shape[3] != output_shape[3]) {
+        const int32_t * op_params = context.get_output_op_params();
+        const int32_t mode_flags = op_params != nullptr ? op_params[0] : 0;
+
+        const int op_case = context.get_op_case();
+        const bool align_corners = (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) != 0;
+        const bool antialias = (mode_flags & GGML_SCALE_FLAG_ANTIALIAS) != 0;
+
+        if (op_case == 2 && antialias) { // GGML_SCALE_MODE_BILINEAR with antialias
+            const size_t H_in = input_shape[2];
+            const size_t W_in = input_shape[3];
+            const size_t H_out = output_shape[2];
+            const size_t W_out = output_shape[3];
+
+            const float pixel_offset = 0.5f;
+
+            // Height projection: [H_out, H_in]
+            std::vector<float> Wy_data(H_out * H_in, 0.0f);
+            const float sf1 = static_cast<float>(H_out) / H_in;
+            const float support1 = std::max(1.0f, 1.0f / sf1);
+            const float invscale1 = 1.0f / support1;
+
+            for (size_t i1 = 0; i1 < H_out; ++i1) {
+                const float y = (static_cast<float>(i1) + pixel_offset) / sf1;
+                const auto y_start = static_cast<int64_t>(y - support1 + pixel_offset);
+                const size_t y_min = y_start > 0 ? static_cast<size_t>(y_start) : 0;
+                const auto y_end = static_cast<int64_t>(y + support1 + pixel_offset);
+                const size_t y_max = y_end > 0 ? std::min<size_t>(static_cast<size_t>(y_end), H_in) : 0;
+
+                float total_weight = 0.0f;
+                for (size_t sy = y_min; sy < y_max; ++sy) {
+                    float diff = std::abs((static_cast<float>(sy) - y + pixel_offset) * invscale1);
+                    float weight = std::max(1.0f - diff, 0.0f);
+                    Wy_data[i1 * H_in + sy] = weight;
+                    total_weight += weight;
+                }
+                if (total_weight > 0.0f) {
+                    for (size_t sy = y_min; sy < y_max; ++sy) {
+                        Wy_data[i1 * H_in + sy] /= total_weight;
+                    }
+                }
+            }
+
+            // Width projection: [W_in, W_out]
+            std::vector<float> Wx_data(W_in * W_out, 0.0f);
+            const float sf0 = static_cast<float>(W_out) / W_in;
+            const float support0 = std::max(1.0f, 1.0f / sf0);
+            const float invscale0 = 1.0f / support0;
+
+            for (size_t i0 = 0; i0 < W_out; ++i0) {
+                const float x = (static_cast<float>(i0) + pixel_offset) / sf0;
+                const auto x_start = static_cast<int64_t>(x - support0 + pixel_offset);
+                const size_t x_min = x_start > 0 ? static_cast<size_t>(x_start) : 0;
+                const auto x_end = static_cast<int64_t>(x + support0 + pixel_offset);
+                const size_t x_max = x_end > 0 ? std::min<size_t>(static_cast<size_t>(x_end), W_in) : 0;
+
+                float total_weight = 0.0f;
+                for (size_t sx = x_min; sx < x_max; ++sx) {
+                    float diff = std::abs((static_cast<float>(sx) - x + pixel_offset) * invscale0);
+                    float weight = std::max(1.0f - diff, 0.0f);
+                    Wx_data[sx * W_out + i0] = weight;
+                    total_weight += weight;
+                }
+                if (total_weight > 0.0f) {
+                    for (size_t sx = x_min; sx < x_max; ++sx) {
+                        Wx_data[sx * W_out + i0] /= total_weight;
+                    }
+                }
+            }
+
+            auto Wy_node = ov::op::v0::Constant::create(ov::element::f32, {H_out, H_in}, Wy_data);
+            auto Wx_node = ov::op::v0::Constant::create(ov::element::f32, {W_in, W_out}, Wx_data);
+
+            auto res_y = std::make_shared<ov::op::v0::MatMul>(Wy_node, res);
+            res = std::make_shared<ov::op::v0::MatMul>(res_y, Wx_node);
+        } else {
+            Interpolate::InterpolateAttrs attrs;
+            attrs.shape_calculation_mode = Interpolate::ShapeCalcMode::SIZES;
+            attrs.antialias = antialias;
+            attrs.pads_begin = {0, 0, 0, 0};
+            attrs.pads_end = {0, 0, 0, 0};
+
+            switch (op_case) {
+            case 1:  // GGML_SCALE_MODE_NEAREST
+                attrs.mode = Interpolate::InterpolateMode::NEAREST;
+                attrs.nearest_mode = Interpolate::NearestMode::FLOOR;
+                attrs.coordinate_transformation_mode = Interpolate::CoordinateTransformMode::ASYMMETRIC;
+                break;
+            case 2:  // GGML_SCALE_MODE_BILINEAR
+                attrs.mode = Interpolate::InterpolateMode::LINEAR;
+                attrs.coordinate_transformation_mode = align_corners
+                    ? Interpolate::CoordinateTransformMode::ALIGN_CORNERS
+                    : Interpolate::CoordinateTransformMode::HALF_PIXEL;
+                break;
+            case 3:  // GGML_SCALE_MODE_BICUBIC
+                attrs.mode = Interpolate::InterpolateMode::CUBIC;
+                attrs.cube_coeff = -0.75;
+                attrs.coordinate_transformation_mode = align_corners
+                    ? Interpolate::CoordinateTransformMode::ALIGN_CORNERS
+                    : Interpolate::CoordinateTransformMode::HALF_PIXEL;
+                break;
+            default:
+                FRONT_END_OP_CONVERSION_CHECK(false, "Unsupported upscale op_case: ", op_case);
+            }
+
+            std::vector<int64_t> target_shape_vec = {
+                static_cast<int64_t>(output_shape[2]),
+                static_cast<int64_t>(output_shape[3]),
+            };
+            std::vector<float> scales_vec = {
+                static_cast<float>(output_shape[2]) / static_cast<float>(input_shape[2]),
+                static_cast<float>(output_shape[3]) / static_cast<float>(input_shape[3]),
+            };
+
+            auto target_shape_node = ov::op::v0::Constant::create(ov::element::i64, {2}, target_shape_vec);
+            auto scales_node = ov::op::v0::Constant::create(ov::element::f32, {2}, scales_vec);
+            auto axes_node = ov::op::v0::Constant::create(ov::element::i64, {2}, {2, 3});
+
+            res = std::make_shared<Interpolate>(
+                res, target_shape_node, scales_node, axes_node, attrs);
+        }
+    }
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op/view.cpp

@@ -0,0 +1,259 @@
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <openvino/op/concat.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/gather.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/shape_of.hpp>
+#include <openvino/op/slice.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_view(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+
+    if (context.get_op_case() == 1) {
+        // Static-mode identity pass-through for VIEWs over a GATED_DELTA_NET combined output or
+        // the conv_input CONCAT; the consuming op (CPY/RMS_NORM) does its own runtime-correct
+        // slicing on the full tensor (see ggml-decoder.cpp compute_op_case, GGML_OP_VIEW).
+        return {context.get_input(0)};
+    }
+
+    if (!context.is_static()) {
+        // On the stateless/non-static path VIEW is normally a no-op (consumers re-slice).
+        // EXCEPTION: the MoE expert aggregation slices each expert plane out of
+        // ffn_moe_weighted [n_embd, n_expert_used, n_tokens] with ggml_view_2d and then
+        // sums the planes with a chain of ADDs (llama-graph.cpp). Those ADDs read this
+        // VIEW node directly from the tensor map and do NOT re-slice, so a no-op here
+        // makes every plane the full tensor and the expert sum collapses. Materialize the
+        // single-expert slice here. Gated by name (ffn_moe_weighted...view) so it can't
+        // affect any other view.
+        const std::string & vname = context.get_name();
+        if (vname.find("ffn_moe_weighted") != std::string::npos) {
+            auto src_ps = context.get_input_shape(0);
+            auto dst_ps = context.get_output_shape();
+            if (src_ps.rank().is_static() && dst_ps.rank().is_static() && src_ps.rank() == dst_ps.rank() &&
+                src_ps.is_static() && dst_ps.is_static()) {
+                auto sst = context.get_input_stride(0);
+                auto dst = context.get_output_stride();
+                size_t voff = context.get_output_op_offset();
+                auto ss = src_ps.to_shape();
+                auto dd = dst_ps.to_shape();
+                const size_t nd = ss.size();
+                // Stateful models drop the leading size-1 batch dim, so the real OV tensor can
+                // be one rank lower than the ggml shape metadata above. Axis indices derived
+                // from that metadata must be shifted down by the difference before they are
+                // used as OV axes. Bail out if the dims we would drop are not all 1.
+                const auto in_rank = context.get_input(0).get_partial_shape().rank();
+                const int axis_shift =
+                    in_rank.is_static() ? (int) nd - (int) in_rank.get_length() : 0;
+                bool shift_ok = axis_shift >= 0 && (size_t) axis_shift < nd;
+                for (int a = 0; a < axis_shift && shift_ok; ++a) {
+                    shift_ok = (ss[a] == 1 && dd[a] == 1);
+                }
+                if (shift_ok && sst.size() == nd && dst.size() == nd) {
+                    // Map each dst axis of size>1 to a src axis with equal (size,stride);
+                    // the unmatched src axis of size>1 is the indexed expert axis.
+                    // dst_to_src[d] records which src axis each dst axis came from, so we can
+                    // later pull the dynamic (token) dim from the right source axis at runtime.
+                    std::vector<bool> used(nd, false);
+                    std::vector<int> dst_to_src(nd, -1);
+                    bool ok = true;
+                    for (size_t d = 0; d < nd; ++d) {
+                        if (dd[d] == 1) {
+                            continue;
+                        }
+                        int found = -1;
+                        for (size_t s = 0; s < nd; ++s) {
+                            if (!used[s] && ss[s] == dd[d] && sst[s] == dst[d]) {
+                                found = (int) s;
+                                break;
+                            }
+                        }
+                        if (found < 0) {
+                            ok = false;
+                            break;
+                        }
+                        used[found] = true;
+                        dst_to_src[d] = found;
+                    }
+                    int dropped = -1;
+                    if (ok) {
+                        for (size_t s = 0; s < nd; ++s) {
+                            if (!used[s] && ss[s] > 1) {
+                                if (dropped >= 0) {
+                                    ok = false;
+                                    break;
+                                }
+                                dropped = (int) s;
+                            }
+                        }
+                    }
+                    if (ok && dropped >= 0) {
+                        const size_t dstr = sst[dropped];
+                        const int64_t dsz = (int64_t) ss[dropped];
+                        if (dstr > 0 && voff % dstr == 0) {
+                            const int64_t sel = (int64_t) (voff / dstr);
+                            if (sel >= 0 && sel < dsz) {
+                                ov::Output<ov::Node> sl = std::make_shared<ov::op::v8::Slice>(
+                                    context.get_input(0),
+                                    ov::op::v0::Constant::create(ov::element::i64, {1}, {sel}),
+                                    ov::op::v0::Constant::create(ov::element::i64, {1}, {sel + 1}),
+                                    ov::op::v0::Constant::create(ov::element::i64, {1}, {1}),
+                                    ov::op::v0::Constant::create(ov::element::i64, {1},
+                                                                 {dropped - axis_shift}));
+                                // Build the reshape target from the (concrete) dst shape, but
+                                // keep the dynamic token axis dynamic instead of freezing it
+                                // to the captured n_tokens. Without this the constant dst
+                                // shape bakes in the prefill token count and the static value
+                                // flows downstream, turning every later decoder layer static
+                                // (the GPU in-place-concat KV-cache bug). The token axis is
+                                // PERMUTED between the sliced input and the dst (e.g. input
+                                // [1,tok,expert,emb] -> dst [1,1,tok,emb]), so special_zero
+                                // (which copies the same-position dim) is not enough: pull the
+                                // dynamic dim from the correct SOURCE axis via ShapeOf+Gather
+                                // and place it at the dst token position.
+                                const int32_t dyn = context.get_op_dynamic_dim();  // output ggml axis, -1 if none
+                                // still in ggml metadata axis space; get_shape() reverses ggml order
+                                int dst_ov_axis = (dyn != -1) ? ((int) nd - 1 - (int) dyn) : -1;
+                                int src_ov_axis = (dst_ov_axis >= axis_shift && dst_ov_axis < (int) nd)
+                                                      ? dst_to_src[dst_ov_axis]
+                                                      : -1;
+                                if (dst_ov_axis >= 0 && src_ov_axis >= axis_shift) {
+                                    // target = concat of per-axis scalars; the token axis is a
+                                    // runtime Gather of the slice's shape, the rest are constants.
+                                    auto sl_shape = std::make_shared<ov::op::v3::ShapeOf>(sl, ov::element::i64);
+                                    auto tok_dim = std::make_shared<ov::op::v8::Gather>(
+                                        sl_shape,
+                                        ov::op::v0::Constant::create(ov::element::i64, {1},
+                                                                     {src_ov_axis - axis_shift}),
+                                        ov::op::v0::Constant::create(ov::element::i64, {}, {0}));
+                                    ov::OutputVector parts;
+                                    for (int a = axis_shift; a < (int) nd; ++a) {
+                                        if (a == dst_ov_axis) {
+                                            parts.push_back(tok_dim);
+                                        } else {
+                                            parts.push_back(ov::op::v0::Constant::create(
+                                                ov::element::i64, {1}, {(int64_t) dd[a]}));
+                                        }
+                                    }
+                                    auto dc = std::make_shared<ov::op::v0::Concat>(parts, 0);
+                                    auto rs = std::make_shared<ov::op::v1::Reshape>(sl, dc, false);
+                                    return rename_outputs_with_suffix({rs}, context.get_name());
+                                }
+                                std::vector<int64_t> dd_ov(dd.begin() + axis_shift, dd.end());
+                                auto dc = ov::op::v0::Constant::create(ov::element::i64, {dd_ov.size()}, dd_ov);
+                                auto rs = std::make_shared<ov::op::v1::Reshape>(sl, dc, false);
+                                return rename_outputs_with_suffix({rs}, context.get_name());
+                            }
+                        }
+                    }
+                }
+            }
+        }
+        return {context.get_input(0)};
+    }
+
+    auto input = context.get_input(0);
+    auto src_shape = context.get_input_shape(0);
+    auto dst_shape = context.get_output_shape();
+
+    if (src_shape.rank().is_dynamic() || dst_shape.rank().is_dynamic()) {
+        return {input};
+    }
+
+    int64_t src_elems = 1;
+    int64_t dst_elems = 1;
+    for (int64_t i = 0; i < src_shape.rank().get_length(); ++i) {
+        if (src_shape[i].is_dynamic()) {
+            return {input};
+        }
+        src_elems *= src_shape[i].get_length();
+    }
+    for (int64_t i = 0; i < dst_shape.rank().get_length(); ++i) {
+        if (dst_shape[i].is_dynamic()) {
+            return {input};
+        }
+        dst_elems *= dst_shape[i].get_length();
+    }
+
+    if (dst_elems >= src_elems) {
+        return {input};
+    }
+
+    auto src_stride = context.get_input_stride(0);
+    auto dst_stride = context.get_output_stride();
+    size_t view_offset = context.get_output_op_offset();
+
+    bool same_stride = (src_stride.size() == dst_stride.size());
+    if (same_stride) {
+        for (size_t i = 0; i < src_stride.size(); ++i) {
+            if (src_stride[i] != dst_stride[i]) {
+                same_stride = false;
+                break;
+            }
+        }
+    }
+
+    if (!same_stride) {
+        return {input};
+    }
+
+    auto src_ov_shape = src_shape.to_shape();
+    auto dst_ov_shape = dst_shape.to_shape();
+    size_t ndims = src_ov_shape.size();
+    if (dst_ov_shape.size() != ndims) {
+        return {input};
+    }
+
+    std::vector<int> diff_dims;
+    for (size_t i = 0; i < ndims; ++i) {
+        if (src_ov_shape[i] != dst_ov_shape[i]) {
+            diff_dims.push_back(static_cast<int>(i));
+        }
+    }
+
+    if (diff_dims.size() != 1) {
+        return {input};
+    }
+
+    int slice_dim = diff_dims[0];
+    int64_t dim_size = static_cast<int64_t>(src_ov_shape[slice_dim]);
+
+    size_t ov_stride_for_dim = 1;
+    for (size_t i = slice_dim + 1; i < ndims; ++i) {
+        ov_stride_for_dim *= src_ov_shape[i];
+    }
+    size_t elem_size = src_stride.back();
+    if (elem_size == 0) {
+        elem_size = 1;
+    }
+
+    int64_t begin_val = 0;
+    if (ov_stride_for_dim > 0 && elem_size > 0) {
+        begin_val = static_cast<int64_t>((view_offset / elem_size) / ov_stride_for_dim);
+    }
+    int64_t end_val = begin_val + static_cast<int64_t>(dst_ov_shape[slice_dim]);
+
+    if (begin_val < 0 || end_val > dim_size) {
+        return {input};
+    }
+
+    auto sliced =
+        std::make_shared<ov::op::v8::Slice>(input, ov::op::v0::Constant::create(ov::element::i64, {1}, {begin_val}),
+                                            ov::op::v0::Constant::create(ov::element::i64, {1}, {end_val}),
+                                            ov::op::v0::Constant::create(ov::element::i64, {1}, {1}),
+                                            ov::op::v0::Constant::create(ov::element::i64, {1}, {slice_dim}));
+
+    sliced->set_friendly_name(context.get_output_name());
+    return {sliced->output(0)};
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op_table.cpp

@@ -0,0 +1,122 @@
+#include "op_table.h"
+
+#include "utils.h"
+
+#include <openvino/op/abs.hpp>
+#include <openvino/op/add.hpp>
+#include <openvino/op/ceiling.hpp>
+#include <openvino/op/cos.hpp>
+#include <openvino/op/divide.hpp>
+#include <openvino/op/exp.hpp>
+#include <openvino/op/floor.hpp>
+#include <openvino/op/gather.hpp>
+#include <openvino/op/gelu.hpp>
+#include <openvino/op/hswish.hpp>
+#include <openvino/op/log.hpp>
+#include <openvino/op/matmul.hpp>
+#include <openvino/op/multiply.hpp>
+#include <openvino/op/negative.hpp>
+#include <openvino/op/relu.hpp>
+#include <openvino/op/sigmoid.hpp>
+#include <openvino/op/sign.hpp>
+#include <openvino/op/sin.hpp>
+#include <openvino/op/softplus.hpp>
+#include <openvino/op/subtract.hpp>
+#include <openvino/op/swish.hpp>
+#include <openvino/op/tanh.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+
+std::unordered_map<std::string, CreatorFunction> get_supported_ops() {
+    using namespace ov::op;
+    return {
+        {"GGML_OP_ADD",             op::translate_add                              },
+        {"GGML_OP_ADD1",            op::translate_1to1_match_2_inputs<v1::Add>     },
+        {"GGML_OP_ADD_ID",          op::translate_add_id                           },
+        {"GGML_OP_CONCAT",          op::translate_concat                           },
+        {"GGML_OP_CONT",            op::translate_cont                             },
+        {"GGML_OP_DUP",             op::translate_cont                             },
+        {"GGML_OP_DIV",             op::translate_div                              },
+        {"GGML_OP_FILL",            op::translate_fill                             },
+        {"GGML_OP_GET_ROWS",        op::translate_get_rows                         },
+        {"GGML_OP_IM2COL",          op::translate_im2col                           },
+        {"GGML_OP_IM2COL_3D",       op::translate_im2col_3d                        },
+        {"GGML_OP_MUL",             op::translate_1to1_match_2_inputs<v1::Multiply>},
+        {"GGML_OP_MUL_MAT",         op::translate_mulmat                           },
+        {"GGML_OP_MUL_MAT_ID",      op::translate_mul_mat_id                       },
+        {"GGML_OP_PERMUTE",         op::translate_permute                          },
+        {"GGML_OP_RESHAPE",         op::translate_reshape                          },
+        {"GGML_OP_RMS_NORM",        op::translate_rms_norm                         },
+        {"GGML_OP_NORM",            op::translate_norm                             },
+        {"GGML_OP_L2_NORM",         op::translate_l2_norm                          },
+        {"GGML_OP_SUM_ROWS",        op::translate_sum_rows                         },
+        {"GGML_OP_ROPE",            op::translate_rope                             },
+        {"GGML_OP_SCALE",           op::translate_scale                            },
+        {"GGML_OP_SQR",             op::translate_sqr                              },
+        {"GGML_OP_SQRT",            op::translate_sqrt                             },
+        {"GGML_OP_SOFT_MAX",        op::translate_soft_max                         },
+        {"GGML_OP_ARGSORT",         op::translate_argsort                          },
+        {"GGML_OP_SUB",             op::translate_1to1_match_2_inputs<v1::Subtract>},
+        {"GGML_OP_TRANSPOSE",       op::translate_transpose                        },
+        {"GGML_OP_SIN",             op::translate_1to1_match_1_input<v0::Sin>      },
+        {"GGML_OP_COS",             op::translate_1to1_match_1_input<v0::Cos>      },
+        {"GGML_OP_LOG",             op::translate_1to1_match_1_input<v0::Log>      },
+        {"GGML_OP_MEAN",            op::translate_mean                             },
+        {"GGML_OP_SUM",             op::translate_sum                              },
+        {"GGML_UNARY_OP_GELU",      op::translate_unary_gelu                       },
+        {"GGML_UNARY_OP_GELU_ERF",  op::translate_unary_gelu_erf                   },
+        {"GGML_UNARY_OP_GELU_QUICK", op::translate_unary_gelu_quick                },
+        {"GGML_UNARY_OP_ELU",       op::translate_unary_elu                        },
+        {"GGML_UNARY_OP_HARDSWISH", op::translate_1to1_match_1_input<v4::HSwish>   },
+        {"GGML_UNARY_OP_HARDSIGMOID", op::translate_unary_hardsigmoid              },
+        {"GGML_UNARY_OP_STEP",      op::translate_unary_step                       },
+        {"GGML_UNARY_OP_ABS",       op::translate_1to1_match_1_input<v0::Abs>      },
+        {"GGML_UNARY_OP_SGN",       op::translate_1to1_match_1_input<v0::Sign>     },
+        {"GGML_UNARY_OP_FLOOR",     op::translate_1to1_match_1_input<v0::Floor>    },
+        {"GGML_UNARY_OP_CEIL",      op::translate_1to1_match_1_input<v0::Ceiling>  },
+        {"GGML_UNARY_OP_ROUND",     op::translate_unary_round                      },
+        {"GGML_UNARY_OP_EXPM1",     op::translate_unary_expm1                      },
+        {"GGML_UNARY_OP_SIGMOID",   op::translate_1to1_match_1_input<v0::Sigmoid>  },
+        {"GGML_UNARY_OP_SILU",      op::translate_1to1_match_1_input<v4::Swish>    },
+        {"GGML_UNARY_OP_SOFTPLUS",  op::translate_unary_softplus                   },
+        {"GGML_UNARY_OP_TANH",      op::translate_1to1_match_1_input<v0::Tanh>     },
+        {"GGML_UNARY_OP_EXP",       op::translate_1to1_match_1_input<v0::Exp>      },
+        {"GGML_UNARY_OP_NEG",       op::translate_1to1_match_1_input<v0::Negative> },
+        {"GGML_UNARY_OP_RELU",      op::translate_1to1_match_1_input<v0::Relu>     },
+        {"GGML_OP_VIEW",            op::translate_view                             },
+        {"GGML_GLU_OP_SWIGLU",      op::translate_glu_swiglu                       },
+        {"GGML_GLU_OP_SWIGLU_OAI",  op::translate_glu_swiglu_oai                   },
+        {"GGML_GLU_OP_SWIGLU_CLAMP", op::translate_glu_swiglu_clamp                },
+        {"GGML_GLU_OP_GEGLU",       op::translate_glu_geglu                        },
+        {"GGML_GLU_OP_GEGLU_QUICK", op::translate_glu_geglu_quick                  },
+        {"GGML_OP_SET_ROWS",        op::translate_set_rows                         },
+        {"GGML_OP_CPY",             op::translate_cpy                              },
+        {"GGML_OP_FLASH_ATTN_EXT",  op::translate_flash_attn_ext                   },
+        {"GGML_OP_CLAMP",           op::translate_clamp                            },
+        {"GGML_OP_PAD",             op::translate_pad                              },
+        {"GGML_OP_SSM_CONV",        op::translate_ssm_conv                         },
+        {"GGML_OP_GATED_DELTA_NET", op::translate_gated_delta_net                  },
+        {"GGML_OP_REPEAT",          op::translate_repeat                           },
+        {"GGML_OP_CUMSUM",          op::translate_cumsum                           },
+        {"GGML_OP_FILL",            op::translate_fill                             },
+        {"GGML_OP_DIAG",            op::translate_diag                             },
+        {"GGML_OP_TRI",             op::translate_tri                              },
+        {"GGML_OP_SET",             op::translate_set                              },
+        {"GGML_OP_POOL_2D",         op::translate_pool_2d                          },
+        {"GGML_OP_ROLL",            op::translate_roll                             },
+        {"GGML_OP_UPSCALE",         op::translate_upscale                          },
+        {"GGML_OP_CONV_2D",         op::translate_conv_2d                          },
+        {"GGML_OP_CONV_2D_DW",      op::translate_conv_2d_dw                       },
+        {"GGML_OP_CONV_TRANSPOSE_1D", op::translate_conv_transpose_1d             },
+        {"GGML_OP_CONV_TRANSPOSE_2D", op::translate_conv_transpose_2d             },
+        {"GGML_OP_CONV_3D",         op::translate_conv_3d                          },
+        // solve_tri has accuracy issues on GPU
+        // {"GGML_OP_SOLVE_TRI",       op::translate_solve_tri                        },
+    };
+}
+
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/op_table.h

@@ -0,0 +1,83 @@
+#pragma once
+
+#include "node_context.h"
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+
+namespace op {
+
+#define GGML_OP_CONVERTER(op) OutputVector op(const NodeContext & context)
+
+GGML_OP_CONVERTER(translate_add);
+GGML_OP_CONVERTER(translate_cont);
+GGML_OP_CONVERTER(translate_concat);
+GGML_OP_CONVERTER(translate_add_id);
+GGML_OP_CONVERTER(translate_div);
+GGML_OP_CONVERTER(translate_fill);
+GGML_OP_CONVERTER(translate_get_rows);
+GGML_OP_CONVERTER(translate_im2col);
+GGML_OP_CONVERTER(translate_im2col_3d);
+GGML_OP_CONVERTER(translate_mulmat);
+GGML_OP_CONVERTER(translate_mul_mat_id);
+GGML_OP_CONVERTER(translate_permute);
+GGML_OP_CONVERTER(translate_reshape);
+GGML_OP_CONVERTER(translate_rms_norm);
+GGML_OP_CONVERTER(translate_norm);
+GGML_OP_CONVERTER(translate_l2_norm);
+GGML_OP_CONVERTER(translate_sum_rows);
+GGML_OP_CONVERTER(translate_sqr);
+GGML_OP_CONVERTER(translate_rope);
+GGML_OP_CONVERTER(translate_scale);
+GGML_OP_CONVERTER(translate_sqrt);
+GGML_OP_CONVERTER(translate_unary_softplus);
+GGML_OP_CONVERTER(translate_soft_max);
+GGML_OP_CONVERTER(translate_transpose);
+GGML_OP_CONVERTER(translate_view);
+GGML_OP_CONVERTER(translate_glu_swiglu);
+GGML_OP_CONVERTER(translate_glu_swiglu_oai);
+GGML_OP_CONVERTER(translate_glu_swiglu_clamp);
+GGML_OP_CONVERTER(translate_glu_geglu);
+GGML_OP_CONVERTER(translate_glu_geglu_quick);
+GGML_OP_CONVERTER(translate_set_rows);
+GGML_OP_CONVERTER(translate_cpy);
+GGML_OP_CONVERTER(translate_argsort);
+GGML_OP_CONVERTER(translate_flash_attn_ext);
+GGML_OP_CONVERTER(translate_clamp);
+GGML_OP_CONVERTER(translate_pad);
+GGML_OP_CONVERTER(translate_ssm_conv);
+GGML_OP_CONVERTER(translate_gated_delta_net);
+GGML_OP_CONVERTER(translate_repeat);
+GGML_OP_CONVERTER(translate_cumsum);
+GGML_OP_CONVERTER(translate_fill);
+GGML_OP_CONVERTER(translate_set);
+GGML_OP_CONVERTER(translate_diag);
+GGML_OP_CONVERTER(translate_tri);
+GGML_OP_CONVERTER(translate_solve_tri);
+GGML_OP_CONVERTER(translate_pool_2d);
+GGML_OP_CONVERTER(translate_roll);
+GGML_OP_CONVERTER(translate_upscale);
+GGML_OP_CONVERTER(translate_mean);
+GGML_OP_CONVERTER(translate_sum);
+GGML_OP_CONVERTER(translate_unary_gelu);
+GGML_OP_CONVERTER(translate_unary_gelu_erf);
+GGML_OP_CONVERTER(translate_unary_gelu_quick);
+GGML_OP_CONVERTER(translate_unary_elu);
+GGML_OP_CONVERTER(translate_unary_hardsigmoid);
+GGML_OP_CONVERTER(translate_unary_step);
+GGML_OP_CONVERTER(translate_unary_round);
+GGML_OP_CONVERTER(translate_unary_expm1);
+GGML_OP_CONVERTER(translate_conv_2d);
+GGML_OP_CONVERTER(translate_conv_2d_dw);
+GGML_OP_CONVERTER(translate_conv_transpose_1d);
+GGML_OP_CONVERTER(translate_conv_transpose_2d);
+GGML_OP_CONVERTER(translate_conv_3d);
+
+}  // namespace op
+
+std::unordered_map<std::string, CreatorFunction> get_supported_ops();
+
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/pass/align_eltwise_ranks.cpp

@@ -0,0 +1,96 @@
+#include "align_eltwise_ranks.h"
+
+#include <numeric>
+#include <openvino/op/add.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/divide.hpp>
+#include <openvino/op/multiply.hpp>
+#include <openvino/op/sqrt.hpp>
+#include <openvino/op/subtract.hpp>
+#include <openvino/op/unsqueeze.hpp>
+#include <openvino/pass/pattern/op/wrap_type.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace pass {
+
+namespace {
+
+// True for an RMS-norm output, x * (1 / sqrt(mean(x^2) + eps)) as translate_rms_norm builds it, optionally
+// scaled by the norm weight.
+bool is_rms_norm_output(const ov::Output<ov::Node> & value, int depth = 1) {
+    const auto * node = value.get_node();
+    if (!ov::is_type<ov::op::v1::Multiply>(node)) {
+        return false;
+    }
+    for (const auto & input : node->input_values()) {
+        const auto * src = input.get_node();
+        if (ov::is_type<ov::op::v1::Divide>(src) && ov::is_type<ov::op::v0::Sqrt>(src->get_input_node_ptr(1))) {
+            return true;
+        }
+        if (depth > 0 && is_rms_norm_output(input, depth - 1)) {
+            return true;
+        }
+    }
+    return false;
+}
+
+}  // namespace
+
+AlignEltwiseOperandRanks::AlignEltwiseOperandRanks() {
+    auto eltwise_m = ov::pass::pattern::wrap_type<ov::op::v1::Add, ov::op::v1::Multiply, ov::op::v1::Subtract>();
+
+    const auto callback = [this](ov::pass::pattern::Matcher & m) {
+        auto node = m.get_match_root();
+        if (node->get_input_size() != 2) {
+            return false;
+        }
+
+        auto lhs = node->input_value(0);
+        auto rhs = node->input_value(1);
+
+        // A tensor-vs-Constant mismatch is the norm's own eps / 1-over-sqrt arithmetic. That
+        // folds into the `rms` primitive itself instead of becoming a fused eltwise post-op,
+        // so it is not affected and is left alone.
+        if (ov::is_type<ov::op::v0::Constant>(lhs.get_node()) ||
+            ov::is_type<ov::op::v0::Constant>(rhs.get_node())) {
+            return false;
+        }
+
+        const auto lhs_rank = lhs.get_partial_shape().rank();
+        const auto rhs_rank = rhs.get_partial_shape().rank();
+        if (lhs_rank.is_dynamic() || rhs_rank.is_dynamic() || lhs_rank == rhs_rank) {
+            return false;
+        }
+
+        const size_t shorter_idx = lhs_rank.get_length() < rhs_rank.get_length() ? 0 : 1;
+        const auto & shorter = shorter_idx == 0 ? lhs : rhs;
+
+        // The defect is with the norm output as the higher-rank operand. When the norm output is the
+        // lower-rank one (gemma-3 adds it to a rank-4 residual), unsqueezing it puts the Unsqueeze between
+        // `rms` and its post-op, and the GPU plugin then computes the layer wrongly.
+        if (is_rms_norm_output(shorter)) {
+            return false;
+        }
+        const int64_t diff = std::abs(lhs_rank.get_length() - rhs_rank.get_length());
+
+        std::vector<int64_t> axes(static_cast<size_t>(diff));
+        std::iota(axes.begin(), axes.end(), 0);
+        auto unsqueeze = std::make_shared<ov::op::v0::Unsqueeze>(
+            shorter, ov::op::v0::Constant::create(ov::element::i64, ov::Shape{ axes.size() }, axes));
+
+        node->input(shorter_idx).replace_source_output(unsqueeze->output(0));
+        register_new_node(unsqueeze);
+        return true;
+    };
+
+    register_matcher(
+        std::make_shared<ov::pass::pattern::Matcher>(eltwise_m, "ov::frontend::ggml::pass::AlignEltwiseOperandRanks"),
+        callback);
+}
+
+}  // namespace pass
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/pass/align_eltwise_ranks.h

@@ -0,0 +1,39 @@
+#pragma once
+
+#include <openvino/pass/matcher_pass.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace pass {
+
+// Give a binary eltwise op's two operands the same rank, by unsqueezing leading axes onto the
+// shorter one.
+//
+// This is semantically a no-op -- NUMPY broadcasting already left-pads the lower-rank operand
+// with 1s, and the result rank is max(rank_a, rank_b) either way. It exists purely to work
+// around an OpenVINO GPU-plugin defect: an eltwise op whose operands differ in rank is computed
+// incorrectly once the plugin fuses it as a post-op into an `rms` primitive. gemma-4 dense hits
+// this under stateful execution, where the layer tail adds a rank-3 residual to the rank-4
+// RMS-norm output; the model then degenerates into repeated tokens on GPU while CPU is correct.
+// Equalising the ranks keeps the fusion and makes it compute the right answer.
+//
+// Deliberately a graph pass rather than something the op translators do: rank is load-bearing
+// during translation (several translators and later passes read operand ranks), and rewriting
+// operands mid-translate breaks the attention path. Running after the graph is complete avoids
+// that entirely.
+//
+// Only applies when both operands are real tensors -- a tensor-vs-Constant mismatch is the
+// RMS norm's own eps/rsqrt arithmetic, which folds into the `rms` primitive rather than
+// becoming a fused post-op, and is not affected by the defect. The norm output itself is never
+// unsqueezed: when it is the lower-rank operand (gemma-3), that breaks the fused path instead.
+class AlignEltwiseOperandRanks : public ov::pass::MatcherPass {
+public:
+    OPENVINO_MATCHER_PASS_RTTI("ov::frontend::ggml::pass::AlignEltwiseOperandRanks")
+    AlignEltwiseOperandRanks();
+};
+
+}  // namespace pass
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/pass/fuse_argsort_topk.cpp

@@ -0,0 +1,70 @@
+#include "fuse_argsort_topk.h"
+
+#include <openvino/op/constant.hpp>
+#include <openvino/op/slice.hpp>
+#include <openvino/op/topk.hpp>
+#include <openvino/pass/pattern/op/wrap_type.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace pass {
+
+FuseArgsortTopK::FuseArgsortTopK() {
+    // GGML argsort materializes the full ordering before a VIEW keeps the prefix.
+    // Let TopK produce only that prefix when all index consumers are such views.
+    auto pattern = ov::pass::pattern::wrap_type<ov::op::v11::TopK>();
+
+    const auto callback = [](ov::pass::pattern::Matcher & m) {
+        auto topk = ov::as_type_ptr<ov::op::v11::TopK>(m.get_match_root());
+        if (!topk->output(0).get_target_inputs().empty() || topk->output(1).get_target_inputs().empty() ||
+            topk->get_sort_type() != ov::op::v11::TopK::SortType::SORT_VALUES) {
+            return false;
+        }
+
+        const auto shape = topk->get_output_partial_shape(1);
+        if (shape.rank().is_dynamic()) {
+            return false;
+        }
+        const int64_t rank = shape.rank().get_length();
+        const int64_t axis = topk->get_axis();
+        if (shape[axis].is_dynamic()) {
+            return false;
+        }
+
+        int64_t prefix = -1;
+        for (const auto & input : topk->output(1).get_target_inputs()) {
+            const auto * slice = ov::as_type<ov::op::v8::Slice>(input.get_node());
+            if (!slice || input.get_index() != 0 || slice->get_input_size() != 5) {
+                return false;
+            }
+            int64_t values[4];
+            for (size_t i = 0; i < 4; ++i) {
+                auto value = ov::as_type_ptr<ov::op::v0::Constant>(slice->get_input_node_shared_ptr(i + 1));
+                if (!value || ov::shape_size(value->get_shape()) != 1) {
+                    return false;
+                }
+                values[i] = value->cast_vector<int64_t>()[0];
+            }
+            const int64_t slice_axis = values[3] < 0 ? values[3] + rank : values[3];
+            if (values[0] != 0 || values[2] != 1 || slice_axis != axis || values[1] <= 0 ||
+                values[1] >= shape[axis].get_length() || (prefix != -1 && prefix != values[1])) {
+                return false;
+            }
+            prefix = values[1];
+        }
+
+        // ggml_argsort_top_k sorts the full row, then exposes only its first k indices.
+        // Keep other TopK forms unchanged because their ordering can be observable.
+        topk->set_argument(1, ov::op::v0::Constant::create(ov::element::i64, ov::Shape{}, {prefix}));
+        topk->validate_and_infer_types();
+        return true;
+    };
+
+    register_matcher(std::make_shared<ov::pass::pattern::Matcher>(pattern, "ov::frontend::ggml::pass::FuseArgsortTopK"), callback);
+}
+
+}  // namespace pass
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/pass/fuse_argsort_topk.h

@@ -0,0 +1,19 @@
+#pragma once
+
+#include <openvino/pass/matcher_pass.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace pass {
+
+class FuseArgsortTopK : public ov::pass::MatcherPass {
+public:
+    OPENVINO_MATCHER_PASS_RTTI("ov::frontend::ggml::pass::FuseArgsortTopK")
+    FuseArgsortTopK();
+};
+
+}  // namespace pass
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/pass/fuse_moe_compressed.cpp

@@ -0,0 +1,553 @@
+#include "fuse_moe_compressed.h"
+
+#include <cstring>
+#include <limits>
+#include <set>
+#include <memory>
+#include <openvino/core/graph_util.hpp>
+#include <openvino/core/rt_info.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/gelu.hpp>
+#include <openvino/op/multiply.hpp>
+#include <openvino/op/reduce_sum.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/slice.hpp>
+#include <openvino/op/squeeze.hpp>
+#include <openvino/op/subtract.hpp>
+#include <openvino/op/swish.hpp>
+#include <openvino/op/transpose.hpp>
+#include <openvino/op/unsqueeze.hpp>
+#include <openvino/pass/constant_folding.hpp>
+#include <openvino/pass/pattern/op/optional.hpp>
+#include <openvino/pass/pattern/op/wrap_type.hpp>
+
+#include "../op/gather_matmul.hpp"
+#include "../op/moe_compressed.hpp"
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace pass {
+
+namespace {
+
+struct dequant_inputs {
+    ov::Output<ov::Node> weight;
+    ov::Output<ov::Node> scale;
+    ov::Output<ov::Node> zp;
+    bool has_zp = false;
+    bool ok = false;
+};
+
+// Peel the chain built by make_int4_weights/make_int8_weights back to its Constant inputs.
+// Grouped weights keep the pre-Reshape rank-4 form [n_expert, n, k/group, group] with scale
+// and zp at [n_expert, n, k/group, 1], which is the layout MOECompressed expects. Channel-wise
+// weights stay rank-3 with a rank-3 scale and carry no zp.
+dequant_inputs unwrap_dequant(const ov::Output<ov::Node> & b) {
+    dequant_inputs res;
+
+    auto node = b.get_node_shared_ptr();
+    while (ov::is_type<ov::op::v0::Convert>(node) || ov::is_type<ov::op::v1::Reshape>(node)) {
+        node = node->get_input_node_shared_ptr(0);
+    }
+
+    auto mul = ov::as_type_ptr<ov::op::v1::Multiply>(node);
+    if (!mul) {
+        return res;
+    }
+    res.scale = mul->input_value(1);
+
+    auto lhs = mul->get_input_node_shared_ptr(0);
+    if (auto sub = ov::as_type_ptr<ov::op::v1::Subtract>(lhs)) {
+        // Take the zero point down to its Constant: an integer zp is wrapped in a Convert to f16,
+        // and the op wants the integer form. A natively quantized expert instead carries an exact
+        // f16 zp (-min/scale) with no integer behind it, which the MoE kernel does not accept.
+        auto zp_node = sub->get_input_node_shared_ptr(1);
+        while (ov::is_type<ov::op::v0::Convert>(zp_node)) {
+            zp_node = zp_node->get_input_node_shared_ptr(0);
+        }
+        res.zp = zp_node->output(0);
+        res.has_zp = true;
+        lhs = sub->get_input_node_shared_ptr(0);
+    }
+    while (ov::is_type<ov::op::v0::Convert>(lhs)) {
+        lhs = lhs->get_input_node_shared_ptr(0);
+    }
+    if (!ov::is_type<ov::op::v0::Constant>(lhs)) {
+        return res;
+    }
+
+    res.weight = lhs->output(0);
+    res.ok = res.scale.get_partial_shape().is_static() && res.weight.get_partial_shape().is_static();
+    return res;
+}
+
+size_t logical_k(const ov::Shape & shape) {
+    return shape.size() == 4 ? shape[2] * shape[3] : shape.back();
+}
+
+// Slice a [n_expert, m, ...] compressed weight/scale/zp Constant along axis 1 (m), [begin, end).
+// Built by copying raw bytes directly instead of a graph Slice op: CommonOptimizations rewrites
+// v8::Slice into v1::StridedSlice for constant folding regardless of disable_constant_folding
+// (the rewrite doesn't carry the marking over), and that reference evaluator crashes on
+// sub-byte (u4/i4) element types. Each axis-1 "row" (the trailing dims) is confirmed
+// byte-aligned here -- weight/scale/zp trailing sizes are always whole groups -- so a
+// byte-range memcpy per row is exact for both regular and sub-byte element types.
+ov::Output<ov::Node> slice_experts_dim1(const ov::Output<ov::Node> & tensor, int64_t begin, int64_t end) {
+    auto constant = ov::as_type_ptr<ov::op::v0::Constant>(tensor.get_node_shared_ptr());
+    OPENVINO_ASSERT(constant, "slice_experts_dim1 expects a Constant input");
+
+    const auto & shape = constant->get_shape();
+    OPENVINO_ASSERT(shape.size() >= 2, "slice_experts_dim1 expects rank >= 2");
+    const auto & type = constant->get_element_type();
+
+    size_t row_elems = 1;
+    for (size_t i = 2; i < shape.size(); ++i) {
+        row_elems *= shape[i];
+    }
+    const size_t row_bits = row_elems * type.bitwidth();
+    OPENVINO_ASSERT(row_bits % 8 == 0, "slice_experts_dim1: row is not byte-aligned");
+    const size_t row_bytes = row_bits / 8;
+
+    ov::Shape out_shape = shape;
+    out_shape[1] = static_cast<size_t>(end - begin);
+
+    std::vector<uint8_t> out_data(shape[0] * static_cast<size_t>(end - begin) * row_bytes);
+    const auto * src = static_cast<const uint8_t *>(constant->get_data_ptr());
+    const size_t src_row_bytes = shape[1] * row_bytes;
+    for (size_t e = 0; e < shape[0]; ++e) {
+        std::memcpy(out_data.data() + e * static_cast<size_t>(end - begin) * row_bytes,
+                   src + e * src_row_bytes + static_cast<size_t>(begin) * row_bytes,
+                   static_cast<size_t>(end - begin) * row_bytes);
+    }
+
+    return std::make_shared<ov::op::v0::Constant>(type, out_shape, out_data.data());
+}
+
+}  // namespace
+
+FuseMoeCompressed::FuseMoeCompressed() {
+    using namespace ov::pass::pattern;
+
+    // The gate and up projections each get their own Reshape/Transpose of the hidden state and
+    // their own Reshape of the routing ids, so every branch needs its own sub-pattern. On GPU
+    // mul_mat_id also converts the activations to f16 before the op and back to f32 after it,
+    // so those Converts are matched as optional.
+    auto hidden_gate_m = any_input();
+    auto a_gate_reshape_m = wrap_type<ov::op::v1::Reshape>({ hidden_gate_m, any_input() });
+    auto a_gate_m =
+        wrap_type<ov::op::v1::Transpose>({ optional<ov::op::v0::Convert>({ a_gate_reshape_m }), any_input() });
+    auto hidden_up_m = any_input();
+    auto a_up_m = wrap_type<ov::op::v1::Transpose>(
+        { optional<ov::op::v0::Convert>({ wrap_type<ov::op::v1::Reshape>({ hidden_up_m, any_input() }) }),
+          any_input() });
+
+    auto gate_w_m = any_input();
+    auto up_w_m = any_input();
+    auto down_w_m = any_input();
+    auto ids_gate_m = any_input();
+    auto ids_up_m = any_input();
+    auto ids_down_m = any_input();
+
+    auto bgm_gate_m = wrap_type<ov::op::internal::GatherMatmul>({ a_gate_m, gate_w_m, ids_gate_m, any_input() });
+    auto gate_u_m = optional<ov::op::v0::Convert>({ wrap_type<ov::op::v0::Unsqueeze>(
+        { wrap_type<ov::op::v1::Transpose>({ bgm_gate_m, any_input() }), any_input() }) });
+
+    auto silu_m = wrap_type<ov::op::v4::Swish>({ gate_u_m });
+
+    auto bgm_up_m = wrap_type<ov::op::internal::GatherMatmul>({ a_up_m, up_w_m, ids_up_m, any_input() });
+    auto up_u_m = optional<ov::op::v0::Convert>({ wrap_type<ov::op::v0::Unsqueeze>(
+        { wrap_type<ov::op::v1::Transpose>({ bgm_up_m, any_input() }), any_input() }) });
+    auto swiglu_m = wrap_type<ov::op::v1::Multiply>({ silu_m, up_u_m });
+
+    auto d_t_m = wrap_type<ov::op::v1::Transpose>(
+        { optional<ov::op::v0::Convert>({ wrap_type<ov::op::v1::Reshape>({ swiglu_m, any_input() }) }),
+          any_input() });
+    auto bgm_down_m = wrap_type<ov::op::internal::GatherMatmul>({ d_t_m, down_w_m, ids_down_m, any_input() });
+    auto down_u_m = optional<ov::op::v0::Convert>({ wrap_type<ov::op::v0::Unsqueeze>(
+        { wrap_type<ov::op::v1::Transpose>({ bgm_down_m, any_input() }), any_input() }) });
+
+    auto routing_m = any_input();
+    auto weighted_m = wrap_type<ov::op::v1::Multiply>({ down_u_m, routing_m });
+    auto root_m = wrap_type<ov::op::v1::ReduceSum>({ weighted_m, any_input() });
+
+    const auto callback = [=](Matcher & m) {
+        auto & pm = m.get_pattern_value_map();
+
+        const auto gate = unwrap_dequant(pm.at(gate_w_m));
+        const auto up = unwrap_dequant(pm.at(up_w_m));
+        const auto down = unwrap_dequant(pm.at(down_w_m));
+        if (!gate.ok || !up.ok || !down.ok) {
+            return false;
+        }
+
+        const auto gate_shape = gate.weight.get_shape();
+        const auto up_shape = up.weight.get_shape();
+        const auto down_shape = down.weight.get_shape();
+        if (gate_shape != up_shape || gate_shape.size() < 3 || down_shape.size() < 3) {
+            return false;
+        }
+
+        // MOECompressed carries one group_size and one has_zp for all three projections, so a
+        // model whose down-proj is quantized differently from gate/up cannot be described. This
+        // happens when ggml requantizes Q5_K/Q6_K experts to channel-wise int8.
+        if (gate.has_zp != down.has_zp || gate_shape.size() != down_shape.size()) {
+            return false;
+        }
+
+        // The kernel only takes an integer zero point (moe_3gemm_swiglu_opt validate_impl).
+        if (gate.has_zp) {
+            static const std::set<ov::element::Type> int_zp_types = { ov::element::u4, ov::element::i4,
+                                                                     ov::element::u8, ov::element::i8 };
+            if (int_zp_types.count(gate.zp.get_element_type()) == 0 ||
+                int_zp_types.count(down.zp.get_element_type()) == 0) {
+                return false;
+            }
+        }
+
+        // Config holds a single group_size for all three projections.
+        const auto group_of = [](const dequant_inputs & w) {
+            const auto s = w.weight.get_shape();
+            return s.size() == 4 ? s[3] : logical_k(s);
+        };
+        if (group_of(gate) != group_of(up) || group_of(gate) != group_of(down)) {
+            return false;
+        }
+
+        // all three branches must route the same hidden state through the same experts
+        if (pm.at(hidden_gate_m) != pm.at(hidden_up_m)) {
+            return false;
+        }
+
+        auto ids = pm.at(ids_down_m);
+        const auto ids_pshape = ids.get_partial_shape();
+        if (ids_pshape.rank().is_dynamic() || ids_pshape[ids_pshape.rank().get_length() - 1].is_dynamic()) {
+            return false;
+        }
+        const size_t top_k = ids_pshape[ids_pshape.rank().get_length() - 1].get_length();
+
+        // routing weights arrive as [1, n_tokens, top_k, 1]; the op wants [..., top_k]
+        auto routing = pm.at(routing_m);
+        const auto routing_pshape = routing.get_partial_shape();
+        if (routing_pshape.rank().is_dynamic() || routing_pshape.rank().get_length() != 4 ||
+            routing_pshape[3] != 1) {
+            return false;
+        }
+        // MOE requires routing weights and ids to have the same shape. Drop the trailing 1 of the
+        // routing weights and give the ids the leading batch dim, so both become [1, n_tokens, top_k].
+        routing = std::make_shared<ov::op::v0::Squeeze>(
+            routing, ov::op::v0::Constant::create(ov::element::i64, ov::Shape{ 1 }, { 3 }));
+        if (ids_pshape.rank().get_length() == 2) {
+            ids = std::make_shared<ov::op::v0::Unsqueeze>(
+                ids, ov::op::v0::Constant::create(ov::element::i64, ov::Shape{ 1 }, { 0 }));
+        }
+        if (routing.get_partial_shape() != ids.get_partial_shape()) {
+            return false;
+        }
+
+        const size_t down_k = logical_k(down_shape);
+        const auto down_scale_shape = down.scale.get_shape();
+        const size_t down_groups = down_scale_shape.size() >= 3 ? down_scale_shape[2] : 1;
+
+        ov::op::internal::MOECompressed::Config config;
+        config.expert_type = ov::op::internal::MOE::Expert_type::GEMM3_SWIGLU;
+        config.activation_type = ov::op::internal::MOE::Activation_type::SWIGLU;
+        config.expert_alpha = 0.0f;
+        config.expert_beta = 1.0f;
+        config.gate_idx = 0;
+        config.hidden_size = logical_k(gate_shape);
+        config.inter_size = gate_shape[1];
+        config.num_expert = gate_shape[0];
+        config.num_shared_expert = 0;
+        config.top_k = top_k;
+        config.group_size = down_groups <= 1 ? std::numeric_limits<size_t>::max() : down_k / down_groups;
+        config.has_batch_dim = true;
+        config.has_zp = gate.has_zp;
+        // dynamic makes the output follow the hidden state, so the plugin can lower this region
+        // to f16 together with the rest of the graph
+        config.out_type = ov::element::dynamic;
+
+        auto absent_zp = [] {
+            auto zp = std::make_shared<ov::op::v0::Constant>(ov::element::dynamic, ov::Shape{ 0 });
+            ov::pass::disable_constant_folding(zp);
+            return zp->output(0);
+        };
+
+        // MOE takes its output type from the hidden state. Transpose the activations before the
+        // f16 Convert that mul_mat_id adds on GPU, so the op stays f32 like the block it replaces
+        // and the plugin can lower the whole region uniformly.
+        const auto a_transpose = pm.at(a_gate_m).get_node_shared_ptr();
+        ov::Output<ov::Node> hidden =
+            std::make_shared<ov::op::v1::Transpose>(pm.at(a_gate_reshape_m), a_transpose->input_value(1));
+
+        const ov::OutputVector args = {
+            hidden, routing, ids,
+            gate.weight,  gate.scale,   gate.has_zp ? gate.zp : absent_zp(),
+            up.weight,    up.scale,     up.has_zp ? up.zp : absent_zp(),
+            down.weight,  down.scale,   down.has_zp ? down.zp : absent_zp(),
+        };
+
+        auto moe = std::make_shared<ov::op::internal::MOECompressed>(args, config);
+
+        // MOE takes its output type from the hidden state, which is f16 on GPU, while the rest of
+        // the ggml graph works in f32.
+        ov::Output<ov::Node> result = moe->output(0);
+        const auto root_type = m.get_match_root()->get_output_element_type(0);
+        if (result.get_element_type() != root_type) {
+            result = std::make_shared<ov::op::v0::Convert>(result, root_type);
+        }
+
+        result.get_node_shared_ptr()->set_friendly_name(m.get_match_root()->get_friendly_name());
+        ov::copy_runtime_info(m.get_matched_nodes(), result.get_node_shared_ptr());
+        ov::replace_node(m.get_match_root(), result.get_node_shared_ptr());
+        register_new_node(moe);
+        return true;
+    };
+
+    register_matcher(std::make_shared<Matcher>(root_m, "ov::frontend::ggml::pass::FuseMoeCompressed"), callback);
+}
+
+FuseMoeCompressedFusedGateUp::FuseMoeCompressedFusedGateUp() {
+    using namespace ov::pass::pattern;
+
+    // A single GatherMatmul computes the fused gate+up projection; the gate/up split happens
+    // AFTER the GEMM via two Slice ops on the last (feature) axis (see get_glu_inputs in
+    // op/glu_geglu.cpp), unlike FuseMoeCompressed's two-separate-GatherMatmul models.
+    auto hidden_m = any_input();
+    auto a_reshape_m = wrap_type<ov::op::v1::Reshape>({ hidden_m, any_input() });
+    auto a_m = wrap_type<ov::op::v1::Transpose>({ optional<ov::op::v0::Convert>({ a_reshape_m }), any_input() });
+
+    auto gate_up_w_m = any_input();
+    auto ids_gate_up_m = any_input();
+    auto bgm_fused_m = wrap_type<ov::op::internal::GatherMatmul>({ a_m, gate_up_w_m, ids_gate_up_m, any_input() });
+    // Two graph shapes reach here. The rank-4 (stateless) graph restores the batch dim with an
+    // Unsqueeze after the Transpose and may Convert afterwards:
+    //     GatherMatmul -> Transpose -> Unsqueeze -> [Convert] -> Slice
+    // The rank-3 (stateful) graph never drops to a batch dim at all, so there is no Unsqueeze
+    // here - it appears later, between the GEGLU and the down Reshape - and the Convert sits on
+    // the other side of the Transpose:
+    //     GatherMatmul -> Convert -> Transpose -> Slice
+    // optional<T>({a, b}) matches T(a, b) or bare a, so one pattern covers both.
+    auto gu_t_m =
+        wrap_type<ov::op::v1::Transpose>({ optional<ov::op::v0::Convert>({ bgm_fused_m }), any_input() });
+    auto gu_u_m =
+        optional<ov::op::v0::Convert>({ optional<ov::op::v0::Unsqueeze>({ gu_t_m, any_input() }) });
+
+    auto gate_slice_m = wrap_type<ov::op::v8::Slice>({ gu_u_m, any_input(), any_input(), any_input(), any_input() });
+    auto up_slice_m = wrap_type<ov::op::v8::Slice>({ gu_u_m, any_input(), any_input(), any_input(), any_input() });
+    auto gelu_m = wrap_type<ov::op::v7::Gelu>({ gate_slice_m });
+    auto geglu_m = wrap_type<ov::op::v1::Multiply>({ gelu_m, up_slice_m });
+
+    // The rank-3 graph inserts the Unsqueeze the gate/up branch lacked right here, before the
+    // Reshape that feeds the down projection; the rank-4 graph goes straight from the GEGLU
+    // into the Reshape.
+    auto d_t_m = wrap_type<ov::op::v1::Transpose>(
+        { optional<ov::op::v0::Convert>({ wrap_type<ov::op::v1::Reshape>(
+              { optional<ov::op::v0::Unsqueeze>({ geglu_m, any_input() }), any_input() }) }),
+          any_input() });
+    auto down_w_m = any_input();
+    auto ids_down_m = any_input();
+    auto bgm_down_m = wrap_type<ov::op::internal::GatherMatmul>({ d_t_m, down_w_m, ids_down_m, any_input() });
+    // Same two shapes as the gate/up branch above.
+    auto down_t_m =
+        wrap_type<ov::op::v1::Transpose>({ optional<ov::op::v0::Convert>({ bgm_down_m }), any_input() });
+    auto down_u_m =
+        optional<ov::op::v0::Convert>({ optional<ov::op::v0::Unsqueeze>({ down_t_m, any_input() }) });
+
+    // gemma-4 applies an extra per-expert output scale to the down projection before the
+    // router-weight multiply (llama-graph.cpp's ffn_down_exps.scale); FuseMoeCompressed's
+    // Qwen-shaped models have no such scale, so this node is specific to this pattern.
+    auto down_scale_m = any_input();
+    auto down_scaled_m = wrap_type<ov::op::v1::Multiply>({ down_u_m, down_scale_m });
+
+    auto routing_m = any_input();
+    auto weighted_m = wrap_type<ov::op::v1::Multiply>({ down_scaled_m, routing_m });
+    auto root_m = wrap_type<ov::op::v1::ReduceSum>({ weighted_m, any_input() });
+
+    const auto callback = [=](Matcher & m) {
+        auto & pm = m.get_pattern_value_map();
+
+        const auto gate_up = unwrap_dequant(pm.at(gate_up_w_m));
+        const auto down = unwrap_dequant(pm.at(down_w_m));
+        if (!gate_up.ok || !down.ok) {
+            return false;
+        }
+
+        // oneDNN's weight-decompression GEMM needs a real zero point, so a symmetric (no zp)
+        // projection cannot be fused here -- e.g. the GPU-only Q4_0 fallback in
+        // ggml_openvino_get_extracted_layout for grouped 8-bit expert weights.
+        if (!gate_up.has_zp || !down.has_zp) {
+            return false;
+        }
+
+        const auto gate_up_shape = gate_up.weight.get_shape();
+        const auto down_shape = down.weight.get_shape();
+        if (gate_up_shape.size() < 3 || down_shape.size() < 3 || gate_up_shape.size() != down_shape.size()) {
+            return false;
+        }
+
+        // The op reads the zero point straight off a weight port, so it must already be an
+        // integer Constant. Requantized experts (GGML_OPENVINO_REQUANT_KQUANT=q4_asym64_all)
+        // are; a native Q4_K expert keeps an exact f16 zp and is left to the unfused path
+        // rather than rounded here -- rounding would put a live node on that port and the
+        // expert GEMM would read the wrong zero point.
+        static const std::set<ov::element::Type> int_types = { ov::element::u4, ov::element::i4,
+                                                               ov::element::u8, ov::element::i8 };
+        if (int_types.count(gate_up.weight.get_element_type()) == 0 ||
+            int_types.count(down.weight.get_element_type()) == 0 ||
+            int_types.count(gate_up.zp.get_element_type()) == 0 ||
+            int_types.count(down.zp.get_element_type()) == 0) {
+            return false;
+        }
+
+        const auto group_of = [](const dequant_inputs & w) {
+            const auto s = w.weight.get_shape();
+            return s.size() == 4 ? s[3] : logical_k(s);
+        };
+        if (group_of(gate_up) != group_of(down)) {
+            return false;
+        }
+
+        auto gelu_node = ov::as_type_ptr<ov::op::v7::Gelu>(pm.at(gelu_m).get_node_shared_ptr());
+        if (!gelu_node || gelu_node->get_approximation_mode() != ov::op::GeluApproximationMode::ERF) {
+            return false;
+        }
+
+        // Read the actual gate/up split point from the matched Slice nodes instead of assuming
+        // inter_size/2, so this stays correct if the model's ffn_dim convention ever changes.
+        auto gate_slice = ov::as_type_ptr<ov::op::v8::Slice>(pm.at(gate_slice_m).get_node_shared_ptr());
+        auto up_slice = ov::as_type_ptr<ov::op::v8::Slice>(pm.at(up_slice_m).get_node_shared_ptr());
+        auto gate_end_c = ov::as_type_ptr<ov::op::v0::Constant>(gate_slice->input_value(2).get_node_shared_ptr());
+        auto up_begin_c = ov::as_type_ptr<ov::op::v0::Constant>(up_slice->input_value(1).get_node_shared_ptr());
+        if (!gate_end_c || !up_begin_c) {
+            return false;
+        }
+        const auto gate_end = gate_end_c->cast_vector<int64_t>().at(0);
+        const auto up_begin = up_begin_c->cast_vector<int64_t>().at(0);
+        if (gate_end != up_begin || gate_end <= 0 || gate_end >= static_cast<int64_t>(gate_up_shape[1])) {
+            return false;
+        }
+        const int64_t mid = gate_end;
+        const int64_t full = static_cast<int64_t>(gate_up_shape[1]);
+
+        auto ids = pm.at(ids_down_m);
+        const auto ids_pshape = ids.get_partial_shape();
+        if (ids_pshape.rank().is_dynamic() || ids_pshape[ids_pshape.rank().get_length() - 1].is_dynamic()) {
+            return false;
+        }
+        const size_t top_k = ids_pshape[ids_pshape.rank().get_length() - 1].get_length();
+
+        // Routing weights arrive as [1, n_tokens, top_k, 1] on the rank-4 (stateless) graph and
+        // as [n_tokens, top_k, 1] on the rank-3 (stateful) one, which is the same thing without
+        // the leading batch dim. Normalise the rank-3 form up to the rank-4 one so everything
+        // below - and the op's own config - stays in the shape that is already validated on the
+        // stateless path; the batch dim is taken back off the result at the end.
+        const auto routing_in = pm.at(routing_m);
+        const auto routing_pshape = routing_in.get_partial_shape();
+        const auto routing_rank = routing_pshape.rank();
+        if (routing_rank.is_dynamic() || (routing_rank.get_length() != 4 && routing_rank.get_length() != 3) ||
+            routing_pshape[routing_rank.get_length() - 1] != 1) {
+            return false;
+        }
+        const bool batchless = routing_rank.get_length() == 3;
+
+        // Fold gemma-4's per-expert output scale into the routing weights: the reduction is
+        // sum_e(routing[e] * scale[e] * down_out[e]), and MOECompressed only takes one
+        // per-expert weight, so pre-multiply it into routing here (same [.., top_k, 1] shape).
+        auto down_scale = pm.at(down_scale_m);
+        if (down_scale.get_partial_shape() != routing_pshape) {
+            return false;
+        }
+        ov::Output<ov::Node> routing = std::make_shared<ov::op::v1::Multiply>(routing_in, down_scale);
+        if (batchless) {
+            routing = std::make_shared<ov::op::v0::Unsqueeze>(
+                routing, ov::op::v0::Constant::create(ov::element::i64, ov::Shape{ 1 }, { 0 }));
+        }
+        routing = std::make_shared<ov::op::v0::Squeeze>(
+            routing, ov::op::v0::Constant::create(ov::element::i64, ov::Shape{ 1 }, { 3 }));
+        if (ids_pshape.rank().get_length() == 2) {
+            ids = std::make_shared<ov::op::v0::Unsqueeze>(
+                ids, ov::op::v0::Constant::create(ov::element::i64, ov::Shape{ 1 }, { 0 }));
+        }
+        if (routing.get_partial_shape() != ids.get_partial_shape()) {
+            return false;
+        }
+
+        const size_t down_k = logical_k(down_shape);
+        const auto down_scale_shape = down.scale.get_shape();
+        const size_t down_groups = down_scale_shape.size() >= 3 ? down_scale_shape[2] : 1;
+
+        // Split the fused gate_up weight/scale/zp on the output axis. slice_experts_dim1 copies
+        // raw bytes out of the Constant, so each half stays a Constant -- which is what the op
+        // needs on its weight ports.
+        const ov::Output<ov::Node> gate_weight = slice_experts_dim1(gate_up.weight, 0, mid);
+        const ov::Output<ov::Node> up_weight = slice_experts_dim1(gate_up.weight, mid, full);
+        const ov::Output<ov::Node> gate_scale = slice_experts_dim1(gate_up.scale, 0, mid);
+        const ov::Output<ov::Node> up_scale = slice_experts_dim1(gate_up.scale, mid, full);
+        const ov::Output<ov::Node> gate_zp = slice_experts_dim1(gate_up.zp, 0, mid);
+        const ov::Output<ov::Node> up_zp = slice_experts_dim1(gate_up.zp, mid, full);
+
+        ov::op::internal::MOECompressed::Config config;
+        config.expert_type = ov::op::internal::MOE::Expert_type::GEMM3_SWIGLU;
+        config.activation_type = ov::op::internal::MOE::Activation_type::GEGLU_ERF;
+        config.expert_alpha = 0.0f;
+        config.expert_beta = 1.0f;
+        config.gate_idx = 0;
+        config.hidden_size = logical_k(gate_up_shape);
+        config.inter_size = static_cast<size_t>(mid);
+        config.num_expert = gate_up_shape[0];
+        config.num_shared_expert = 0;
+        config.top_k = top_k;
+        config.group_size = down_groups <= 1 ? std::numeric_limits<size_t>::max() : down_k / down_groups;
+        config.has_batch_dim = true;
+        config.has_zp = true;
+        config.out_type = ov::element::dynamic;
+
+        // Rebuild the activations Transpose before mul_mat_id's f16 GPU Convert, so the op
+        // stays f32 like the block it replaces (mirrors FuseMoeCompressed).
+        const auto a_transpose = pm.at(a_m).get_node_shared_ptr();
+        ov::Output<ov::Node> hidden =
+            std::make_shared<ov::op::v1::Transpose>(pm.at(a_reshape_m), a_transpose->input_value(1));
+
+        // w0 is the activated (gate) lane and w1 the multiplied (up) lane -- see
+        // moe_3gemm_swiglu_opt.cpp: "scratch.up = up(x) * silu(gate(x))".
+        const ov::OutputVector args = {
+            hidden,      routing,     ids,
+            gate_weight, gate_scale,  gate_zp,
+            up_weight,   up_scale,    up_zp,
+            down.weight, down.scale,  down.zp,
+        };
+
+        auto moe = std::make_shared<ov::op::internal::MOECompressed>(args, config);
+
+        ov::Output<ov::Node> result = moe->output(0);
+        // The op was fed the batched form, so it produces [1, n_tokens, hidden]. On the rank-3
+        // graph the ReduceSum being replaced is [n_tokens, hidden], so drop the batch dim again.
+        if (batchless) {
+            result = std::make_shared<ov::op::v0::Squeeze>(
+                result, ov::op::v0::Constant::create(ov::element::i64, ov::Shape{ 1 }, { 0 }));
+        }
+        const auto root_type = m.get_match_root()->get_output_element_type(0);
+        if (result.get_element_type() != root_type) {
+            result = std::make_shared<ov::op::v0::Convert>(result, root_type);
+        }
+        if (result.get_partial_shape() != m.get_match_root()->get_output_partial_shape(0)) {
+            return false;
+        }
+
+        result.get_node_shared_ptr()->set_friendly_name(m.get_match_root()->get_friendly_name());
+        ov::copy_runtime_info(m.get_matched_nodes(), result.get_node_shared_ptr());
+        ov::replace_node(m.get_match_root(), result.get_node_shared_ptr());
+        register_new_node(moe);
+        return true;
+    };
+
+    register_matcher(
+        std::make_shared<Matcher>(root_m, "ov::frontend::ggml::pass::FuseMoeCompressedFusedGateUp"), callback);
+}
+
+}  // namespace pass
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/pass/fuse_moe_compressed.h

@@ -0,0 +1,30 @@
+#include "openvino/pass/matcher_pass.hpp"
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace pass {
+
+// Folds the MoE expert block emitted for MUL_MAT_ID (3 GatherMatmul + SwiGLU + routing
+// weighting + expert reduction) into a single ov::op::internal::MOECompressed.
+class FuseMoeCompressed : public ov::pass::MatcherPass {
+public:
+    OPENVINO_MATCHER_PASS_RTTI("ov::frontend::ggml::pass::FuseMoeCompressed")
+    FuseMoeCompressed();
+};
+
+// Folds the MoE expert block emitted for a model whose gate and up projections share one
+// fused MUL_MAT_ID weight (gemma-4: one GatherMatmul + Slice/Slice split, GEGLU activation)
+// into a single ov::op::internal::MOECompressed, GEMM3_SWIGLU/GEGLU_ERF. Splits the fused
+// weight/scale/zp into gate/up halves so it lands on the same 3-GEMM expert-grouped kernel
+// FuseMoeCompressed uses for models with separate gate/up weights.
+class FuseMoeCompressedFusedGateUp : public ov::pass::MatcherPass {
+public:
+    OPENVINO_MATCHER_PASS_RTTI("ov::frontend::ggml::pass::FuseMoeCompressedFusedGateUp")
+    FuseMoeCompressedFusedGateUp();
+};
+
+}  // namespace pass
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/pass/fuse_moe_router.cpp

@@ -0,0 +1,153 @@
+#include "fuse_moe_router.h"
+
+#include <openvino/core/graph_util.hpp>
+#include <openvino/core/rt_info.hpp>
+#include <openvino/op/broadcast.hpp>
+#include <openvino/op/clamp.hpp>
+#include <openvino/op/concat.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/divide.hpp>
+#include <openvino/op/gather.hpp>
+#include <openvino/op/matmul.hpp>
+#include <openvino/op/reduce_sum.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/shape_of.hpp>
+#include <openvino/op/slice.hpp>
+#include <openvino/op/softmax.hpp>
+#include <openvino/op/squeeze.hpp>
+#include <openvino/op/tile.hpp>
+#include <openvino/op/topk.hpp>
+#include <openvino/op/unsqueeze.hpp>
+#include <openvino/pass/pattern/op/wrap_type.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace pass {
+
+namespace {
+
+bool constant_is(const ov::Output<ov::Node> & output, const std::vector<int64_t> & values) {
+    auto node = ov::as_type_ptr<ov::op::v0::Constant>(output.get_node_shared_ptr());
+    return node && node->get_element_type().is_integral_number() && node->cast_vector<int64_t>() == values;
+}
+
+}  // namespace
+
+FuseMoeRouter::FuseMoeRouter() {
+    using namespace ov::pass::pattern;
+    using namespace ov::op;
+
+    // Match the GGML rank-4 routing chain. The GPU plugin recognizes the resulting
+    // Softmax -> TopK -> ReduceSum -> Divide subgraph as MoERouterFused.
+    auto logits = wrap_type<v0::MatMul>();
+    auto softmax = wrap_type<v8::Softmax>({logits});
+    auto topk = wrap_type<v11::TopK>({softmax, any_input()});
+    topk->set_output_size(2);
+    auto ids = wrap_type<v8::Slice>({topk->output(1), any_input(), any_input(), any_input(), any_input()});
+    auto ids_2d = wrap_type<v0::Squeeze>({ids, any_input()});
+    auto probs = wrap_type<v1::Reshape>({softmax, any_input()});
+    auto data = wrap_type<v0::Squeeze>({probs, any_input()});
+    auto data_batch = wrap_type<v8::Gather>({wrap_type<v3::ShapeOf>({data}), any_input(), any_input()});
+    auto ids_count = wrap_type<v8::Gather>({wrap_type<v3::ShapeOf>({ids_2d}), any_input(), any_input()});
+    auto target = wrap_type<v0::Concat>({data_batch, ids_count});
+    auto broadcast = wrap_type<v3::Broadcast>({ids_2d, target});
+    auto gather = wrap_type<v8::Gather>({data, broadcast, any_input()});
+    auto weights_4d = wrap_type<v0::Unsqueeze>({gather, any_input()});
+    auto weights = wrap_type<v1::Reshape>({weights_4d, any_input()});
+    auto sum = wrap_type<v1::ReduceSum>({weights, any_input()});
+    auto clamp = wrap_type<v0::Clamp>({sum});
+    auto tile = wrap_type<v0::Tile>({clamp, any_input()});
+    auto norm = wrap_type<v1::Divide>({weights, tile});
+
+    const auto callback = [=](Matcher & m) {
+        const auto & pm = m.get_pattern_value_map();
+        const auto node = [&](const std::shared_ptr<ov::Node> & p) { return pm.at(p).get_node_shared_ptr(); };
+        const auto input_is = [&](const std::shared_ptr<ov::Node> & p, size_t i, const std::vector<int64_t> & v) {
+            return constant_is(node(p)->input_value(i), v);
+        };
+        auto mm = ov::as_type_ptr<v0::MatMul>(node(logits));
+        auto sm = ov::as_type_ptr<v8::Softmax>(node(softmax));
+        auto tk = ov::as_type_ptr<v11::TopK>(node(topk));
+        auto reduce = ov::as_type_ptr<v1::ReduceSum>(node(sum));
+        auto limit = ov::as_type_ptr<v0::Clamp>(node(clamp));
+        const auto shape = mm->get_output_partial_shape(0);
+        if (shape.rank() != 4 || shape[0] != 1 || shape[1] != 1 || shape[3].is_dynamic() ||
+            mm->get_transpose_a() || mm->get_input_partial_shape(0).rank() != 4 ||
+            mm->get_input_partial_shape(1).rank() != 2 || (sm->get_axis() != -1 && sm->get_axis() != 3) ||
+            tk->get_axis() != 3 || tk->get_mode() != v11::TopK::Mode::MAX ||
+            tk->get_sort_type() != v11::TopK::SortType::SORT_VALUES || tk->get_stable() ||
+            tk->get_index_element_type() != ov::element::i32 ||
+            !tk->output(0).get_target_inputs().empty()) {
+            return false;
+        }
+        const int64_t experts = shape[3].get_length();
+        auto end = ov::as_type_ptr<v0::Constant>(node(ids)->get_input_node_shared_ptr(2));
+        if (!end || !end->get_element_type().is_integral_number() || ov::shape_size(end->get_shape()) != 1) {
+            return false;
+        }
+        const int64_t k = end->cast_vector<int64_t>()[0];
+        const auto sorted_shape = tk->get_output_partial_shape(1);
+        if (k <= 0 || k > experts || sorted_shape[3].is_dynamic() || sorted_shape[3].get_length() < k ||
+            !input_is(probs, 1, {1, -1, experts, 1}) ||
+            !input_is(data, 1, {0}) || !input_is(ids_2d, 1, {0, 1}) ||
+            !input_is(weights_4d, 1, {0}) || !input_is(weights, 1, {1, 1, -1, k}) ||
+            ov::as_type_ptr<v1::Reshape>(node(probs))->get_special_zero() ||
+            ov::as_type_ptr<v1::Reshape>(node(weights))->get_special_zero() ||
+            !input_is(gather, 2, {1}) || ov::as_type_ptr<v8::Gather>(node(gather))->get_batch_dims() != 1 ||
+            !input_is(data_batch, 1, {0}) || !input_is(data_batch, 2, {0}) ||
+            !input_is(ids_count, 1, {1}) || !input_is(ids_count, 2, {0}) ||
+            ov::as_type_ptr<v8::Gather>(node(data_batch))->get_batch_dims() != 0 ||
+            ov::as_type_ptr<v8::Gather>(node(ids_count))->get_batch_dims() != 0 ||
+            ov::as_type_ptr<v0::Concat>(node(target))->get_axis() != 0 ||
+            ov::as_type_ptr<v3::Broadcast>(node(broadcast))->get_broadcast_spec().m_type != ov::op::BroadcastType::BIDIRECTIONAL ||
+            !reduce->get_keep_dims() || (!input_is(sum, 1, {-1}) && !input_is(sum, 1, {3})) ||
+            !input_is(tile, 1, {1, 1, 1, k}) ||
+            ov::as_type_ptr<v1::Divide>(node(norm))->get_autob().m_type != ov::op::AutoBroadcastType::NUMPY) {
+            return false;
+        }
+        // The top-k softmax sum is at least k / experts. Leave margin for rounding.
+        // This guard is only valid for unbiased softmax routing.
+        if (!(limit->get_min() <= 0.5 * double(k) / experts && limit->get_max() >= 2.0)) {
+            return false;
+        }
+        std::vector<std::shared_ptr<ov::Node>> slices;
+        for (const auto & input : tk->output(1).get_target_inputs()) {
+            auto slice = ov::as_type_ptr<v8::Slice>(input.get_node()->shared_from_this());
+            if (!slice || input.get_index() != 0 || slice->get_input_size() != 5 ||
+                !constant_is(slice->input_value(1), {0}) || !constant_is(slice->input_value(2), {k}) ||
+                !constant_is(slice->input_value(3), {1}) ||
+                (!constant_is(slice->input_value(4), {3}) && !constant_is(slice->input_value(4), {-1}))) {
+                return false;
+            }
+            slices.push_back(slice);
+        }
+
+        // Remove GGML's leading singleton dimension while building the plugin pattern,
+        // then restore it on both outputs for the following GGML nodes.
+        auto axis0 = v0::Constant::create(ov::element::i64, ov::Shape{1}, {0});
+        auto hidden = std::make_shared<v0::Squeeze>(mm->input_value(0), axis0);
+        auto routing = std::make_shared<v0::MatMul>(hidden, mm->input_value(1), false, mm->get_transpose_b());
+        auto probabilities = std::make_shared<v8::Softmax>(routing, -1);
+        auto selected = std::make_shared<v11::TopK>(probabilities,
+            v0::Constant::create(ov::element::i64, ov::Shape{}, {k}), 2,
+            v11::TopK::Mode::MAX, v11::TopK::SortType::SORT_VALUES, tk->get_index_element_type(), false);
+        auto total = std::make_shared<v1::ReduceSum>(selected->output(0),
+            v0::Constant::create(ov::element::i64, ov::Shape{1}, {-1}), true);
+        auto normalized = std::make_shared<v1::Divide>(selected->output(0), total);
+        auto weights_out = std::make_shared<v0::Unsqueeze>(normalized, axis0);
+        auto ids_out = std::make_shared<v0::Unsqueeze>(selected->output(1), axis0);
+        ov::copy_runtime_info({mm, sm, tk, node(norm)}, {hidden, routing, probabilities, selected, total, normalized, weights_out, ids_out});
+        ov::replace_node(node(norm), weights_out);
+        for (const auto & slice : slices) {
+            ov::replace_node(slice, ids_out);
+        }
+        return true;
+    };
+    register_matcher(std::make_shared<Matcher>(norm, "ov::frontend::ggml::pass::FuseMoeRouter"), callback);
+}
+
+}  // namespace pass
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/pass/fuse_moe_router.h

@@ -0,0 +1,19 @@
+#pragma once
+
+#include <openvino/pass/matcher_pass.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace pass {
+
+class FuseMoeRouter : public ov::pass::MatcherPass {
+public:
+    OPENVINO_MATCHER_PASS_RTTI("ov::frontend::ggml::pass::FuseMoeRouter")
+    FuseMoeRouter();
+};
+
+}  // namespace pass
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/pass/fuse_to_conv.cpp

@@ -0,0 +1,283 @@
+#include "fuse_to_conv.h"
+
+#include <openvino/core/graph_util.hpp>
+#include <openvino/core/rt_info.hpp>
+#include <openvino/op/add.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/convolution.hpp>
+#include <openvino/op/extractimagepatches.hpp>
+#include <openvino/op/group_conv.hpp>
+#include <openvino/op/matmul.hpp>
+#include <openvino/op/pad.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/transpose.hpp>
+#include <openvino/pass/pattern/op/pattern.hpp>
+#include <openvino/pass/pattern/op/wrap_type.hpp>
+#include <utility>
+
+namespace opp = ov::pass::pattern;
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace pass {
+
+FuseToConv::FuseToConv() {
+    const auto m_in0 = opp::any_input();
+    const auto m_in1 = opp::any_input();
+    const auto m_matmul = opp::wrap_type<ov::op::v0::MatMul>({m_in0, m_in1});
+
+    const auto callback = [=](opp::Matcher & m) {
+        auto matmul_node = ov::as_type_ptr<ov::op::v0::MatMul>(m.get_match_root());
+        if (!matmul_node) {
+            return false;
+        }
+
+        auto unwrap = [](ov::Output<Node> n) {
+            while (ov::is_type<ov::op::v0::Convert>(n.get_node_shared_ptr()) ||
+                   ov::is_type<ov::op::v1::Reshape>(n.get_node_shared_ptr())) {
+                n = n.get_node_shared_ptr()->input_value(0);
+            }
+            return n;
+        };
+
+        auto get_im2col = [&](ov::Output<Node> trace)
+            -> std::pair<std::shared_ptr<ov::op::v3::ExtractImagePatches>, std::shared_ptr<ov::op::v1::Pad>> {
+            auto t2 = ov::as_type_ptr<ov::op::v1::Transpose>(unwrap(std::move(trace)).get_node_shared_ptr());
+            auto r1 = t2 ? ov::as_type_ptr<ov::op::v1::Reshape>(t2->get_input_node_shared_ptr(0)) : nullptr;
+            auto t1 = r1 ? ov::as_type_ptr<ov::op::v1::Transpose>(r1->get_input_node_shared_ptr(0)) : nullptr;
+            auto eip =
+                t1 ? ov::as_type_ptr<ov::op::v3::ExtractImagePatches>(t1->get_input_node_shared_ptr(0)) : nullptr;
+            auto pad = eip ? ov::as_type_ptr<ov::op::v1::Pad>(eip->get_input_node_shared_ptr(0)) : nullptr;
+            return {eip, pad};
+        };
+
+        bool weight_is_in0 = true;
+        auto [eip, pad] = get_im2col(matmul_node->input_value(1));
+        ov::Output<Node> w_trace = matmul_node->input_value(0);
+        if (!pad) {
+            std::tie(eip, pad) = get_im2col(matmul_node->input_value(0));
+            w_trace = matmul_node->input_value(1);
+            weight_is_in0 = false;
+            if (!pad) {
+                return false;
+            }
+        }
+
+        auto pb_const = ov::as_type_ptr<ov::op::v0::Constant>(pad->get_input_node_shared_ptr(1));
+        auto pe_const = ov::as_type_ptr<ov::op::v0::Constant>(pad->get_input_node_shared_ptr(2));
+        if (!pb_const || !pe_const) {
+            return false;
+        }
+
+        const auto pb = pb_const->cast_vector<std::ptrdiff_t>();
+        const auto pe = pe_const->cast_vector<std::ptrdiff_t>();
+        if (pb.size() < 4 || pe.size() < 4) {
+            return false;
+        }
+
+        auto image_input = pad->input_value(0);
+        auto image_shape = image_input.get_partial_shape();
+        if (image_shape.rank() != 4 || image_shape[1].is_dynamic()) {
+            return false;
+        }
+        const size_t IC = static_cast<size_t>(image_shape[1].get_length());
+
+        w_trace = unwrap(w_trace);
+        auto weight_pshape = w_trace.get_partial_shape();
+        if (!weight_pshape.is_static()) {
+            return false;
+        }
+
+        const auto & ws = weight_pshape.to_shape();
+        const size_t KH = eip->get_sizes()[0];
+        const size_t KW = eip->get_sizes()[1];
+        const size_t kernel_spatial_ic = IC * KH * KW;
+
+        size_t groups = 0;
+        if (IC == 1 && image_shape[0].is_static() && image_shape[2].is_static() && image_shape[3].is_static()) {
+            if ((ws.size() == 4 && ws[1] == 1 && ws[2] == KH && ws[3] == KW && ws[0] > 1) ||
+                (ws.size() == 2 && ws[1] == KH * KW && ws[0] > 1) ||
+                (ws.size() == 3 && ws[1] == 1 && ws[2] == KH * KW && ws[0] > 1)) {
+                groups = ws[0];
+            } else if (ws.size() == 4 && ws[0] == 1 && ws[2] == 1 && ws[3] == KH * KW && ws[1] > 1) {
+                groups = ws[1];
+            }
+        }
+
+        const bool is_depthwise = groups > 1 && (image_shape[0].get_length() % groups == 0);
+
+        ov::Output<Node> conv_out;
+        size_t OC = 0;
+
+        if (is_depthwise) {
+            const size_t N = static_cast<size_t>(image_shape[0].get_length()) / groups;
+            const size_t IH = static_cast<size_t>(image_shape[2].get_length());
+            const size_t IW = static_cast<size_t>(image_shape[3].get_length());
+            OC = groups;
+
+            auto img_shape_const = register_new_node<ov::op::v0::Constant>(
+                ov::element::i64, ov::Shape{4},
+                std::vector<int64_t>{static_cast<int64_t>(N), static_cast<int64_t>(groups), static_cast<int64_t>(IH),
+                                     static_cast<int64_t>(IW)});
+            ov::Output<Node> image_reshaped =
+                register_new_node<ov::op::v1::Reshape>(image_input, img_shape_const, false);
+
+            const ov::Shape conv_w_shape = {groups, 1, 1, KH, KW};
+            ov::Output<Node> weight_input;
+            if (auto weight_const = ov::as_type_ptr<ov::op::v0::Constant>(w_trace.get_node_shared_ptr())) {
+                weight_input = register_new_node<ov::op::v0::Constant>(weight_const->get_element_type(), conv_w_shape,
+                                                                       weight_const->get_data_ptr());
+            } else {
+                auto shape_const = register_new_node<ov::op::v0::Constant>(
+                    ov::element::i64, ov::Shape{5},
+                    std::vector<int64_t>{static_cast<int64_t>(groups), 1, 1, static_cast<int64_t>(KH),
+                                         static_cast<int64_t>(KW)});
+                weight_input = register_new_node<ov::op::v1::Reshape>(w_trace, shape_const, false);
+            }
+
+            if (weight_input.get_element_type() != image_reshaped.get_element_type()) {
+                weight_input = register_new_node<ov::op::v0::Convert>(weight_input, image_reshaped.get_element_type());
+            }
+
+            conv_out = register_new_node<ov::op::v1::GroupConvolution>(
+                image_reshaped, weight_input, eip->get_strides(), ov::CoordinateDiff{pb[2], pb[3]},
+                ov::CoordinateDiff{pe[2], pe[3]}, ov::Strides{eip->get_rates()[0], eip->get_rates()[1]},
+                ov::op::PadType::EXPLICIT);
+        } else {
+            if ((ws.size() == 4 && ws[1] == IC && ws[2] == KH && ws[3] == KW) ||
+                (ws.size() == 3 && ws[1] == IC && ws[2] == KW) || (ws.size() == 2 && ws[1] == kernel_spatial_ic)) {
+                OC = ws[0];
+            } else if ((ws.size() == 4 && ws[0] == 1 && ws[2] == IC && ws[3] == KW) ||
+                       (ws.size() == 2 && ws[0] == kernel_spatial_ic)) {
+                OC = ws[1];
+            } else if (ws.size() == 3 && ws[0] == 1 && ws[1] == IC && ws[2] == KW) {
+                OC = 1;
+            } else if (ws.size() == 4 && ws[3] == kernel_spatial_ic) {
+                OC = ws[2];
+            } else if (ws.size() == 4 && ws[2] == kernel_spatial_ic) {
+                OC = ws[3];
+            } else if (kernel_spatial_ic && ov::shape_size(ws) % kernel_spatial_ic == 0) {
+                OC = ov::shape_size(ws) / kernel_spatial_ic;
+            } else {
+                return false;
+            }
+
+            const ov::Shape conv_w_shape = {OC, IC, KH, KW};
+
+            ov::Output<Node> weight_input;
+            if (auto weight_const = ov::as_type_ptr<ov::op::v0::Constant>(w_trace.get_node_shared_ptr())) {
+                weight_input = register_new_node<ov::op::v0::Constant>(weight_const->get_element_type(), conv_w_shape,
+                                                                       weight_const->get_data_ptr());
+            } else {
+                auto shape_const = register_new_node<ov::op::v0::Constant>(
+                    ov::element::i64, ov::Shape{4},
+                    std::vector<int64_t>{static_cast<int64_t>(OC), static_cast<int64_t>(IC), static_cast<int64_t>(KH),
+                                         static_cast<int64_t>(KW)});
+                weight_input = register_new_node<ov::op::v1::Reshape>(w_trace, shape_const, false);
+            }
+
+            if (weight_input.get_element_type() != image_input.get_element_type()) {
+                weight_input = register_new_node<ov::op::v0::Convert>(weight_input, image_input.get_element_type());
+            }
+
+            conv_out = register_new_node<ov::op::v1::Convolution>(
+                image_input, weight_input, eip->get_strides(), ov::CoordinateDiff{pb[2], pb[3]},
+                ov::CoordinateDiff{pe[2], pe[3]}, ov::Strides{eip->get_rates()[0], eip->get_rates()[1]},
+                ov::op::PadType::EXPLICIT);
+        }
+
+        constexpr auto target_type = ov::element::f32;
+        if (conv_out.get_element_type() != target_type) {
+            conv_out = register_new_node<ov::op::v0::Convert>(conv_out, target_type);
+        }
+
+        std::shared_ptr<ov::op::v1::Add> add_node;
+        ov::Output<Node> bias_input;
+
+        auto try_fuse_bias = [&](const std::shared_ptr<Node> & n) {
+            auto add = ov::as_type_ptr<ov::op::v1::Add>(n);
+            if (!add) {
+                return false;
+            }
+            for (size_t i = 0; i < 2; ++i) {
+                if (ov::is_type<ov::op::v0::Constant>(add->get_input_node_shared_ptr(i))) {
+                    bias_input = add->input_value(i);
+                    add_node = add;
+                    return true;
+                }
+            }
+            return false;
+        };
+
+        for (const auto & consumer : matmul_node->output(0).get_target_inputs()) {
+            auto n = consumer.get_node()->shared_from_this();
+            if (try_fuse_bias(n)) {
+                break;
+            }
+            if (ov::is_type<ov::op::v0::Convert>(n) || ov::is_type<ov::op::v1::Reshape>(n)) {
+                for (const auto & next : n->output(0).get_target_inputs()) {
+                    if (try_fuse_bias(next.get_node()->shared_from_this())) {
+                        break;
+                    }
+                }
+            }
+            if (add_node) {
+                break;
+            }
+        }
+
+        ov::Output<Node> final_out = conv_out;
+        std::shared_ptr<Node> target_node = matmul_node;
+
+        if (add_node) {
+            ov::Output<Node> bias = bias_input;
+            if (bias.get_element_type() != target_type) {
+                bias = register_new_node<ov::op::v0::Convert>(bias, target_type);
+            }
+            auto bias_shape = register_new_node<ov::op::v0::Constant>(
+                ov::element::i64, ov::Shape{4}, std::vector<int64_t>{1, static_cast<int64_t>(OC), 1, 1});
+            bias = register_new_node<ov::op::v1::Reshape>(bias, bias_shape, false);
+            final_out = register_new_node<ov::op::v1::Add>(conv_out, bias);
+            target_node = add_node;
+        }
+
+        if (!is_depthwise) {
+            auto perm = register_new_node<ov::op::v0::Constant>(
+                ov::element::i64, ov::Shape{4},
+                weight_is_in0 ? std::vector<int64_t>{1, 0, 2, 3} : std::vector<int64_t>{0, 2, 3, 1});
+            final_out = register_new_node<ov::op::v1::Transpose>(final_out, perm);
+        }
+
+        auto orig_shape = target_node->get_output_partial_shape(0);
+        if (orig_shape.is_static() && final_out.get_partial_shape().is_static()) {
+            if (ov::shape_size(orig_shape.to_shape()) != ov::shape_size(final_out.get_shape())) {
+                return false;
+            }
+        }
+        if (orig_shape.is_static() && final_out.get_partial_shape() != orig_shape) {
+            auto shape_const = register_new_node<ov::op::v0::Constant>(ov::element::i64, ov::Shape{orig_shape.size()},
+                                                                       orig_shape.to_shape());
+            final_out = register_new_node<ov::op::v1::Reshape>(final_out, shape_const, false);
+        }
+
+        auto orig_type = target_node->get_output_element_type(0);
+        if (final_out.get_element_type() != orig_type) {
+            final_out = register_new_node<ov::op::v0::Convert>(final_out, orig_type);
+        }
+
+        final_out.get_node_shared_ptr()->set_friendly_name(target_node->get_friendly_name());
+        ov::copy_runtime_info(m.get_matched_nodes(), final_out.get_node_shared_ptr());
+        ov::replace_node(target_node, final_out.get_node_shared_ptr());
+
+        return true;
+    };
+
+    register_matcher(std::make_shared<opp::Matcher>(m_matmul, "ov::frontend::ggml::pass::FuseToConv"), callback);
+}
+
+}  // namespace pass
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/pass/fuse_to_conv.h

@@ -0,0 +1,17 @@
+#include "openvino/pass/matcher_pass.hpp"
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace pass {
+
+class FuseToConv : public ov::pass::MatcherPass {
+public:
+    OPENVINO_MATCHER_PASS_RTTI("ov::frontend::ggml::pass::FuseToConv")
+    FuseToConv();
+};
+
+}  // namespace pass
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.cpp

@@ -0,0 +1,60 @@
+#include "fuse_to_sdpa.h"
+
+#include <openvino/core/graph_util.hpp>
+#include <openvino/core/rt_info.hpp>
+#include <openvino/op/add.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/matmul.hpp>
+#include <openvino/op/multiply.hpp>
+#include <openvino/op/scaled_dot_product_attention.hpp>
+#include <openvino/op/softmax.hpp>
+#include <openvino/op/transpose.hpp>
+#include <openvino/pass/pattern/op/label.hpp>
+#include <openvino/pass/pattern/op/pattern.hpp>
+#include <openvino/pass/pattern/op/wrap_type.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace pass {
+
+FuseToSDPA::FuseToSDPA() {
+    // Not maintained since FLASH_ATTN_EXT has replaced this pattern
+    const auto m_k = ov::pass::pattern::any_input();
+    const auto m_q = ov::pass::pattern::any_input();
+    const auto m_qk = ov::pass::pattern::wrap_type<ov::op::v0::MatMul>({m_q, m_k});
+    const auto m_qk_f32 = ov::pass::pattern::wrap_type<ov::op::v0::Convert>({m_qk});
+    const auto m_scale = ov::pass::pattern::any_input();
+    const auto m_scaled_qk = ov::pass::pattern::wrap_type<ov::op::v1::Multiply>({m_qk_f32, m_scale});
+    const auto m_mask = ov::pass::pattern::any_input();
+    const auto m_masked_qk = ov::pass::pattern::wrap_type<ov::op::v1::Add>({m_scaled_qk, m_mask});
+    const auto m_softmax_qk = ov::pass::pattern::wrap_type<ov::op::v8::Softmax>({m_masked_qk});
+    const auto m_softmax_qk_f16 = ov::pass::pattern::wrap_type<ov::op::v0::Convert>({m_softmax_qk});
+    const auto m_v = ov::pass::pattern::any_input();
+    const auto m_qkv = ov::pass::pattern::wrap_type<ov::op::v0::MatMul>({m_softmax_qk_f16, m_v});
+
+    const auto callback = [=](ov::pass::pattern::Matcher & m) {
+        auto & pattern_to_output = m.get_pattern_value_map();
+        auto k = pattern_to_output[m_k];
+        auto q = pattern_to_output[m_q];
+        auto v = pattern_to_output[m_v];
+        auto mask = pattern_to_output[m_mask];
+        auto scale = pattern_to_output[m_scale];
+
+        auto mask_f16 = register_new_node<ov::op::v0::Convert>(mask, ov::element::f16);
+        auto scale_f16 = register_new_node<ov::op::v0::Convert>(scale, ov::element::f16);
+        auto sdpa = std::make_shared<ov::op::v13::ScaledDotProductAttention>(q, k, v, mask_f16, scale_f16, false);
+
+        ov::replace_node(m.get_match_root(), sdpa);
+        ov::copy_runtime_info(m.get_matched_nodes(), sdpa);
+
+        return true;
+    };
+    register_matcher(std::make_shared<ov::pass::pattern::Matcher>(m_qkv, "ov::frontend::ggml::pass::FuseToSDPA"),
+                     callback);
+}
+
+}  // namespace pass
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.h

@@ -0,0 +1,17 @@
+#include "openvino/pass/matcher_pass.hpp"
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace pass {
+
+class FuseToSDPA : public ov::pass::MatcherPass {
+public:
+    OPENVINO_MATCHER_PASS_RTTI("ov::frontend::ggml::pass::FuseToSDPA")
+    FuseToSDPA();
+};
+
+}  // namespace pass
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/pass/kv_state_seq_axis.cpp

@@ -0,0 +1,114 @@
+#include "kv_state_seq_axis.h"
+
+#include <memory>
+#include <openvino/core/graph_util.hpp>
+#include <openvino/op/assign.hpp>
+#include <openvino/op/concat.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/read_value.hpp>
+#include <openvino/op/transpose.hpp>
+#include <vector>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace pass {
+
+namespace {
+
+const std::vector<int64_t> & seq_axis_perm() {
+    // [1, seq, n_heads_kv, head_size] <-> [1, n_heads_kv, seq, head_size]
+    static const std::vector<int64_t> perm{0, 2, 1, 3};
+    return perm;
+}
+
+// True when the state still has the frontend's stateful KV layout, so the sequence axis
+// can be moved: rank 4, batch and both head dims static, and seq the only dynamic dim,
+// at dim 1. Any KV head count is fine. With a single head the rewrite is pure metadata
+// ([1, seq, 1, head] and [1, 1, seq, head] are the same memory); with several heads it
+// also drops the reader-side transpose of the whole accumulated state, which is where
+// most of the gain comes from at depth.
+bool can_move_seq_axis(const ov::PartialShape & shape) {
+    return shape.rank().is_static() && shape.rank().get_length() == 4 && shape[0].is_static() &&
+           shape[1].is_dynamic() && shape[2].is_static() && shape[3].is_static();
+}
+
+std::shared_ptr<ov::op::v0::Concat> match_kv_append(const std::shared_ptr<ov::op::v6::Assign> & assign) {
+    auto concat = ov::as_type_ptr<ov::op::v0::Concat>(assign->get_input_node_shared_ptr(0));
+    if (!concat || concat->get_input_size() != 2 || concat->get_axis() != 1) {
+        return nullptr;
+    }
+    auto read_value = ov::as_type_ptr<ov::op::v6::ReadValue>(concat->get_input_node_shared_ptr(0));
+    if (!read_value || read_value->get_variable() != assign->get_variable()) {
+        return nullptr;
+    }
+    if (!can_move_seq_axis(read_value->get_output_partial_shape(0))) {
+        return nullptr;
+    }
+    return concat;
+}
+
+}  // namespace
+
+bool KVStateSeqAxis::run_on_model(const std::shared_ptr<ov::Model> & model) {
+    std::vector<std::shared_ptr<ov::op::v6::Assign>> assigns;
+    for (const auto & op : model->get_ops()) {
+        if (auto assign = ov::as_type_ptr<ov::op::v6::Assign>(op)) {
+            assigns.push_back(assign);
+        }
+    }
+
+    bool changed = false;
+    for (const auto & assign : assigns) {
+        auto concat = match_kv_append(assign);
+        if (!concat) {
+            continue;
+        }
+        auto read_value = ov::as_type_ptr<ov::op::v6::ReadValue>(concat->get_input_node_shared_ptr(0));
+
+        auto variable = read_value->get_variable();
+        auto info = variable->get_info();
+        const auto & shape = info.data_shape;
+        info.data_shape = ov::PartialShape{shape[0], shape[2], shape[1], shape[3]};
+        variable->update(info);
+        read_value->validate_and_infer_types();
+
+        auto readers = concat->output(0).get_target_inputs();
+
+        auto new_rows = concat->input_value(1);
+        auto perm_in = ov::op::v0::Constant::create(ov::element::i64, {4}, seq_axis_perm());
+        concat->set_argument(1, std::make_shared<ov::op::v1::Transpose>(new_rows, perm_in));
+        concat->set_axis(2);
+        concat->validate_and_infer_types();
+
+        // Readers still expect seq at dim 1. A reader that is itself the inverse
+        // Transpose wanted seq at dim 2 all along, so drop it; give anything else the
+        // inverse Transpose so its input is unchanged.
+        for (const auto & reader : readers) {
+            auto * node = reader.get_node();
+            if (ov::is_type<ov::op::v6::Assign>(node)) {
+                continue;
+            }
+            bool dropped = false;
+            if (auto * transpose = ov::as_type<ov::op::v1::Transpose>(node)) {
+                auto order = ov::as_type_ptr<ov::op::v0::Constant>(transpose->get_input_node_shared_ptr(1));
+                if (order && order->cast_vector<int64_t>() == seq_axis_perm()) {
+                    ov::replace_output_update_name(transpose->output(0), concat->output(0));
+                    dropped = true;
+                }
+            }
+            if (!dropped) {
+                auto perm_out = ov::op::v0::Constant::create(ov::element::i64, {4}, seq_axis_perm());
+                reader.replace_source_output(std::make_shared<ov::op::v1::Transpose>(concat->output(0), perm_out));
+            }
+        }
+        changed = true;
+    }
+
+    return changed;
+}
+
+}  // namespace pass
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/pass/kv_state_seq_axis.h

@@ -0,0 +1,24 @@
+#include "openvino/pass/pass.hpp"
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace pass {
+
+// Moves the sequence axis of the stateful KV cache from dim 1 to dim 2, i.e. from
+// [1, seq, n_heads_kv, head_size] to [1, n_heads_kv, seq, head_size], and updates the
+// Concat that appends to it. Two wins: the GPU plugin only appends new tokens in place
+// when the growing axis is a spatial axis, and the reader no longer has to transpose the
+// whole accumulated state every token (that cost grows with context length, so it is the
+// larger win at depth for a model with several KV heads). Only rewrites states that still
+// match the frontend layout, so it no-ops if that layout ever changes.
+class KVStateSeqAxis : public ov::pass::ModelPass {
+public:
+    OPENVINO_MODEL_PASS_RTTI("ov::frontend::ggml::pass::KVStateSeqAxis")
+    bool run_on_model(const std::shared_ptr<ov::Model> & model) override;
+};
+
+}  // namespace pass
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/pass/mark_decompression_convert_constant_folding.h

@@ -0,0 +1,29 @@
+#pragma once
+
+#include "mark_decompression_convert_constant_folding.h"
+#include "openvino/core/visibility.hpp"
+#include "openvino/pass/matcher_pass.hpp"
+
+#ifdef OPENVINO_STATIC_LIBRARY
+#    define TRANSFORMATIONS_API
+#else
+#    ifdef IMPLEMENT_OPENVINO_API
+#        define TRANSFORMATIONS_API OPENVINO_CORE_EXPORTS
+#    else
+#        define TRANSFORMATIONS_API OPENVINO_CORE_IMPORTS
+#    endif  // IMPLEMENT_OPENVINO_API
+#endif      // OPENVINO_STATIC_LIBRARY
+
+namespace ov {
+namespace pass {
+
+class TRANSFORMATIONS_API MarkCompressedFloatConstants;
+
+}  // namespace pass
+}  // namespace ov
+
+class ov::pass::MarkCompressedFloatConstants : public MatcherPass {
+public:
+    OPENVINO_MATCHER_PASS_RTTI("MarkCompressedFloatConstants")
+    MarkCompressedFloatConstants();
+};

llama.cpp/ggml/src/ggml-openvino/openvino/pass/mark_dequantization_subgraph.h

@@ -0,0 +1,44 @@
+// Copyright (C) 2018-2026 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+// Local mirror of OpenVINO's ov::pass::MarkDequantization pass declaration.
+//
+// The pass body is provided by the linked libopenvino.so; only the declaration is needed here so
+// we can register it directly in our own TranslateSession::apply_transformations (same approach as
+// MarkCompressedFloatConstants's local mirror in mark_decompression_convert_constant_folding.h). This
+// lets us mark our GatherMatmul dequantization chain with disable_constant_folding regardless of the
+// CPU/GPU plugin's own is_decompression_multiply() consumer allowlist.
+// The class layout must stay in sync with
+//   openvino/src/common/transformations/include/transformations/low_precision/mark_dequantization_subgraph.hpp
+
+#pragma once
+
+#include "openvino/core/type/element_type.hpp"
+#include "openvino/core/visibility.hpp"
+#include "openvino/pass/matcher_pass.hpp"
+
+#ifdef OPENVINO_STATIC_LIBRARY
+#    define TRANSFORMATIONS_API
+#else
+#    ifdef IMPLEMENT_OPENVINO_API
+#        define TRANSFORMATIONS_API OPENVINO_CORE_EXPORTS
+#    else
+#        define TRANSFORMATIONS_API OPENVINO_CORE_IMPORTS
+#    endif  // IMPLEMENT_OPENVINO_API
+#endif      // OPENVINO_STATIC_LIBRARY
+
+namespace ov {
+namespace pass {
+
+class TRANSFORMATIONS_API MarkDequantization;
+
+}  // namespace pass
+}  // namespace ov
+
+class ov::pass::MarkDequantization : public MatcherPass {
+public:
+    OPENVINO_MATCHER_PASS_RTTI("MarkDequantization")
+    explicit MarkDequantization(const element::TypeVector & precisions,
+                                bool fold_subtract_const = false,
+                                bool fold_multiply_const = true);
+};

llama.cpp/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.cpp

@@ -0,0 +1,59 @@
+#include "squeeze_matmul.h"
+
+#include <openvino/core/graph_util.hpp>
+#include <openvino/core/rt_info.hpp>
+#include <openvino/core/type.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/matmul.hpp>
+#include <openvino/op/squeeze.hpp>
+#include <openvino/op/unsqueeze.hpp>
+#include <openvino/pass/pattern/op/label.hpp>
+#include <openvino/pass/pattern/op/pattern.hpp>
+#include <openvino/pass/pattern/op/wrap_type.hpp>
+
+namespace opp = ov::pass::pattern;
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace pass {
+
+// For quantized models, NPUW expects the activation to be 3d in DQ(DynamicQuantization) opt, e.g. DQMatMulGQ2i
+SqueezeMatmul::SqueezeMatmul() {
+    auto m_act = opp::any_input();
+    auto m_wei = opp::any_input();
+    auto m_matmul = opp::wrap_type<ov::op::v0::MatMul>({m_act, m_wei});
+
+    const auto callback = [=](ov::pass::pattern::Matcher & m) {
+        const auto & pattern_map = m.get_pattern_value_map();
+        auto matmul_node =
+            ov::as_type_ptr<ov::op::v0::MatMul>(pattern_map.at(m_matmul).get_node_shared_ptr());
+        auto act = pattern_map.at(m_act);
+        auto wei = pattern_map.at(m_wei);
+        auto act_shape = act.get_partial_shape();
+        auto wei_shape = wei.get_partial_shape();
+        if (act_shape.rank().is_dynamic() || wei_shape.rank().is_dynamic()) {
+            return false;
+        }
+        if (act_shape.rank().get_length() == 4 && wei_shape.rank().get_length() == 2) {
+            auto axis = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{1}, {0});
+            auto squeezed_act = std::make_shared<ov::op::v0::Squeeze>(act, axis);
+            auto new_matmul = std::make_shared<ov::op::v0::MatMul>(squeezed_act, wei, matmul_node->get_transpose_a(),
+                                                                   matmul_node->get_transpose_b());
+            auto unsqueezed_output = std::make_shared<ov::op::v0::Unsqueeze>(new_matmul, axis);
+            unsqueezed_output->set_friendly_name(matmul_node->get_friendly_name());
+            ov::copy_runtime_info(matmul_node, {squeezed_act, new_matmul, unsqueezed_output});
+            ov::replace_node(matmul_node, unsqueezed_output);
+            return true;
+        }
+        return false;
+    };
+
+    register_matcher(std::make_shared<ov::pass::pattern::Matcher>(m_matmul, "ov::frontend::ggml::pass::SqueezeMatmul"),
+                     callback);
+}
+
+}  // namespace pass
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.h

@@ -0,0 +1,17 @@
+#include "openvino/pass/matcher_pass.hpp"
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace pass {
+
+class SqueezeMatmul : public ov::pass::MatcherPass {
+public:
+    OPENVINO_MATCHER_PASS_RTTI("ov::frontend::ggml::pass::SqueezeMatmul")
+    SqueezeMatmul();
+};
+
+}  // namespace pass
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/rt_info/weightless_caching_attributes.hpp

@@ -0,0 +1,41 @@
+// Copyright (C) 2018-2026 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include <openvino/core/core_visibility.hpp>
+#include <openvino/core/node.hpp>
+#include <openvino/core/runtime_attribute.hpp>
+
+namespace ov {
+
+/**
+ * @brief Holds weightless caching attributes of a single constant.
+ *
+ * WeightlessCacheAttribute class represents runtime info attribute that holds
+ * the values of original size of the constant in bytes and the binary offset of the
+ * constant's data in the weights file used by the weightless caching mechanism. It's
+ * not copyable in case the data was changed (the original node was replaced by a new
+ * one produced during the tranformation pipeline) - in that case weightless caching
+ * can't be used for that constant.
+ */
+class OPENVINO_API WeightlessCacheAttribute : public RuntimeAttribute {
+public:
+    OPENVINO_RTTI("WeightlessCacheAttribute", "0", RuntimeAttribute)
+
+    WeightlessCacheAttribute() = delete;
+
+    WeightlessCacheAttribute(size_t original_size, size_t bin_offset, ov::element::Type original_dtype)
+        : original_size(original_size),
+          bin_offset(bin_offset),
+          original_dtype(original_dtype) {}
+
+    bool is_copyable() const override;
+
+    size_t original_size;
+    size_t bin_offset;
+    ov::element::Type original_dtype;
+};
+
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/translate_session.cpp

@@ -0,0 +1,546 @@
+#include "translate_session.h"
+
+#include "ggml-impl.h"
+#include "ggml-openvino/ggml-openvino-extra.h"
+#include "ggml-openvino/openvino/node_context.h"
+#include "ggml-openvino/openvino/utils.h"
+#include "input_model.h"
+#include "pass/fuse_argsort_topk.h"
+#include "pass/fuse_moe_router.h"
+#include "pass/align_eltwise_ranks.h"
+#include "pass/fuse_moe_compressed.h"
+#include "pass/fuse_to_conv.h"
+#include "pass/kv_state_seq_axis.h"
+#include "pass/mark_decompression_convert_constant_folding.h"
+#include "pass/mark_dequantization_subgraph.h"
+#include "pass/squeeze_matmul.h"
+#include "rt_info/weightless_caching_attributes.hpp"
+
+#include <algorithm>
+#include <cstdint>
+#include <cstdlib>
+#include <map>
+#include <memory>
+#include <openvino/core/node.hpp>
+#include <openvino/core/preprocess/pre_post_process.hpp>
+#include <openvino/core/shape.hpp>
+#include <openvino/core/type.hpp>
+#include <openvino/core/type/element_type.hpp>
+#include <openvino/op/add.hpp>
+#include <openvino/op/broadcast.hpp>
+#include <openvino/op/concat.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/convert_like.hpp>
+#include <openvino/op/cos.hpp>
+#include <openvino/op/divide.hpp>
+#include <openvino/op/gather.hpp>
+#include <openvino/op/greater_eq.hpp>
+#include <openvino/op/less.hpp>
+#include <openvino/op/logical_and.hpp>
+#include <openvino/op/multiply.hpp>
+#include <openvino/op/parameter.hpp>
+#include <openvino/op/range.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/result.hpp>
+#include <openvino/op/select.hpp>
+#include <openvino/op/sin.hpp>
+#include <openvino/op/slice.hpp>
+#include <openvino/op/squeeze.hpp>
+#include <openvino/op/strided_slice.hpp>
+#include <openvino/op/subtract.hpp>
+#include <openvino/op/transpose.hpp>
+#include <openvino/op/unsqueeze.hpp>
+#include <openvino/pass/constant_folding.hpp>
+#include <openvino/pass/make_stateful.hpp>
+#include <limits>
+#include <sstream>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+
+using namespace ov::op;
+
+namespace {
+
+std::shared_ptr<ov::op::v0::Parameter> create_parameter(const std::string & name,
+                                                        const ModelInputInfo & input_info) {
+    auto param_node = std::make_shared<ov::op::v0::Parameter>(input_info.type, input_info.shape);
+    param_node->set_friendly_name(name);
+    param_node->output(0).get_tensor().set_names({name});
+    return param_node;
+}
+
+std::shared_ptr<ov::Node> create_extra_input(const std::string & name, const ModelExtraInputInfo & input_info) {
+    if (input_info.is_parameter) {
+        auto param_node = std::make_shared<ov::op::v0::Parameter>(input_info.type, input_info.shape);
+        param_node->set_friendly_name(name);
+        param_node->output(0).get_tensor().set_names({name});
+        return param_node;
+    }
+
+    auto constant = std::make_shared<ov::op::v0::Constant>(input_info.type, input_info.shape,
+                                                          std::vector<int64_t>{input_info.value});
+    constant->set_friendly_name(name);
+    return constant;
+}
+
+ov::pass::MakeStateful::ParamResPairs get_kv_param_res_pairs(
+    const std::shared_ptr<ov::Model> & model,
+    const std::map<std::string, std::string> & kv_param_res_names) {
+    ov::pass::MakeStateful::ParamResPairs pairs;
+    const auto & params = model->get_parameters();
+    const auto & results = model->get_results();
+
+    for (const auto & param_res : kv_param_res_names) {
+        const auto & param_name = param_res.first;
+        const auto & res_name = param_res.second;
+
+        auto param_it = std::find_if(params.begin(), params.end(), [&](const std::shared_ptr<v0::Parameter> & node) {
+            return node->get_friendly_name() == param_name;
+        });
+
+        OPENVINO_ASSERT(param_it != params.end(), "The tensor name ", param_name,
+                        " is not associated with any of "
+                        "Parameters in the network.");
+
+        auto res_it = std::find_if(results.begin(), results.end(), [&](const std::shared_ptr<v0::Result> & node) {
+            return node->get_friendly_name() == res_name;
+        });
+
+        OPENVINO_ASSERT(res_it != results.end(), "The tensor name ", res_name,
+                        " is not associated with any of "
+                        "Results in the network.");
+
+        std::shared_ptr<ov::op::v0::Parameter> param = *param_it;
+        std::shared_ptr<ov::op::v0::Result> res = *res_it;
+        pairs.emplace_back(param, res);
+    }
+    return pairs;
+}
+
+void add_sliced_mask_stateful(TensorMap & tensor_map, bool imrope) {
+    auto create_sliced_mask = [&](const std::string & mask_name, const std::string & sliced_name) {
+        if ((tensor_map.find(mask_name) != tensor_map.end()) &&
+            (tensor_map.find("token_len_per_seq") != tensor_map.end()) &&
+            (tensor_map.find("inp_pos") != tensor_map.end())) {
+            auto token_len_per_seq = tensor_map.at("token_len_per_seq").get_node_shared_ptr();
+            auto mask = tensor_map.at(mask_name).get_node_shared_ptr();
+            std::shared_ptr<ov::Node> mask_sliced = mask;
+            auto one = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+            auto zero = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
+            auto three = ov::op::v0::Constant::create(ov::element::i64, {1}, {3});
+            auto neg_one = ov::op::v0::Constant::create(ov::element::i64, {1}, {-1});
+
+            auto step = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+            auto axes = ov::op::v0::Constant::create(ov::element::i64, {1}, {-1});
+
+            auto inp_pos = tensor_map.at("inp_pos").get_node_shared_ptr();
+            // IMROPE's fourth position plane is zero for text; use the token's first plane.
+            ov::Output<ov::Node> last_index = neg_one;
+            if (imrope) {
+                last_index = std::make_shared<ov::op::v1::Add>(token_len_per_seq, neg_one);
+            }
+            auto last_inp_pos = std::make_shared<ov::op::v8::Gather>(inp_pos, last_index, three);
+            auto last_inp_pos_1d = std::make_shared<ov::op::v1::Reshape>(
+                last_inp_pos, ov::op::v0::Constant::create(ov::element::i64, {1}, {1}), false);
+            auto last_inp_pos_cvt = std::make_shared<ov::op::v0::Convert>(last_inp_pos_1d, ov::element::i64);
+            auto last_inp_pos_inc = std::make_shared<ov::op::v1::Add>(last_inp_pos_cvt, one);
+
+            mask_sliced = std::make_shared<ov::op::v8::Slice>(mask, zero, last_inp_pos_inc, step, axes);
+            mask_sliced = std::make_shared<ov::op::v0::Convert>(mask_sliced, ov::element::f16);
+            mask_sliced->set_friendly_name(sliced_name);
+
+            tensor_map.insert({sliced_name, mask_sliced->output(0)});
+        }
+    };
+
+    create_sliced_mask("self_kq_mask", "KQ_mask_sliced");
+    create_sliced_mask("KQ_mask", "KQ_mask_sliced");
+    create_sliced_mask("self_kq_mask_swa", "KQ_mask_swa_sliced");
+}
+
+// Rebuild the sliding-window mask from absolute positions.
+// ggml caps self_kq_mask_swa at the size of its own SWA cache, but the stateful KV state is
+// Concat-appended and grows without bound, so past that cap the two disagree on length and the
+// mask add fails. A pure-Concat state is ordered by position, so positions can rebuild the mask.
+// swa_window holds the real n_swa, read back from the ggml mask in ggml-decoder.cpp.
+// No-op when the graph has no SWA mask, or when the window could not be read back.
+void add_position_mask_stateful_swa(TensorMap & tensor_map) {
+    if (tensor_map.find("self_kq_mask_swa") == tensor_map.end() || tensor_map.find("inp_pos") == tensor_map.end() ||
+        tensor_map.find("swa_window") == tensor_map.end()) {
+        return;
+    }
+
+    auto inp_pos = tensor_map.at("inp_pos").get_node_shared_ptr();
+
+    auto zero_i64 = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
+    auto one_i64 = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+    auto three = ov::op::v0::Constant::create(ov::element::i64, {1}, {3});
+    auto neg_one = ov::op::v0::Constant::create(ov::element::i64, {1}, {-1});
+
+    auto query_pos = std::make_shared<ov::op::v0::Convert>(inp_pos, ov::element::i64);
+    auto query_pos_1d = std::make_shared<ov::op::v1::Reshape>(
+        query_pos, ov::op::v0::Constant::create(ov::element::i64, {1}, {-1}), false);
+
+    auto last_pos = std::make_shared<ov::op::v8::Gather>(inp_pos, neg_one, three);
+    auto last_pos_1d = std::make_shared<ov::op::v1::Reshape>(last_pos, one_i64, false);
+    auto last_pos_cvt = std::make_shared<ov::op::v0::Convert>(last_pos_1d, ov::element::i64);
+    auto total_len = std::make_shared<ov::op::v1::Add>(last_pos_cvt, one_i64);
+    auto total_len_scalar = std::make_shared<ov::op::v0::Squeeze>(total_len);
+
+    auto cached_pos = std::make_shared<ov::op::v4::Range>(
+        ov::op::v0::Constant::create(ov::element::i64, {}, {0}), total_len_scalar,
+        ov::op::v0::Constant::create(ov::element::i64, {}, {1}), ov::element::i64);
+
+    auto query_col = std::make_shared<ov::op::v1::Reshape>(
+        query_pos_1d, ov::op::v0::Constant::create(ov::element::i64, {2}, {-1, 1}), false);
+    auto cached_row = std::make_shared<ov::op::v1::Reshape>(
+        cached_pos, ov::op::v0::Constant::create(ov::element::i64, {2}, {1, -1}), false);
+    auto diff = std::make_shared<ov::op::v1::Subtract>(query_col, cached_row);
+
+    auto swa_window = tensor_map.at("swa_window").get_node_shared_ptr();
+    auto window = std::make_shared<ov::op::v0::Convert>(swa_window, ov::element::i64);
+    auto causal_ok = std::make_shared<ov::op::v1::GreaterEqual>(diff, zero_i64);
+    auto window_ok = std::make_shared<ov::op::v1::Less>(diff, window);
+    auto keep = std::make_shared<ov::op::v1::LogicalAnd>(causal_ok, window_ok);
+
+    auto zero_f = ov::op::v0::Constant::create(ov::element::f32, {}, {0.0f});
+    auto neg_inf_f = ov::op::v0::Constant::create(ov::element::f32, {}, {-std::numeric_limits<float>::infinity()});
+    std::shared_ptr<ov::Node> mask = std::make_shared<ov::op::v1::Select>(keep, zero_f, neg_inf_f);
+
+    auto batch_axis = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
+    mask = std::make_shared<ov::op::v0::Unsqueeze>(mask, batch_axis);
+    mask = std::make_shared<ov::op::v0::Unsqueeze>(mask, batch_axis);
+    mask = std::make_shared<ov::op::v0::Convert>(mask, ov::element::f16);
+    mask->set_friendly_name("KQ_mask_swa_sliced");
+
+    tensor_map["KQ_mask_swa_sliced"] = mask->output(0);
+}
+
+void add_rope_sin_cos(TensorMap & tensor_map, GgmlDecoder & ggml_model_decoder) {
+    // When ROPE ops in the graph have divergent op_params (e.g. gemma4's mixed
+    // SWA/non-SWA layers with different n_dims or freq_base), a shared sin/cos
+    // precompute cannot broadcast across every ROPE use. Skip it here and let
+    // translate_rope() build sin/cos per-op from its own op_params.
+    if (ggml_model_decoder.has_mixed_rope_params()) {
+        return;
+    }
+    int32_t * rope_params = ggml_model_decoder.get_rope_params();
+    if (tensor_map.find("inp_pos") == tensor_map.end() || rope_params == nullptr) {
+        return;
+    }
+    auto inp_pos = tensor_map.at("inp_pos").get_node_shared_ptr();
+    std::shared_ptr<ov::Node> rope_freqs_weight;
+    if (tensor_map.find("rope_freqs.weight") != tensor_map.end()) {
+        rope_freqs_weight = tensor_map.at("rope_freqs.weight").get_node_shared_ptr();
+    }
+
+    auto sin_cos = make_sin_cos(rope_params, inp_pos, rope_freqs_weight);
+    auto sin_theta = sin_cos.first;
+    auto cos_theta = sin_cos.second;
+
+    cos_theta.get_node_shared_ptr()->set_friendly_name("rope_cos");
+    sin_theta.get_node_shared_ptr()->set_friendly_name("rope_sin");
+    tensor_map.insert({"rope_cos", cos_theta});
+    tensor_map.insert({"rope_sin", sin_theta});
+}
+
+// Create common patterns
+void preprocess(TensorMap & tensor_map, GgmlDecoder & ggml_model_decoder) {
+    if (ggml_model_decoder.is_stateful()) {
+        add_sliced_mask_stateful(tensor_map, ggml_model_decoder.get_rope_params()[2] == GGML_ROPE_TYPE_IMROPE);
+        add_position_mask_stateful_swa(tensor_map);
+    }
+    // This optimization is error-prone
+    // add_rope_sin_cos(tensor_map, ggml_model_decoder);
+}
+
+}  // namespace
+
+TranslateSession::TranslateSession(const frontend::InputModel::Ptr & input_model,
+                                   const std::unordered_map<std::string, CreatorFunction> & translator_map,
+                                   bool naive) :
+    m_input_model(input_model),
+    m_translator_map(translator_map),
+    m_ov_model(nullptr),
+    m_naive(naive) {}
+
+std::shared_ptr<Model> TranslateSession::get_converted_model() {
+    if (m_ov_model) {
+        return m_ov_model;
+    }
+    m_ov_model = translate_graph(m_input_model);
+    return m_ov_model;
+}
+
+std::shared_ptr<Model> TranslateSession::translate_graph(const frontend::InputModel::Ptr & input_model) {
+    ov::ParameterVector params;
+    ov::ResultVector results;
+    auto tensor_map = std::make_shared<TensorMap>();
+    std::shared_ptr<Model> resulting_model;
+
+    const auto & ggml_model = ov::as_type_ptr<InputModel>(input_model);
+    std::shared_ptr<GgmlDecoder> ggml_model_decoder = ggml_model->get_model_decoder();
+
+    for (const auto & it : ggml_model_decoder->get_model_inputs()) {
+        auto param_node = create_parameter(it.first, it.second);
+        params.push_back(param_node);
+        (*tensor_map)[it.first] = param_node;
+    }
+
+    for (const auto & it : ggml_model_decoder->get_model_extra_inputs()) {
+        auto input_node = create_extra_input(it.first, it.second);
+        if (it.second.is_parameter) {
+            params.push_back(ov::as_type_ptr<ov::op::v0::Parameter>(input_node));
+        }
+        (*tensor_map)[it.first] = input_node;
+    }
+
+    for (const auto & it : ggml_model_decoder->get_model_weights()) {
+        (*tensor_map)[it.first] = it.second;
+    }
+
+    auto translate_node = [&](const std::shared_ptr<GgmlDecoder> & decoder, int node_idx) {
+        auto operation_type = decoder->get_op_type(node_idx);
+        if (operation_type == "GGML_OP_NONE") {
+            return ov::OutputVector{};
+        }
+
+        const auto & node_output_names = decoder->get_output_names(node_idx);
+        if (operation_type == "GGML_OP_VIEW" && decoder->get_op_case(node_idx) == 2 && node_output_names.size() == 1) {
+            auto direct_output = tensor_map->find(node_output_names[0]);
+            if (direct_output != tensor_map->end()) {
+                // GDN publishes its native attention/state outputs under the two GGML VIEW names.
+                // Keep those mappings instead of rebuilding slices of a packed temporary.
+                return ov::OutputVector{direct_output->second};
+            }
+        }
+
+        auto it = m_translator_map.find(operation_type);
+        FRONT_END_OP_CONVERSION_CHECK(it != m_translator_map.end(), "Translation for operation type ", operation_type,
+                                      " is not implemented.");
+        NodeContext node_context(decoder, tensor_map, node_idx, this);
+        ov::OutputVector converted_outputs = it->second(node_context);
+
+        FRONT_END_OP_CONVERSION_CHECK(node_output_names.size() == converted_outputs.size(), "Number of ",
+                                      operation_type, " outputs greater than number of converted outputs, which are ",
+                                      node_output_names.size(), " and ", converted_outputs.size(), " respectively.");
+
+        for (size_t i = 0; i < node_output_names.size(); ++i) {
+            auto output_name = node_output_names[i];
+            if (i < converted_outputs.size() && converted_outputs[i].get_node_shared_ptr() != nullptr) {
+                (*tensor_map)[output_name] = converted_outputs[i];
+            }
+        }
+        return converted_outputs;
+    };
+
+    // To handle cases like this
+    // 3: [ 18432,     1,     1,     1] RESHAPE              cache_r_l0 (reshaped)#3
+    //     [ 18432,     1,     1,     1]            0: NONE        cache_r_l0
+    // 4: [     0,     1,     1,     1] VIEW                 cache_r_l0 (reshaped) (view)#4
+    //     [ 18432,     1,     1,     1]            0: RESHAPE     cache_r_l0 (reshaped)#3
+    // 5: [     0,     1,     1,     1] SCALE                cache_r_l0 (reshaped) (view) (view)#5
+    //     [     0,     1,     1,     1]            0: VIEW        cache_r_l0 (reshaped) (view)#4
+    // 6: [     1,     1,     1,     1] VIEW                  (view)#6
+    //     [     1,     1,     1,     1]            0: NONE        leaf_5
+    // 7: [ 18432,     1,     1,     1] GET_ROWS             conv_states-0#7
+    //     [ 18432,     1,     1,     1]            0: RESHAPE     cache_r_l0 (reshaped)#3
+    //     [     1,     1,     1,     1]            1: VIEW         (view)#6
+    // The scale is in-place which modifies cache_r_l0 (reshaped)#3
+    // The translation of scale overwrites cache_r in the tensor_map,
+    // but we also need to overwrite the old cache_r_l0 (reshaped)#3
+    auto refresh_inplace_aliases = [&](const std::shared_ptr<GgmlDecoder> & decoder, int inplace_node_idx,
+                                       const std::string & view_src_name) {
+        for (int node_idx = 0; node_idx < inplace_node_idx; node_idx++) {
+            if (decoder->is_view_like_alias_of(node_idx, view_src_name)) {
+                translate_node(decoder, node_idx);
+            }
+        }
+    };
+
+    auto node_visitor = [&](const std::shared_ptr<GgmlDecoder> & decoder, int node_idx) {
+        auto converted_outputs = translate_node(decoder, node_idx);
+        if (converted_outputs.empty()) {
+            return;
+        }
+        const auto inplace_src = decoder->get_inplace_op_src(node_idx);
+        if (inplace_src.empty()) {
+            return;
+        }
+        if (converted_outputs[0].get_node_shared_ptr() != nullptr) {
+            (*tensor_map)[inplace_src] = converted_outputs[0];
+        }
+        refresh_inplace_aliases(decoder, node_idx, inplace_src);
+    };
+
+    if (!m_naive) {
+        preprocess(*tensor_map, *ggml_model_decoder);
+    }
+    ggml_model_decoder->visit_subgraph(node_visitor);
+
+    for (const auto & name : ggml_model_decoder->get_model_output_names()) {
+        FRONT_END_GENERAL_CHECK(tensor_map->find(name) != tensor_map->end(),
+                                "Output name not found in tensor map: ", name);
+        auto result = std::make_shared<v0::Result>(tensor_map->at(name));
+        result->set_friendly_name(name);
+        results.push_back(result);
+    }
+
+    // Debug-only hook: GGML_OPENVINO_DEBUG_NODE=<name1>,<name2>,... adds extra
+    // Result nodes for arbitrary intermediate tensors (looked up by name in
+    // tensor_map), on top of the real model outputs above. These debug
+    // Results are deliberately NOT added to ggml_decoder's model outputs, so
+    // the caller (ov_graph_compute_dynamic in utils.cpp) will not bind them
+    // to any ggml tensor buffer -- OpenVINO allocates its own tensor for
+    // them. This avoids the risk of reading a ggml buffer that has since
+    // been overwritten by a later in-place op (ggml aggressively reuses
+    // buffers), which can happen if trying to inspect an intermediate value
+    // via GGML_OPENVINO_DEBUG_OUTPUT by hacking it into a real output.
+    //
+    // tensor_map keys are usually the plain ggml tensor name (e.g. "embd"),
+    // but tensors that are recomputed multiple times in the same cgraph
+    // (GGML_TENSOR_FLAG_COMPUTE) are disambiguated with a "#<hash>" suffix
+    // (e.g. "cache_k_l0#4853", see get_tensor_ov_name()) which is not
+    // predictable ahead of time. To keep the env var usable, a requested
+    // name is matched either exactly, or as the "name" part before "#" of a
+    // suffixed key (first match wins; ambiguous requests should include the
+    // full "name#hash" form seen in a previous run's log/dump).
+    if (const char * debug_nodes = ggml_openvino_getenv_str("GGML_OPENVINO_DEBUG_NODE")) {
+        std::stringstream ss(debug_nodes);
+        std::string name;
+        while (std::getline(ss, name, ',')) {
+            auto it = tensor_map->find(name);
+            if (it == tensor_map->end()) {
+                it = std::find_if(tensor_map->begin(), tensor_map->end(), [&](const auto & entry) {
+                    return entry.first.compare(0, name.size(), name) == 0 && entry.first.size() > name.size() &&
+                           entry.first[name.size()] == '#';
+                });
+            }
+            if (it == tensor_map->end()) {
+                GGML_LOG_WARN("GGML_OPENVINO_DEBUG_NODE: node '%s' not found in tensor map, skipping\n", name.c_str());
+                continue;
+            }
+            auto result = std::make_shared<v0::Result>(it->second);
+            result->set_friendly_name("__debug_" + it->first);
+            results.push_back(result);
+        }
+    }
+
+    ov::ParameterVector used_params;
+    for (const auto & param : params) {
+        if (!param->output(0).get_target_inputs().empty()) {
+            used_params.push_back(param);
+        }
+    }
+    // if (auto diff = params.size() - used_params.size()) {
+    //     GGML_LOG_INFO("%zu parameters are not used in the model.", diff);
+    // }
+    resulting_model = std::make_shared<Model>(results, used_params);
+
+    apply_transformations(resulting_model);
+
+    // Set WeightlessCacheAttribute on large constants to avoid unnecessary memory copies
+    // in the NPUW plugin. Without this attribute, NPUW's LazyTensor constructor
+    // (lazy_tensor.cpp, op::Const::Const) will memcpy every constant "in case export
+    // occurs", doubling memory usage per compile_model call.
+    //
+    // The bin_offset field serves as a unique key (not a real file offset) — this is
+    // the same convention the GPU plugin uses for non-IR models (see
+    // Plugin::set_weightless_cache_attributes in intel_gpu/src/plugin/plugin.cpp).
+    // Each constant must have a distinct bin_offset, otherwise GPU's weightless cache
+    // import will map multiple constants to the same data.
+    //
+    // Small constants (< 16 elements) are excluded since they may be introduced by
+    // optimization patterns and the overhead is negligible.
+    //
+    // Note: use shape_size() rather than byte_size()/element_type().size() - GatherMatmul's default
+    // bias is a Constant(element::dynamic, Shape{0}), whose element_type().size() is 0 and would
+    // divide by zero.
+    size_t offset = 0;
+    for (auto & node : resulting_model->get_ordered_ops()) {
+        if (auto cnst = ov::as_type_ptr<ov::op::v0::Constant>(node); cnst && ov::shape_size(cnst->get_shape()) >= 16) {
+            auto & rt_info = cnst->get_rt_info();
+            if (rt_info.find(ov::WeightlessCacheAttribute::get_type_info_static()) == rt_info.end()) {
+                rt_info[ov::WeightlessCacheAttribute::get_type_info_static()] =
+                    ov::WeightlessCacheAttribute(cnst->get_byte_size(), offset++, cnst->get_element_type());
+            }
+        }
+    }
+    return resulting_model;
+}
+
+std::shared_ptr<Model> TranslateSession::apply_transformations(std::shared_ptr<Model> model) {
+    auto ggml_model_decoder = ov::as_type_ptr<InputModel>(m_input_model)->get_model_decoder();
+    {
+        ov::pass::Manager manager;
+        manager.set_per_pass_validation(true);
+        manager.register_pass<ov::pass::MarkCompressedFloatConstants>();
+        // Marks the Convert/Subtract/Multiply nodes of our GatherMatmul dequantization chain
+        // (make_int4_weights/make_int8_weights, for_gather_matmul=true) with disable_constant_folding,
+        // so it survives ConstantFolding regardless of whether the target plugin's own
+        // is_decompression_multiply() recognizes GatherMatmul as a valid consumer.
+        manager.register_pass<ov::pass::MarkDequantization>(
+            std::vector<ov::element::Type>{ov::element::u8, ov::element::i8, ov::element::u4, ov::element::i4});
+        manager.register_pass<pass::FuseToConv>();
+        manager.register_pass<pass::FuseArgsortTopK>();
+
+        // MOECompressed has no CPU plugin implementation, so enable it by default only on GPU.
+        // GGML_OPENVINO_MOE_OP=0 keeps the unfused GatherMatmul path for fallback/debugging.
+        if (ggml_openvino_is_gpu() &&
+            ggml_openvino_getenv_int("GGML_OPENVINO_MOE_OP", 1) != 0) {
+            manager.register_pass<pass::FuseMoeRouter>();
+            manager.register_pass<pass::FuseMoeCompressed>();
+            // Same fusion for models whose gate and up projections share one fused expert
+            // weight (gemma-4). It only matches that shape and only when the experts carry an
+            // integer zero point, so it is a no-op on the separate-gate/up models above.
+            manager.register_pass<pass::FuseMoeCompressedFusedGateUp>();
+        }
+
+        // Workaround for an OpenVINO GPU-plugin defect: an eltwise op whose operands differ in
+        // rank is computed wrongly once the plugin fuses it as a post-op into `rms`. gemma-4
+        // dense under stateful execution adds a rank-3 residual to the rank-4 norm output and
+        // decodes garbage on GPU while CPU is correct. Equalising the ranks is a no-op for
+        // NUMPY broadcasting and makes the fused path correct.
+        // Remove once the plugin guards that fusion. Opt out with
+        // GGML_OPENVINO_DISABLE_ELTWISE_RANK_ALIGN=1.
+        if (ggml_openvino_is_gpu() && !ggml_openvino_getenv_int("GGML_OPENVINO_DISABLE_ELTWISE_RANK_ALIGN")) {
+            manager.register_pass<pass::AlignEltwiseOperandRanks>();
+        }
+
+        if (ggml_model_decoder->is_stateful()) {
+            const auto kv_param_res_names = ggml_model_decoder->get_kv_param_res_names();
+            const auto kv_param_res_pairs = get_kv_param_res_pairs(model, kv_param_res_names);
+            manager.register_pass<ov::pass::MakeStateful>(kv_param_res_pairs);
+            // Must run after MakeStateful, which is what creates the ReadValue/Assign pairs.
+            if (!ggml_openvino_getenv_int("GGML_OPENVINO_DISABLE_KV_STATE_RELAYOUT")) {
+                manager.register_pass<pass::KVStateSeqAxis>();
+            }
+        }
+
+        if (ggml_model_decoder->is_static()) {
+            manager.register_pass<pass::SqueezeMatmul>();
+        }
+        manager.run_passes(model);
+        if (ggml_model_decoder->is_stateful()) {
+            ov::preprocess::PrePostProcessor ppp(model);
+            for (size_t i = 0; i < model->get_output_size(); i++) {
+                auto model_output_shape = model->output(i).get_partial_shape();
+                if (model_output_shape.rank().is_static() && model_output_shape.rank().get_length() == 3) {
+                    ppp.output(i).postprocess().custom([](const ov::Output<ov::Node>& node) {
+                        auto axes = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{1}, {0});
+                        return std::make_shared<ov::op::v0::Unsqueeze>(node, axes);
+                    });
+                }
+            }
+            model = ppp.build();
+        }
+    }
+    return model;
+}
+
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/translate_session.h

@@ -0,0 +1,29 @@
+#pragma once
+
+#include "input_model.h"
+#include "node_context.h"
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+
+class TranslateSession {
+public:
+    TranslateSession(const frontend::InputModel::Ptr & input_model,
+                     const std::unordered_map<std::string, CreatorFunction> & translator_map,
+                     bool naive = false);
+
+    std::shared_ptr<Model> get_converted_model();
+    std::shared_ptr<Model> translate_graph(const frontend::InputModel::Ptr & input_model);
+
+private:
+    std::shared_ptr<Model> apply_transformations(std::shared_ptr<Model> model);
+    const frontend::InputModel::Ptr m_input_model;
+    const std::unordered_map<std::string, CreatorFunction> & m_translator_map;
+    std::shared_ptr<Model> m_ov_model;
+    bool m_naive;
+};
+
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/utils.cpp

@@ -0,0 +1,987 @@
+#include "utils.h"
+
+#include <cmath>
+#include <cstddef>
+#include <ctime>
+#include <memory>
+#include <openvino/op/add.hpp>
+#include <openvino/op/clamp.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/cos.hpp>
+#include <openvino/op/divide.hpp>
+#include <openvino/op/gather.hpp>
+#include <openvino/op/maximum.hpp>
+#include <openvino/op/multiply.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/shape_of.hpp>
+#include <openvino/op/sin.hpp>
+#include <openvino/op/slice.hpp>
+#include <openvino/op/split.hpp>
+#include <openvino/op/squeeze.hpp>
+#include <openvino/op/subtract.hpp>
+#include <openvino/op/transpose.hpp>
+#include <string>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+
+void num_inputs_check(const NodeContext & context, size_t min_inputs, size_t max_inputs) {
+    auto input_size = context.get_input_size();
+    FRONT_END_OP_CONVERSION_CHECK(input_size >= min_inputs, "Got less inputs than expected");
+    FRONT_END_OP_CONVERSION_CHECK(input_size <= max_inputs, "Got more inputs than expected");
+}
+
+int non_cont_dim(std::vector<size_t> ne, std::vector<size_t> nb) {
+    int dim = nb.size() - 1;
+    size_t bytes = nb[dim];
+    for (int i = dim; i > 0; i--) {
+        bytes *= ne[i];
+        if (bytes != nb[i - 1]) {
+            return i;
+        }
+    }
+    return 0;
+}
+
+std::shared_ptr<ov::Node> get_dimensions(const std::shared_ptr<ov::op::v3::ShapeOf> & shape,
+                                         const std::vector<int> & dims) {
+    using namespace ov::op;
+    const auto zero = v0::Constant::create(ov::element::i32, ov::Shape{}, {0});
+    const auto dims_const = v0::Constant::create(ov::element::i32, ov::Shape{dims.size()}, dims);
+    return std::make_shared<v8::Gather>(shape, dims_const, zero);
+}
+
+std::shared_ptr<ov::Node> get_dimensions(const std::shared_ptr<ov::Node> & node, const std::vector<int> & dims) {
+    return get_dimensions(std::make_shared<ov::op::v3::ShapeOf>(node), dims);
+}
+
+OutputVector rename_outputs_with_suffix(const OutputVector & outputs, const std::string & suffix) {
+    for (const auto & output : outputs) {
+        auto node = output.get_node_shared_ptr();
+        std::string name = node->get_friendly_name();
+        name += "_";
+        name += suffix;
+        node->set_friendly_name(name);
+        // Uncomment to dump every node's inferred shape (used to hunt down dynamic dims on NPU).
+        // std::cout << name << "  " << output.get_partial_shape() << std::endl;
+    }
+    return outputs;
+}
+
+namespace {
+ov::Output<ov::Node> rope_yarn_ramp_mix(int num_elements, const float corr_dims[2], float ext_factor) {
+    std::vector<float> dim_ids_vec(num_elements);
+    std::iota(dim_ids_vec.begin(), dim_ids_vec.end(), 0.0f);
+    auto dim_ids = ov::op::v0::Constant::create(ov::element::f32, Shape{1, 1, 1, (size_t) num_elements}, dim_ids_vec);
+    auto corr_low = ov::op::v0::Constant::create(ov::element::f32, Shape{1, 1, 1, 1}, {corr_dims[0]});
+    auto corr_high = ov::op::v0::Constant::create(ov::element::f32, Shape{1, 1, 1, 1}, {corr_dims[1]});
+    auto denom = std::make_shared<ov::op::v1::Maximum>(
+        std::make_shared<ov::op::v1::Subtract>(corr_high, corr_low),
+        ov::op::v0::Constant::create(ov::element::f32, Shape{1, 1, 1, 1}, {0.001f}));
+    auto ramp_y =
+        std::make_shared<ov::op::v1::Divide>(std::make_shared<ov::op::v1::Subtract>(dim_ids, corr_low), denom);
+    auto ramp_clamped = std::make_shared<ov::op::v0::Clamp>(ramp_y, 0.0f, 1.0f);
+    // rope_yarn_ramp returns (1 - clamp(y)), so invert before scaling
+    auto one = ov::op::v0::Constant::create(ov::element::f32, Shape{1, 1, 1, 1}, {1.0f});
+    auto ramp_inverted = std::make_shared<ov::op::v1::Subtract>(one, ramp_clamped);
+    auto ext_factor_node = ov::op::v0::Constant::create(ov::element::f32, Shape{}, {ext_factor});
+    auto ramp_mix = std::make_shared<ov::op::v1::Multiply>(ramp_inverted, ext_factor_node);
+    return ramp_mix;
+}
+
+float ggml_rope_yarn_corr_dim(int n_dims, int n_ctx_orig, float n_rot, float base) {
+#ifndef M_PI
+#    define M_PI 3.14159265358979323846
+#endif
+    return n_dims * logf(n_ctx_orig / (n_rot * 2 * (float) M_PI)) / (2 * logf(base));
+}
+
+void ggml_rope_yarn_corr_dims(int n_dims,
+                              int n_ctx_orig,
+                              float freq_base,
+                              float beta_fast,
+                              float beta_slow,
+                              float dims[2]) {
+    float start = floorf(ggml_rope_yarn_corr_dim(n_dims, n_ctx_orig, beta_fast, freq_base));
+    float end = ceilf(ggml_rope_yarn_corr_dim(n_dims, n_ctx_orig, beta_slow, freq_base));
+    dims[0] = std::max(0.0f, start);
+    dims[1] = std::min(static_cast<float>(n_dims - 1), end);
+}
+}  // namespace
+
+std::pair<ov::Output<Node>, ov::Output<Node>> make_sin_cos(int32_t * rope_params,
+                                                           std::shared_ptr<ov::Node> inp_pos,
+                                                           std::shared_ptr<ov::Node> rope_freqs_weight,
+                                                           int mode,
+                                                           bool stateful,
+                                                           int64_t head_dim) {
+    constexpr int TYPE_IMROPE = 2;
+    constexpr int TYPE_VISION = 3;
+    constexpr int TYPE_MROPE = 4;
+
+    const bool is_imrope = (mode == TYPE_IMROPE);
+    const bool is_vision = (mode == TYPE_VISION);
+    const bool is_mrope = (mode == TYPE_MROPE);
+    const bool is_multi = is_imrope || is_vision || is_mrope;
+
+    if (stateful) {
+        inp_pos =
+            std::make_shared<ov::op::v0::Squeeze>(inp_pos, ov::op::v0::Constant::create(ov::element::i64, {1}, {0}));
+        inp_pos = std::make_shared<ov::op::v0::Convert>(inp_pos, ov::element::f32);
+        auto pos_perm =
+            std::make_shared<ov::op::v0::Constant>(ov::element::i64, ov::Shape{3}, std::vector<int64_t>{2, 1, 0});
+        inp_pos = std::make_shared<ov::op::v1::Transpose>(inp_pos, pos_perm);
+    } else if (is_multi) {
+        inp_pos = std::make_shared<ov::op::v0::Convert>(inp_pos, ov::element::f32);
+        auto pos_shape = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{5}, {0, 0, 0, 4, -1});
+        inp_pos = std::make_shared<ov::op::v1::Reshape>(inp_pos, pos_shape, true);
+        auto pos_transpose_shape =
+            std::make_shared<ov::op::v0::Constant>(ov::element::i64, ov::Shape{5}, std::vector<int64_t>{0, 1, 2, 4, 3});
+        inp_pos = std::make_shared<ov::op::v1::Transpose>(inp_pos, pos_transpose_shape);
+    } else {
+        inp_pos = std::make_shared<ov::op::v0::Convert>(inp_pos, ov::element::f32);
+        auto pos_perm =
+            std::make_shared<ov::op::v0::Constant>(ov::element::i64, ov::Shape{4}, std::vector<int64_t>{0, 3, 1, 2});
+        inp_pos = std::make_shared<ov::op::v1::Transpose>(inp_pos, pos_perm);
+    }
+
+    float freq_base;
+    float freq_scale;
+    float ext_factor;
+    float attn_factor;
+    float beta_fast;
+    float beta_slow;
+    const int n_dims = rope_params[1];
+    const size_t n_dims_half = n_dims >> 1;
+    const int n_ctx_orig = rope_params[4];
+    memcpy(&freq_base, rope_params + 5, sizeof(float));
+    memcpy(&freq_scale, rope_params + 6, sizeof(float));
+    memcpy(&ext_factor, rope_params + 7, sizeof(float));
+    memcpy(&attn_factor, rope_params + 8, sizeof(float));
+    memcpy(&beta_fast, rope_params + 9, sizeof(float));
+    memcpy(&beta_slow, rope_params + 10, sizeof(float));
+
+    const float theta_scale = powf(freq_base, -2.0f / n_dims);
+
+    Output<Node> theta;
+    float mscale = attn_factor;
+    if (is_multi) {
+        const size_t num_pairs = is_vision ? (n_dims > 0 ? (size_t) n_dims : (size_t) (head_dim / 2)) : n_dims_half;
+        int sections[4];
+        memcpy(sections, rope_params + 11, sizeof(int) * 4);
+        int sect_dims = sections[0] + sections[1] + sections[2] + sections[3];
+        if (sect_dims <= 0) {
+            sect_dims = num_pairs;
+        }
+        int sec_w = sections[0] + sections[1];
+        int sec_e = sec_w + sections[2];
+        std::vector<int64_t> gather_indices(num_pairs);
+        std::vector<float> factor(num_pairs);
+        for (size_t j = 0; j < num_pairs; j++) {
+            int sector = j % sect_dims;
+            int plane = 0;
+            if (is_imrope) {
+                if (sector % 3 == 1 && sector < 3 * sections[1]) {
+                    plane = 1;
+                } else if (sector % 3 == 2 && sector < 3 * sections[2]) {
+                    plane = 2;
+                } else if (sector % 3 == 0 && sector < 3 * sections[0]) {
+                    plane = 0;
+                } else {
+                    plane = 3;
+                }
+                factor[j] = std::pow(theta_scale, j);
+            } else if (is_vision) {
+                int p_idx = 0;
+                if (sector < sections[0]) {
+                    plane = 0;
+                    p_idx = sector;
+                } else if (sector < sections[0] + sections[1]) {
+                    plane = 1;
+                    p_idx = sector - sections[0];
+                } else if (sector < sec_w + sections[2]) {
+                    plane = 2;
+                    p_idx = sector - sec_w;
+                } else {
+                    plane = 3;
+                    p_idx = sector - sec_e;
+                }
+                factor[j] = std::pow(theta_scale, p_idx);
+            } else {
+                if (sector >= sections[0] && sector < sec_w) {
+                    plane = 1;
+                } else if (sector >= sec_w && sector < sec_e) {
+                    plane = 2;
+                } else if (sector >= sec_e) {
+                    plane = 3;
+                } else {
+                    plane = 0;
+                }
+                factor[j] = std::pow(theta_scale, j);
+            }
+            gather_indices[j] = plane;
+        }
+        auto gather_indices_const =
+            std::make_shared<ov::op::v0::Constant>(ov::element::i64, ov::Shape{num_pairs}, gather_indices);
+        auto gather_axis = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{}, {4});
+        inp_pos = std::make_shared<ov::op::v8::Gather>(inp_pos, gather_indices_const, gather_axis);
+        Output<Node> factor_node =
+            std::make_shared<ov::op::v0::Constant>(ov::element::f32, ov::Shape{num_pairs}, factor);
+        if (rope_freqs_weight) {
+            Output<Node> rope_factors = std::make_shared<ov::op::v8::Slice>(
+                rope_freqs_weight,
+                ov::op::v0::Constant::create(ov::element::i64, {1}, {0}),
+                ov::op::v0::Constant::create(ov::element::i64, {1}, {(int64_t) num_pairs}),
+                ov::op::v0::Constant::create(ov::element::i64, {1}, {1}),
+                ov::op::v0::Constant::create(ov::element::i64, {1}, {rope_freqs_weight->get_output_partial_shape(0).rank().get_length() - 1}));
+            rope_factors = std::make_shared<ov::op::v1::Reshape>(
+                rope_factors,
+                ov::op::v0::Constant::create(ov::element::i64, {1}, {(int64_t) num_pairs}),
+                false);
+            factor_node = std::make_shared<ov::op::v1::Divide>(factor_node, rope_factors);
+        }
+        auto theta_extrap = std::make_shared<ov::op::v1::Multiply>(inp_pos, factor_node);
+        auto theta_interp = std::make_shared<ov::op::v1::Multiply>(
+            theta_extrap, ov::op::v0::Constant::create(ov::element::f32, {1}, {freq_scale}));
+        if (ext_factor == 0.0f) {
+            theta = theta_interp;
+        } else {
+            float corr_dims[2];
+            ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
+            auto ramp_mix = rope_yarn_ramp_mix(num_pairs, corr_dims, ext_factor);
+            Output<Node> one = ov::op::v0::Constant::create(ov::element::f32, Shape{1, 1, 1, 1}, {1.0f});
+            auto one_minus_ramp = std::make_shared<ov::op::v1::Subtract>(one, ramp_mix);
+            theta = std::make_shared<ov::op::v1::Add>(
+                std::make_shared<ov::op::v1::Multiply>(theta_interp, one_minus_ramp),
+                std::make_shared<ov::op::v1::Multiply>(theta_extrap, ramp_mix));
+            mscale *= (1.0f + 0.1f * std::log(1.0f / freq_scale));
+        }
+    } else {
+        std::vector<float> factor(n_dims_half);
+        Output<Node> freq_factors;
+        float corr_dims[2];
+        ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
+        factor[0] = 1.0f;
+        for (size_t i = 1; i < factor.size(); i++) {
+            factor[i] = theta_scale * factor[i - 1];
+        }
+        if (stateful) {
+            freq_factors =
+                std::make_shared<ov::op::v0::Constant>(ov::element::f32, ov::Shape{1, 1, factor.size()}, factor);
+        } else {
+            freq_factors =
+                std::make_shared<ov::op::v0::Constant>(ov::element::f32, ov::Shape{1, 1, 1, factor.size()}, factor);
+        }
+        if (rope_freqs_weight) {
+            Output<Node> rope_factors = std::make_shared<ov::op::v8::Slice>(
+                rope_freqs_weight,
+                ov::op::v0::Constant::create(ov::element::i64, {1}, {0}),
+                ov::op::v0::Constant::create(ov::element::i64, {1}, {(int64_t) n_dims_half}),
+                ov::op::v0::Constant::create(ov::element::i64, {1}, {1}),
+                ov::op::v0::Constant::create(ov::element::i64, {1}, {rope_freqs_weight->get_output_partial_shape(0).rank().get_length() - 1}));
+            if (stateful) {
+                rope_factors = std::make_shared<ov::op::v1::Reshape>(
+                    rope_factors,
+                    ov::op::v0::Constant::create(ov::element::i64, {3}, {(int64_t) 1, (int64_t) 1, (int64_t) n_dims_half}),
+                    false);
+            } else {
+                rope_factors = std::make_shared<ov::op::v1::Reshape>(
+                    rope_factors,
+                    ov::op::v0::Constant::create(ov::element::i64, {4}, {(int64_t) 1, (int64_t) 1, (int64_t) 1, (int64_t) n_dims_half}),
+                    false);
+            }
+            freq_factors = std::make_shared<ov::op::v1::Divide>(freq_factors, rope_factors);
+        }
+
+        auto theta_extrap = std::make_shared<ov::op::v1::Multiply>(freq_factors, inp_pos);
+        auto theta_interp = std::make_shared<ov::op::v1::Multiply>(
+            theta_extrap, ov::op::v0::Constant::create(ov::element::f32, {1}, {freq_scale}));
+
+        if (ext_factor == 0.0f) {
+            theta = theta_interp;
+        } else {
+            auto ramp_mix = rope_yarn_ramp_mix(n_dims_half, corr_dims, ext_factor);
+            Output<Node> one;
+            if (stateful) {
+                one = ov::op::v0::Constant::create(ov::element::f32, Shape{1, 1, 1}, {1.0f});
+            } else {
+                one = ov::op::v0::Constant::create(ov::element::f32, Shape{1, 1, 1, 1}, {1.0f});
+            }
+            auto one_minus_ramp = std::make_shared<ov::op::v1::Subtract>(one, ramp_mix);
+
+            theta =
+                std::make_shared<ov::op::v1::Add>(std::make_shared<ov::op::v1::Multiply>(theta_interp, one_minus_ramp),
+                                                  std::make_shared<ov::op::v1::Multiply>(theta_extrap, ramp_mix));
+            mscale *= (1.0f + 0.1f * std::log(1.0f / freq_scale));
+        }
+    }
+
+    Output<Node> cos_theta = std::make_shared<ov::op::v0::Cos>(theta);
+    Output<Node> sin_theta = std::make_shared<ov::op::v0::Sin>(theta);
+
+    if (mscale != 1.0f) {
+        auto mscale_node = ov::op::v0::Constant::create(ov::element::f32, Shape{}, {mscale});
+
+        cos_theta = std::make_shared<ov::op::v1::Multiply>(cos_theta, mscale_node);
+        sin_theta = std::make_shared<ov::op::v1::Multiply>(sin_theta, mscale_node);
+    }
+
+    return std::make_pair(sin_theta, cos_theta);
+}
+
+ov::Output<ov::Node> process_view_input(const NodeContext & context, int input_index, int slice_len, int axis) {
+    // Only works for VIEW operations that does a non-strided slice with optinal reshape on the slice result.
+    // The function only does the slice part, the reshape (if any) should be handled by the caller.
+    // Default axis is -1, which means slicing the last dimension.
+    // If the VIEW reshapes the result, `slice_len` should be provided
+    auto input = context.get_input(input_index);
+    auto * op_params = (size_t *) context.get_input_op_params(input_index);
+    auto src_stride = context.get_input_stride(input_index);
+
+    int64_t slice_start = op_params[0] / src_stride[3];
+    if (slice_len == 0) {
+        slice_len = context.get_input_shape(input_index)[3].get_length();
+    }
+    int64_t slice_end = slice_start + slice_len;
+
+    auto begin = ov::op::v0::Constant::create(ov::element::i64, {1}, {slice_start});
+    auto end = ov::op::v0::Constant::create(ov::element::i64, {1}, {slice_end});
+    auto stride = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+    ov::Output<ov::Node> axes;
+    if (axis == -1) {
+        axes = ov::op::v0::Constant::create(ov::element::i64, {1}, {context.is_stateful() ? 2 : 3});
+    } else {
+        axes = ov::op::v0::Constant::create(ov::element::i64, {1}, {axis});
+    }
+    auto sliced = std::make_shared<ov::op::v8::Slice>(input, begin, end, stride, axes);
+    return sliced;
+}
+
+ov::Output<ov::Node> process_view_input_new(const NodeContext & context, int input_index) {
+    auto input = context.get_input(input_index);
+
+    // Check if this input has view inputs
+    size_t view_input_size = context.get_view_input_size(input_index);
+    if (view_input_size == 0) {
+        // No view inputs, return the input as is
+        return input;
+    }
+
+    // If translate_view already resolved this VIEW (produced a Slice), the input
+    // will already have the expected shape — skip re-slicing.
+    //
+    // Two notions of "matches" are accepted per axis:
+    //   - both dims static and equal, OR
+    //   - both dims dynamic.
+    // The dynamic case matters for the MoE expert-plane views: translate_view now emits a
+    // DYNAMIC-token slice (so the token dim is not frozen). An all-static-only check would
+    // see the dynamic token dim, decide the shapes "don't match", and fall through to
+    // re-slice/flatten the already-resolved view (a Reshape to the full flattened
+    // n_expert_used*n_embd tail, which then conflicts with the single-plane input). Treat a
+    // dynamic-vs-dynamic axis as matching so the already-resolved view is reused as-is.
+    //
+    // A third case matters for split-model MoE fragments: translate_view resolves the
+    // expert-plane view against the fragment's INPUT parameter. When the graph is split
+    // the token axis of that parameter may already be concrete (static n_tokens) even
+    // though get_view_input_ov_shape() still reports it as dynamic (-1). The resolved
+    // view is then static [1,1,n_tokens,n_embd] while `expected` is [1,1,?,n_embd].
+    // An "expected dynamic, actual static" axis is a valid concretization of the SAME
+    // resolved view, so treat it as matching too. Falling through to process_single_view
+    // here would re-slice/re-flatten the already-resolved single-plane view against the
+    // recorded (multi-plane) source strides and emit a constant-target Reshape whose baked
+    // dims no longer divide the concretized input -> "dimensions do not evenly divide".
+    // A fourth case matters for stateful execution: `expected` comes from ggml metadata and is
+    // always GGML_MAX_DIMS ranks, but stateful drops the leading size-1 batch dim, so the
+    // resolved view is one rank lower. Compare the common trailing dims and require the
+    // leading expected dims we skip to be 1. Without this the rank test below never matches on
+    // the stateful path and every already-resolved MoE expert-plane view is re-sliced.
+    auto expected_ov_shape = context.get_view_input_ov_shape(input_index, 0);
+    auto actual_shape = input.get_partial_shape();
+    if (expected_ov_shape.rank().is_static() && actual_shape.rank().is_static() &&
+        expected_ov_shape.rank().get_length() >= actual_shape.rank().get_length()) {
+        const int64_t n_actual = actual_shape.rank().get_length();
+        const int64_t shift = expected_ov_shape.rank().get_length() - n_actual;
+        bool shapes_match = true;
+        for (int64_t i = 0; i < shift; ++i) {
+            if (!expected_ov_shape[i].is_static() || expected_ov_shape[i].get_length() != 1) {
+                shapes_match = false;
+                break;
+            }
+        }
+        for (int64_t i = 0; i < n_actual && shapes_match; ++i) {
+            const auto & exp = expected_ov_shape[i + shift];
+            const bool both_dynamic = exp.is_dynamic() && actual_shape[i].is_dynamic();
+            const bool both_static_equal =
+                exp.is_static() && actual_shape[i].is_static() && exp == actual_shape[i];
+            // expected dynamic, actual static: the resolved view already carries the
+            // concrete size for this fragment; reuse it rather than re-materializing.
+            const bool expected_dyn_actual_static = exp.is_dynamic() && actual_shape[i].is_static();
+            if (!both_dynamic && !both_static_equal && !expected_dyn_actual_static) {
+                shapes_match = false;
+                break;
+            }
+        }
+        if (shapes_match) {
+            return input;
+        }
+    }
+
+    // In static mode, use Split instead of Slice for single-dimension reductions.
+    // This ensures NPUW's FOLD doesn't parametrize per-layer slice indices (which
+    // would introduce dynamic shapes). A shared Split node sits outside the repeated
+    // subgraph boundary; each layer receives one of its output ports.
+    if (context.is_static() && view_input_size == 1) {
+        auto view_stride_v = context.get_view_input_stride(input_index, 0);
+        auto view_src_stride_v = context.get_view_input_src_stride(input_index, 0);
+        auto view_ggml_shape = context.get_view_input_ggml_shape(input_index, 0);
+        auto view_src_ggml_shape = context.get_view_input_src_ggml_shape(input_index, 0);
+        auto view_offset = context.get_view_input_offset(input_index, 0);
+        auto view_src_offset = context.get_view_input_src_offset(input_index, 0);
+
+        size_t ndims = view_ggml_shape.size();
+        std::vector<int> diff_dims;
+        if (view_src_ggml_shape.size() == ndims) {
+            for (size_t i = 0; i < ndims; ++i) {
+                if (view_ggml_shape[i] != view_src_ggml_shape[i]) {
+                    diff_dims.push_back(static_cast<int>(i));
+                }
+            }
+        }
+
+        if (diff_dims.size() == 1) {
+            int split_dim = diff_dims[0];
+            int64_t num_splits = static_cast<int64_t>(view_src_ggml_shape[split_dim]);
+            int64_t chunk_size = static_cast<int64_t>(view_ggml_shape[split_dim]);
+
+            // Only apply when slicing exactly 1 element from a multi-element dimension
+            if (chunk_size == 1 && num_splits > 1) {
+                // Check suffix strides match (dimensions after split_dim)
+                bool suffix_ok = view_stride_v.size() == view_src_stride_v.size();
+                if (suffix_ok) {
+                    for (size_t i = static_cast<size_t>(split_dim) + 1; i < ndims; ++i) {
+                        if (view_stride_v[i] != view_src_stride_v[i]) {
+                            suffix_ok = false;
+                            break;
+                        }
+                    }
+                }
+
+                if (suffix_ok && view_src_stride_v[split_dim] > 0) {
+                    size_t relative_offset = view_offset >= view_src_offset ? view_offset - view_src_offset : 0;
+                    int64_t split_index = static_cast<int64_t>(relative_offset / view_src_stride_v[split_dim]);
+
+                    if (split_index >= 0 && split_index < num_splits) {
+                        auto src_node = input.get_node_shared_ptr();
+                        std::string rt_key = "split_dim_" + std::to_string(split_dim);
+                        auto & rt_info = src_node->get_rt_info();
+
+                        if (rt_info.find(rt_key) == rt_info.end()) {
+                            auto axis_const =
+                                ov::op::v0::Constant::create(ov::element::i64, {}, {static_cast<int64_t>(split_dim)});
+                            auto split_node =
+                                std::make_shared<ov::op::v1::Split>(input, axis_const, static_cast<size_t>(num_splits));
+                            split_node->set_friendly_name(src_node->get_friendly_name() + "_split");
+                            rt_info[rt_key] = split_node;
+                        }
+
+                        auto split_node = rt_info[rt_key].as<std::shared_ptr<ov::op::v1::Split>>();
+                        return split_node->output(static_cast<size_t>(split_index));
+                    }
+                }
+            }
+        }
+    }
+
+    // Lambda function to process a single view operation
+    auto process_single_view =
+        [](ov::Output<ov::Node> current, size_t view_offset, const std::vector<size_t> & view_stride,
+           const ov::Shape & view_ggml_shape, const ov::PartialShape & view_ov_shape, const std::string & view_name,
+           size_t view_src_offset, const std::vector<size_t> & view_src_stride, const ov::Shape & view_src_ggml_shape,
+           const ov::PartialShape & view_src_ov_shape, const std::string & view_src_name) -> ov::Output<ov::Node> {
+        // Stateful execution drops a leading size-1 axis, so `current` can be one rank lower
+        // than the ggml shape metadata (view_stride/view_ggml_shape, always GGML_MAX_DIMS
+        // entries) assumes. Shift any axis index built from that metadata down by the
+        // difference before using it as an OV Slice axis.
+        const auto current_rank = current.get_partial_shape().rank();
+        const int axis_shift = current_rank.is_static() ?
+                                    static_cast<int>(view_stride.size()) - static_cast<int>(current_rank.get_length()) :
+                                    0;
+
+        auto build_reshape_pattern = [](const ov::PartialShape & target_ov_shape,
+                                        const ov::Shape & target_ggml_shape) -> std::vector<int64_t> {
+            const size_t ndims = target_ggml_shape.size();
+            std::vector<int64_t> reshape_pattern(ndims);
+            size_t dynamic_dims = 0;
+
+            if (target_ov_shape.rank().is_static() &&
+                target_ov_shape.rank().get_length() == static_cast<int64_t>(ndims)) {
+                for (size_t i = 0; i < ndims; ++i) {
+                    if (target_ov_shape[i].is_static()) {
+                        reshape_pattern[i] = target_ov_shape[i].get_length();
+                    } else {
+                        reshape_pattern[i] = -1;
+                        ++dynamic_dims;
+                    }
+                }
+            } else {
+                dynamic_dims = 2;
+            }
+
+            if (dynamic_dims > 1) {
+                for (size_t i = 0; i < ndims; ++i) {
+                    reshape_pattern[i] = static_cast<int64_t>(target_ggml_shape[i]);
+                }
+            }
+
+            return reshape_pattern;
+        };
+
+        auto build_prefix_tail_reshape_pattern = [](const ov::PartialShape & target_ov_shape,
+                                                    const ov::Shape & target_ggml_shape, size_t prefix_dims,
+                                                    int64_t tail_dim) -> std::vector<int64_t> {
+            std::vector<int64_t> reshape_pattern(prefix_dims + 1);
+            size_t dynamic_dims = 0;
+
+            if (target_ov_shape.rank().is_static() &&
+                target_ov_shape.rank().get_length() == static_cast<int64_t>(target_ggml_shape.size())) {
+                for (size_t i = 0; i < prefix_dims; ++i) {
+                    if (target_ov_shape[i].is_static()) {
+                        reshape_pattern[i] = target_ov_shape[i].get_length();
+                    } else {
+                        reshape_pattern[i] = -1;
+                        ++dynamic_dims;
+                    }
+                }
+            } else {
+                dynamic_dims = 2;
+            }
+
+            if (dynamic_dims > 1) {
+                for (size_t i = 0; i < prefix_dims; ++i) {
+                    reshape_pattern[i] = static_cast<int64_t>(target_ggml_shape[i]);
+                }
+            }
+
+            reshape_pattern[prefix_dims] = tail_dim;
+            return reshape_pattern;
+        };
+
+        bool same_stride = view_stride.size() == view_src_stride.size();
+        if (same_stride) {
+            for (size_t i = 0; i < view_stride.size(); ++i) {
+                if (view_stride[i] != view_src_stride[i]) {
+                    same_stride = false;
+                    break;
+                }
+            }
+        }
+
+        bool same_ggml_shape = view_ggml_shape.size() == view_src_ggml_shape.size();
+        if (same_ggml_shape) {
+            for (size_t i = 0; i < view_ggml_shape.size(); ++i) {
+                if (view_ggml_shape[i] != view_src_ggml_shape[i]) {
+                    same_ggml_shape = false;
+                    break;
+                }
+            }
+        }
+
+        if (same_stride && same_ggml_shape) {
+            return current;
+        }
+
+        if (same_stride) {
+            const size_t relative_offset = view_offset >= view_src_offset ? view_offset - view_src_offset : 0;
+            const size_t ndims = view_stride.size();
+
+            std::vector<int> diff_dims;
+            if (view_ggml_shape.size() == ndims && view_src_ggml_shape.size() == ndims) {
+                for (size_t i = 0; i < ndims; ++i) {
+                    if (view_ggml_shape[i] != view_src_ggml_shape[i]) {
+                        diff_dims.push_back(static_cast<int>(i));
+                    }
+                }
+            }
+
+            if (diff_dims.size() == 1) {
+                const int slice_dim = diff_dims[0];
+                const int64_t dim_size = static_cast<int64_t>(view_src_ggml_shape[slice_dim]);
+
+                if (view_stride[slice_dim] > 0 && relative_offset % view_stride[slice_dim] == 0) {
+                    const int64_t begin_val = static_cast<int64_t>((relative_offset / view_stride[slice_dim]) %
+                                                                   static_cast<size_t>(dim_size));
+                    const int64_t end_val = begin_val + static_cast<int64_t>(view_ggml_shape[slice_dim]);
+
+                    if (begin_val >= 0 && end_val <= dim_size) {
+                        auto sliced = std::make_shared<ov::op::v8::Slice>(
+                            current, ov::op::v0::Constant::create(ov::element::i64, {1}, {begin_val}),
+                            ov::op::v0::Constant::create(ov::element::i64, {1}, {end_val}),
+                            ov::op::v0::Constant::create(ov::element::i64, {1}, {1}),
+                            ov::op::v0::Constant::create(ov::element::i64, {1}, {slice_dim - axis_shift}));
+
+                        if (view_ov_shape.is_static()) {
+                            auto reshaped = std::make_shared<ov::op::v1::Reshape>(
+                                sliced,
+                                ov::op::v0::Constant::create(ov::element::i64, {ndims}, view_ov_shape.to_shape()),
+                                false);
+                            reshaped->set_friendly_name(view_name);
+                            return reshaped;
+                        }
+
+                        sliced->set_friendly_name(view_name);
+                        return sliced;
+                    }
+                }
+
+                int64_t tail_src_elems = 1;
+                int64_t tail_dst_elems = 1;
+                for (size_t i = slice_dim; i < ndims; ++i) {
+                    tail_src_elems *= static_cast<int64_t>(view_src_ggml_shape[i]);
+                    tail_dst_elems *= static_cast<int64_t>(view_ggml_shape[i]);
+                }
+
+                const size_t elem_stride = view_stride[ndims - 1];
+                int64_t tail_begin = 0;
+                if (elem_stride > 0) {
+                    tail_begin =
+                        static_cast<int64_t>((relative_offset / elem_stride) % static_cast<size_t>(tail_src_elems));
+                }
+                const int64_t tail_end = tail_begin + tail_dst_elems;
+
+                if (tail_begin >= 0 && tail_end <= tail_src_elems) {
+                    std::vector<int64_t> flat_shape;
+                    flat_shape.reserve(slice_dim);
+                    for (int i = 0; i < slice_dim; ++i) {
+                        flat_shape.push_back(static_cast<int64_t>(view_src_ggml_shape[i]));
+                    }
+                    flat_shape.push_back(tail_src_elems);
+                    const size_t flat_ndims = flat_shape.size();
+
+                    auto flat = std::make_shared<ov::op::v1::Reshape>(
+                        current, ov::op::v0::Constant::create(ov::element::i64, {flat_ndims}, flat_shape), false);
+
+                    auto sliced = std::make_shared<ov::op::v8::Slice>(
+                        flat, ov::op::v0::Constant::create(ov::element::i64, {1}, {tail_begin}),
+                        ov::op::v0::Constant::create(ov::element::i64, {1}, {tail_end}),
+                        ov::op::v0::Constant::create(ov::element::i64, {1}, {1}),
+                        ov::op::v0::Constant::create(ov::element::i64, {1}, {slice_dim}));
+
+                    if (view_ov_shape.is_static()) {
+                        auto reshaped = std::make_shared<ov::op::v1::Reshape>(
+                            sliced, ov::op::v0::Constant::create(ov::element::i64, {ndims}, view_ov_shape.to_shape()),
+                            false);
+                        reshaped->set_friendly_name(view_name);
+                        return reshaped;
+                    }
+
+                    sliced->set_friendly_name(view_name);
+                    return sliced;
+                }
+            }
+
+            std::vector<int64_t> begin(ndims, 0);
+            std::vector<int64_t> end(ndims, 0);
+            std::vector<int64_t> step(ndims, 1);
+            std::vector<int64_t> axes(ndims, 0);
+
+            size_t remaining_offset = relative_offset;
+            for (size_t i = 0; i < ndims; ++i) {
+                axes[i] = static_cast<int64_t>(i);
+                if (view_stride[i] > 0) {
+                    begin[i] = static_cast<int64_t>(remaining_offset / view_stride[i]);
+                    remaining_offset %= view_stride[i];
+                }
+                end[i] = begin[i] + static_cast<int64_t>(view_ggml_shape[i]);
+            }
+
+            bool in_bounds = view_src_ggml_shape.size() == ndims && view_ggml_shape.size() == ndims;
+            if (in_bounds) {
+                for (size_t i = 0; i < ndims; ++i) {
+                    if (end[i] > static_cast<int64_t>(view_src_ggml_shape[i])) {
+                        in_bounds = false;
+                        break;
+                    }
+                }
+            }
+
+            if (in_bounds && remaining_offset == 0) {
+                auto sliced = std::make_shared<ov::op::v8::Slice>(
+                    current, ov::op::v0::Constant::create(ov::element::i64, {ndims}, begin),
+                    ov::op::v0::Constant::create(ov::element::i64, {ndims}, end),
+                    ov::op::v0::Constant::create(ov::element::i64, {ndims}, step),
+                    ov::op::v0::Constant::create(ov::element::i64, {ndims}, axes));
+
+                sliced->set_friendly_name(view_name);
+                return sliced;
+            }
+        } else {
+            bool same_rank = view_stride.size() == view_src_stride.size() &&
+                             view_ggml_shape.size() == view_src_ggml_shape.size() &&
+                             view_stride.size() == view_ggml_shape.size();
+            const size_t relative_offset = view_offset >= view_src_offset ? view_offset - view_src_offset : 0;
+
+            if (same_rank) {
+                const size_t ndims = view_ggml_shape.size();
+                std::vector<int> diff_dims;
+                for (size_t i = 0; i < ndims; ++i) {
+                    if (view_ggml_shape[i] != view_src_ggml_shape[i]) {
+                        diff_dims.push_back(static_cast<int>(i));
+                    }
+                }
+
+                if (diff_dims.size() == 1) {
+                    const size_t slice_dim = static_cast<size_t>(diff_dims[0]);
+                    bool suffix_stride_match = true;
+                    for (size_t i = slice_dim + 1; i < ndims; ++i) {
+                        if (view_stride[i] != view_src_stride[i]) {
+                            suffix_stride_match = false;
+                            break;
+                        }
+                    }
+
+                    if (suffix_stride_match && view_src_stride[slice_dim] > 0 &&
+                        relative_offset % view_src_stride[slice_dim] == 0) {
+                        const int64_t begin_val = static_cast<int64_t>(relative_offset / view_src_stride[slice_dim]);
+                        const int64_t end_val = begin_val + static_cast<int64_t>(view_ggml_shape[slice_dim]);
+                        const int64_t dim_size = static_cast<int64_t>(view_src_ggml_shape[slice_dim]);
+
+                        if (begin_val >= 0 && end_val <= dim_size) {
+                            auto sliced = std::make_shared<ov::op::v8::Slice>(
+                                current, ov::op::v0::Constant::create(ov::element::i64, {1}, {begin_val}),
+                                ov::op::v0::Constant::create(ov::element::i64, {1}, {end_val}),
+                                ov::op::v0::Constant::create(ov::element::i64, {1}, {1}),
+                                ov::op::v0::Constant::create(ov::element::i64, {1}, {static_cast<int64_t>(slice_dim)}));
+                            sliced->set_friendly_name(view_name);
+                            return sliced;
+                        }
+                    }
+                }
+            }
+
+            size_t view_elems = 1;
+            size_t src_elems = 1;
+            if (same_rank) {
+                for (size_t i = 0; i < view_ggml_shape.size(); ++i) {
+                    view_elems *= view_ggml_shape[i];
+                    src_elems *= view_src_ggml_shape[i];
+                }
+            }
+
+            bool same_num_elements = same_rank && view_elems == src_elems;
+
+            if (same_rank && relative_offset == 0 && same_num_elements) {
+                auto reshape_pattern = build_reshape_pattern(view_ov_shape, view_ggml_shape);
+
+                auto reshaped = std::make_shared<ov::op::v1::Reshape>(
+                    current, ov::op::v0::Constant::create(ov::element::i64, {reshape_pattern.size()}, reshape_pattern),
+                    false);
+                reshaped->set_friendly_name(view_name);
+                return reshaped;
+            }
+
+            if (same_rank) {
+                const size_t ndims = view_ggml_shape.size();
+
+                // Match views that can be expressed as a regular strided slice over the
+                // already reconstructed source tensor, e.g. offset on one axis plus step > 1
+                // on another axis.
+                bool is_regular_slice = view_src_ggml_shape.size() == ndims;
+                std::vector<int64_t> begin(ndims, 0);
+                std::vector<int64_t> end(ndims, 0);
+                std::vector<int64_t> step(ndims, 1);
+                std::vector<int64_t> axes(ndims, 0);
+                size_t remaining_offset = relative_offset;
+
+                if (is_regular_slice) {
+                    for (size_t i = 0; i < ndims; ++i) {
+                        axes[i] = static_cast<int64_t>(i);
+
+                        if (view_src_stride[i] == 0 || view_stride[i] == 0 ||
+                            view_stride[i] % view_src_stride[i] != 0) {
+                            is_regular_slice = false;
+                            break;
+                        }
+
+                        step[i] = static_cast<int64_t>(view_stride[i] / view_src_stride[i]);
+                        if (step[i] <= 0) {
+                            is_regular_slice = false;
+                            break;
+                        }
+
+                        begin[i] = static_cast<int64_t>(remaining_offset / view_src_stride[i]);
+                        remaining_offset %= view_src_stride[i];
+
+                        if (view_ggml_shape[i] == 0) {
+                            end[i] = begin[i];
+                            continue;
+                        }
+
+                        end[i] = begin[i] + step[i] * static_cast<int64_t>(view_ggml_shape[i] - 1) + 1;
+
+                        if (begin[i] < 0 || end[i] > static_cast<int64_t>(view_src_ggml_shape[i])) {
+                            is_regular_slice = false;
+                            break;
+                        }
+                    }
+                }
+
+                if (is_regular_slice && remaining_offset == 0) {
+                    auto sliced = std::make_shared<ov::op::v8::Slice>(
+                        current, ov::op::v0::Constant::create(ov::element::i64, {ndims}, begin),
+                        ov::op::v0::Constant::create(ov::element::i64, {ndims}, end),
+                        ov::op::v0::Constant::create(ov::element::i64, {ndims}, step),
+                        ov::op::v0::Constant::create(ov::element::i64, {ndims}, axes));
+
+                    sliced->set_friendly_name(view_name);
+                    return sliced;
+                }
+
+                const size_t elem_stride = view_src_stride.back();
+                const bool aligned_offset = elem_stride > 0 && relative_offset % elem_stride == 0;
+
+                if (aligned_offset) {
+                    size_t suffix_start = 0;
+                    size_t expected_stride = elem_stride;
+                    for (int i = static_cast<int>(ndims) - 1; i >= 0; --i) {
+                        if (view_stride[i] != expected_stride) {
+                            suffix_start = static_cast<size_t>(i + 1);
+                            break;
+                        }
+                        expected_stride *= view_ggml_shape[i];
+                    }
+
+                    size_t prefix_elems = 1;
+                    size_t suffix_elems = 1;
+                    for (size_t i = 0; i < suffix_start; ++i) {
+                        prefix_elems *= view_ggml_shape[i];
+                    }
+                    for (size_t i = suffix_start; i < ndims; ++i) {
+                        suffix_elems *= view_ggml_shape[i];
+                    }
+
+                    if (prefix_elems > 0 && src_elems % prefix_elems == 0) {
+                        const size_t src_tail_elems = src_elems / prefix_elems;
+                        const int64_t tail_begin = static_cast<int64_t>(relative_offset / elem_stride);
+                        const int64_t tail_end = tail_begin + static_cast<int64_t>(suffix_elems);
+
+                        if (tail_begin >= 0 && tail_end <= static_cast<int64_t>(src_tail_elems)) {
+                            auto prefix_tail_pattern = build_prefix_tail_reshape_pattern(
+                                view_ov_shape, view_ggml_shape, suffix_start, static_cast<int64_t>(src_tail_elems));
+
+                            auto prefix_tail = std::make_shared<ov::op::v1::Reshape>(
+                                current,
+                                ov::op::v0::Constant::create(ov::element::i64, {prefix_tail_pattern.size()},
+                                                             prefix_tail_pattern),
+                                false);
+
+                            ov::Output<ov::Node> selected = prefix_tail;
+                            if (tail_begin != 0 || tail_end != static_cast<int64_t>(src_tail_elems)) {
+                                selected = std::make_shared<ov::op::v8::Slice>(
+                                    prefix_tail, ov::op::v0::Constant::create(ov::element::i64, {1}, {tail_begin}),
+                                    ov::op::v0::Constant::create(ov::element::i64, {1}, {tail_end}),
+                                    ov::op::v0::Constant::create(ov::element::i64, {1}, {1}),
+                                    ov::op::v0::Constant::create(ov::element::i64, {1},
+                                                                 {static_cast<int64_t>(suffix_start)}));
+                            }
+
+                            auto reshape_pattern = build_reshape_pattern(view_ov_shape, view_ggml_shape);
+                            auto reshaped = std::make_shared<ov::op::v1::Reshape>(
+                                selected,
+                                ov::op::v0::Constant::create(ov::element::i64, {reshape_pattern.size()},
+                                                             reshape_pattern),
+                                false);
+                            reshaped->set_friendly_name(view_name);
+                            return reshaped;
+                        }
+                    }
+                }
+            }
+
+            return current;
+        }
+
+        (void) view_name;
+        (void) view_src_ov_shape;
+        (void) view_src_name;
+
+        return current;
+    };
+
+    // Special case: ggml collapses VIEW-of-VIEW chains so that `view_offs` is always an
+    // ABSOLUTE offset from the true root allocation, regardless of how many VIEW levels
+    // are in between (see ggml_new_tensor_impl). `src[0]` is still the immediate op-graph
+    // parent though, which can be a DIFFERENT (already narrowed) VIEW with the SAME ggml
+    // shape as this one but a different absolute offset -- e.g. a per-layer deepstack
+    // slice `view_2d(embd, n_embd, n_tokens, embd->nb[1], layer*n_embd*sizeof(float))`
+    // whose src[0] ("embd") is itself already a zero-offset VIEW of the true root (the
+    // padded embedding). Chaining through "embd" here would try to re-slice an already
+    // 2-narrowed tensor using a root-relative offset, going out of bounds and silently
+    // falling back to a no-op (returning the wrong, already-resolved sibling slice).
+    // Detect this (same shape as the immediate src, but different absolute offset) and
+    // re-slice directly from the untouched root using the innermost view's absolute
+    // offset against the ROOT's own shape/stride instead of chaining through src[0].
+    {
+        auto innermost_offset = context.get_view_input_offset(input_index, 0);
+        auto innermost_src_offset = context.get_view_input_src_offset(input_index, 0);
+        auto innermost_shape = context.get_view_input_ggml_shape(input_index, 0);
+        auto innermost_src_shape = context.get_view_input_src_ggml_shape(input_index, 0);
+        if (innermost_offset != innermost_src_offset && innermost_shape == innermost_src_shape) {
+            size_t root_view_idx = view_input_size - 1;
+            auto root_ggml_shape = context.get_view_input_src_ggml_shape(input_index, root_view_idx);
+            auto root_stride = context.get_view_input_src_stride(input_index, root_view_idx);
+            auto root_offset = context.get_view_input_src_offset(input_index, root_view_idx);
+            auto root_ov_shape = context.get_view_input_src_ov_shape(input_index, root_view_idx);
+            auto root_name = context.get_view_input_src_name(input_index, root_view_idx);
+            auto innermost_stride = context.get_view_input_stride(input_index, 0);
+            auto innermost_ov_shape = context.get_view_input_ov_shape(input_index, 0);
+            auto innermost_name = context.get_view_input_name(input_index, 0);
+
+            return process_single_view(input, innermost_offset, innermost_stride, innermost_shape, innermost_ov_shape,
+                                       innermost_name, root_offset, root_stride, root_ggml_shape, root_ov_shape,
+                                       root_name);
+        }
+    }
+
+    // Process views from the base tensor (last) to the current view (first)
+    // Start with the base tensor
+    ov::Output<ov::Node> current = input;
+
+    // Process each view in reverse order (from base to current)
+    for (int view_idx = view_input_size - 1; view_idx >= 0; view_idx--) {
+        auto view_offset = context.get_view_input_offset(input_index, view_idx);
+        auto view_stride = context.get_view_input_stride(input_index, view_idx);
+        auto view_ggml_shape = context.get_view_input_ggml_shape(input_index, view_idx);
+        auto view_ov_shape = context.get_view_input_ov_shape(input_index, view_idx);
+        auto view_name = context.get_view_input_name(input_index, view_idx);
+
+        // print view info
+        // std::cout << "View " << view_idx << ": name = " << view_name << ", offset = " << view_offset << ", stride = ["
+        //       << view_stride[0] << "," << view_stride[1] << "," << view_stride[2] << "," << view_stride[3]
+        //       << "], ggml shape = [" << view_ggml_shape[0] << "," << view_ggml_shape[1] << ","
+        //       << view_ggml_shape[2] << "," << view_ggml_shape[3] << "], ov shape = " << view_ov_shape << std::endl;
+
+        auto view_src_offset = context.get_view_input_src_offset(input_index, view_idx);
+        auto view_src_stride = context.get_view_input_src_stride(input_index, view_idx);
+        auto view_src_ggml_shape = context.get_view_input_src_ggml_shape(input_index, view_idx);
+        auto view_src_ov_shape = context.get_view_input_src_ov_shape(input_index, view_idx);
+        auto view_src_name = context.get_view_input_src_name(input_index, view_idx);
+        // print source view info
+        // std::cout << "View " << view_idx << ": source name = " << view_src_name
+        //           << ", source offset = " << view_src_offset << ", source stride = [" << view_src_stride[0] << ","
+        //           << view_src_stride[1] << "," << view_src_stride[2] << "," << view_src_stride[3]
+        //           << "], source ggml shape = [" << view_src_ggml_shape[0] << "," << view_src_ggml_shape[1] << ","
+        //           << view_src_ggml_shape[2] << "," << view_src_ggml_shape[3]
+        //           << "], source ov shape = " << view_src_ov_shape << std::endl;
+
+        current = process_single_view(current, view_offset, view_stride, view_ggml_shape, view_ov_shape, view_name,
+                                      view_src_offset, view_src_stride, view_src_ggml_shape, view_src_ov_shape,
+                                      view_src_name);
+    }
+
+    return current;
+}
+
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/openvino/utils.h

@@ -0,0 +1,102 @@
+#pragma once
+
+#include "node_context.h"
+
+#include <memory>
+#include <openvino/core/node.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/shape_of.hpp>
+#include <openvino/op/slice.hpp>
+#include <utility>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+
+std::string getCurrentTime();
+
+void num_inputs_check(const NodeContext & context, size_t min_inputs, size_t max_inputs);
+
+int non_cont_dim(std::vector<size_t> ne, std::vector<size_t> nb);
+
+template <typename T> std::vector<int> argsort_descend(const std::vector<T> & v) {
+    std::vector<int> idx(v.size());
+    std::iota(idx.begin(), idx.end(), 0);
+    std::sort(idx.begin(), idx.end(), [&v](int i1, int i2) { return v[i1] > v[i2]; });
+    return idx;
+}
+
+template <typename T> std::vector<T> sorted_descend(std::vector<T> v) {
+    std::sort(v.begin(), v.end(), [](T a, T b) { return a > b; });
+    return v;
+}
+
+template <typename T> bool is_permuted(const std::vector<T> & strides) {
+    for (size_t i = 0; i < strides.size() - 1; ++i) {
+        if (strides[i] < strides[i + 1]) {
+            return true;
+        }
+    }
+    return false;
+}
+
+template <typename T> std::vector<T> permute(const std::vector<T> & x, const std::vector<int> & perm) {
+    std::vector<T> result;
+    result.reserve(perm.size());
+    for (int i : perm) {
+        result.push_back(x[i]);
+    }
+    return result;
+}
+
+std::shared_ptr<ov::Node> get_dimensions(const std::shared_ptr<ov::op::v3::ShapeOf> & shape,
+                                         const std::vector<int> & dims);
+std::shared_ptr<ov::Node> get_dimensions(const std::shared_ptr<ov::Node> & node, const std::vector<int> & dims);
+
+OutputVector rename_outputs_with_suffix(const OutputVector & outputs, const std::string & suffix);
+
+std::pair<ov::Output<Node>, ov::Output<Node>> make_sin_cos(int32_t * rope_params,
+                                                           std::shared_ptr<ov::Node> inp_pos,
+                                                           std::shared_ptr<ov::Node> rope_freqs_weight = nullptr,
+                                                           int mode = 0,
+                                                           bool stateful = false,
+                                                           int64_t head_dim = 0);
+
+ov::Output<ov::Node> process_view_input(const NodeContext & context, int input_index, int slice_len = 0, int axis = -1);
+
+ov::Output<ov::Node> process_view_input_new(const NodeContext & context, int input_index);
+
+namespace op {
+template <typename T> OutputVector translate_1to1_match_2_inputs(const NodeContext & context) {
+    num_inputs_check(context, 2, 2);
+    auto input_0 = process_view_input_new(context, 0);
+    auto input_1 = process_view_input_new(context, 1);
+
+    auto output_type = context.get_output_type();
+    if (input_0.get_element_type() != input_1.get_element_type()) {
+        if (input_0.get_element_type() != ov::element::f32) {
+            input_0 = std::make_shared<ov::op::v0::Convert>(input_0, ov::element::f32);
+        }
+        if (input_1.get_element_type() != ov::element::f32) {
+            input_1 = std::make_shared<ov::op::v0::Convert>(input_1, ov::element::f32);
+        }
+    }
+
+    ov::Output<ov::Node> res = std::make_shared<T>(input_0, input_1);
+    if (res.get_element_type() != output_type) {
+        res = std::make_shared<ov::op::v0::Convert>(res, output_type);
+    }
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+template <typename T> OutputVector translate_1to1_match_1_input(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+    auto input = process_view_input_new(context, 0);
+    auto res = std::make_shared<T>(input);
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+}  // namespace op
+
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

llama.cpp/ggml/src/ggml-openvino/utils.h

@@ -0,0 +1,196 @@
+#include "ggml-decoder.h"
+#include "ggml-impl.h"
+#include "ggml-openvino-extra.h"
+
+#include <cstddef>
+#include <cstdint>
+#include <functional>
+#include <memory>
+#include <mutex>
+#include <openvino/runtime/core.hpp>
+#include <openvino/runtime/infer_request.hpp>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include <utility>
+#include <vector>
+
+// Cache key for a translated/compiled graph. Node count plus the two end node names identify a
+// graph during inference, where the same few graphs repeat for the whole session. The list of
+// external input names below tells those apart more precisely, but it walks every node and every
+// src slot, so it is built only when GGML_OPENVINO_FULL_GRAPH_KEY is set. Op tests run many small
+// graphs that can share a node count and end names, and need the full key.
+struct graph_key {
+    int n_nodes;
+    int n_leaves;
+    std::string first_node_name;
+    std::string last_node_name;
+    // Each entry: "<node_idx>:<src_idx>:<op_type>"
+    std::vector<std::string> input_srcs;
+
+    graph_key(const ggml_cgraph * cgraph, bool include_inputs = false) : n_nodes(cgraph->n_nodes), n_leaves(cgraph->n_leafs) {
+        if (n_nodes > 0) {
+            first_node_name = cgraph->nodes[0]->name;
+            last_node_name = cgraph->nodes[n_nodes - 1]->name;
+        }
+
+        static const bool full_key = ggml_openvino_getenv_int("GGML_OPENVINO_FULL_GRAPH_KEY") != 0;
+        if (!full_key && !include_inputs) {
+            return;
+        }
+
+        std::unordered_set<const ggml_tensor *> node_set;
+        node_set.reserve(cgraph->n_nodes);
+        for (int i = 0; i < cgraph->n_nodes; i++) {
+            node_set.insert(cgraph->nodes[i]);
+        }
+
+        for (int node_idx = 0; node_idx < cgraph->n_nodes; node_idx++) {
+            const ggml_tensor * node = cgraph->nodes[node_idx];
+            for (int src_idx = 0; src_idx < GGML_MAX_SRC; src_idx++) {
+                const ggml_tensor * src = node->src[src_idx];
+                if (src == nullptr || src->buffer == nullptr || node_set.count(src) || src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
+                    continue;
+                }
+
+                input_srcs.push_back(std::to_string(node_idx) + ":" + std::to_string(src_idx) + ":" +
+                    GgmlOvDecoder::compute_op_type(node));
+            }
+        }
+    }
+
+    bool operator==(const graph_key & other) const {
+        return n_nodes == other.n_nodes && n_leaves == other.n_leaves &&
+               first_node_name == other.first_node_name &&
+               last_node_name == other.last_node_name && input_srcs == other.input_srcs;
+    }
+};
+
+struct graph_key_hash {
+    size_t operator()(const graph_key & key) const {
+        size_t hash = std::hash<int>{}(key.n_nodes);
+        hash ^= std::hash<int>{}(key.n_leaves) + 0x9e3779b9 + (hash << 6) + (hash >> 2);
+        if (key.n_nodes > 0) {
+            hash ^= std::hash<std::string>{}(key.first_node_name) + 0x9e3779b9 + (hash << 6) + (hash >> 2);
+            hash ^= std::hash<std::string>{}(key.last_node_name) + 0x9e3779b9 + (hash << 6) + (hash >> 2);
+        }
+        for (const auto & s : key.input_srcs) {
+            hash ^= std::hash<std::string>{}(s) + 0x9e3779b9 + (hash << 6) + (hash >> 2);
+        }
+        return hash;
+    }
+};
+
+struct decoder_runtime_ctx {
+    decoder_runtime_ctx(std::shared_ptr<std::mutex> mutex) : mutex(std::move(mutex)) {}
+
+    std::shared_ptr<std::mutex> mutex;
+    std::shared_ptr<GgmlOvDecoder> ptr;
+};
+
+struct ov_compiled_graph {
+    ov::CompiledModel decode;
+    ov::CompiledModel prefill;
+    std::vector<std::string> input_names;
+    std::vector<std::string> output_names;
+};
+
+// Only compilation and cache publication use this mutex. Requests, decoders and
+// sequence state belong to individual backend contexts and never enter this cache.
+struct ov_compiled_model_cache {
+    std::mutex mutex;
+    std::unordered_map<std::string, ov_compiled_graph> graphs;
+    size_t backend_count = 0;
+};
+
+// Private to one backend instance. Only compiled_cache is shared with other
+// instances; clearing these local caches cannot invalidate their requests.
+struct ov_runtime_context {
+    // Serializes calls on this backend only, not inference in other contexts.
+    std::mutex execution_mutex;
+    std::shared_ptr<ov_compiled_model_cache> compiled_cache;
+    mutable std::mutex ctx_mutex;
+    std::string device;
+    bool stateful;
+    std::unordered_map<graph_key, std::shared_ptr<decoder_runtime_ctx>, graph_key_hash> decoder_cache;
+    std::unordered_map<graph_key, std::shared_ptr<ov::InferRequest>, graph_key_hash> infer_request_cache;
+    std::unordered_map<graph_key, std::shared_ptr<ov::InferRequest>, graph_key_hash> infer_request_cache_prefill;
+    std::unordered_map<graph_key, std::vector<std::string>, graph_key_hash> ov_input_names_cache;
+    std::unordered_map<graph_key, std::vector<std::string>, graph_key_hash> ov_output_names_cache;
+    size_t stateful_kv_size;
+    std::map<std::string, std::string> kv_state_input_name_map;
+    // compute-only copies of graphs that carry unselected ggml_build_forward_select() branches
+    struct compute_graph {
+        ggml_cgraph graph;
+        std::vector<ggml_tensor *> nodes;
+    };
+    std::unordered_map<const ggml_cgraph *, compute_graph> compute_graphs;
+
+    ov_runtime_context() : device("CPU"), stateful(false), stateful_kv_size(0) {}
+
+    void clear_caches_locked() {
+        decoder_cache.clear();
+        infer_request_cache.clear();
+        infer_request_cache_prefill.clear();
+        ov_input_names_cache.clear();
+        ov_output_names_cache.clear();
+        kv_state_input_name_map.clear();
+        stateful_kv_size = 0;
+    }
+
+    void clear_caches() {
+        std::lock_guard<std::mutex> lock(ctx_mutex);
+        clear_caches_locked();
+    }
+};
+
+enum ggml_status ov_graph_compute(struct ggml_cgraph * cgraph, ggml_backend_t backend);
+
+size_t checksum(const void * data, size_t size);
+
+bool save_ggml_tensor_data_to_txt(const ggml_tensor * tensor, const std::string & file_path);
+
+void print_input_tensor_info(const std::string & name, const ov::Tensor & tensor);
+
+void print_output_tensor_info(const std::string & name, const ov::Tensor & tensor, const void * output_dst);
+
+template <typename T>
+std::vector<T> pad_input(const T * data,
+                         size_t rows,
+                         size_t cols,
+                         size_t padded_rows,
+                         size_t padded_cols,
+                         T pad_value) {
+    std::vector<T> padded(padded_rows * padded_cols, pad_value);
+
+    for (size_t i = 0; i < std::min(rows, padded_rows); ++i) {
+        for (size_t j = 0; j < std::min(cols, padded_cols); ++j) {
+            padded[i * padded_cols + j] = data[i * cols + j];
+        }
+    }
+
+    return padded;
+}
+
+template <typename T>
+std::vector<T> pad_input(const ggml_tensor * tensor, size_t padded_rows, size_t padded_cols, T pad_value) {
+    return pad_input<T>(reinterpret_cast<const T *>(tensor->data),
+                        static_cast<size_t>(tensor->ne[1]),  // rows
+                        static_cast<size_t>(tensor->ne[0]),  // cols
+                        padded_rows, padded_cols, pad_value);
+}
+
+const ggml_tensor * get_inp_pos_tensor(struct ggml_cgraph * cgraph);
+
+int64_t get_inp_pos_n_tokens(struct ggml_cgraph * cgraph, const ggml_tensor * inp_pos);
+
+bool get_is_prefill(struct ggml_cgraph * cgraph, const ggml_tensor * inp_pos);
+
+bool is_naive(struct ggml_cgraph * cgraph);
+
+/**
+ * @brief Heuristically checks whether the given computation graph is a split-model fragment.
+ * @param cgraph Pointer to the GGML computation graph to analyze.
+ * @return true if the graph is identified as split; otherwise false.
+ */
+bool is_model_splitted(struct ggml_cgraph * cgraph);