initial commit

nvfp4gguf/__init__.py

@@ -0,0 +1,75 @@
+"""Shared NVFP4 quantization used by the GGUF scripts in this project."""
+
+from .quant import (
+    CTQ_NUM_ITER,
+    DEFAULT_CONVERT_TO_QUANT,
+    DEFAULT_LLAMA_CPP,
+    DEFAULT_EARLY_STOP_FLOOR_STALL,
+    DEFAULT_TARGET_LOSS,
+    DEFAULT_EARLY_STOP_STALL,
+    DEFAULT_SCALE_OPTIMIZATION,
+    DEFAULT_SCALE_REFINEMENT_ROUNDS,
+    QUANT_VERSION,
+    ShapeView,
+    TensorQuantizer,
+    add_quant_arguments,
+    bind_exclude,
+    ctq_skips_key,
+    eligible_weight,
+    ensure_cuda_interpreter,
+    import_gguf,
+    load_gguf_modules,
+    load_learned_converter,
+    looks_like_spark,
+    check_scale_policy,
+    must_bake_nvfp4_scale,
+    mtp_layer_tensor,
+    bake_packed_nvfp4,
+    pack_nvfp4,
+    preserve_scale_tensor,
+    quant_profile,
+    quant_settings,
+    quantize_and_pack,
+    report_reconstruction,
+    require_quant_settings,
+    resolve_optimizer,
+    scale_bake_notice,
+    ue4m3_bits,
+)
+
+__all__ = [
+    "CTQ_NUM_ITER",
+    "DEFAULT_CONVERT_TO_QUANT",
+    "DEFAULT_LLAMA_CPP",
+    "DEFAULT_EARLY_STOP_FLOOR_STALL",
+    "DEFAULT_TARGET_LOSS",
+    "DEFAULT_EARLY_STOP_STALL",
+    "DEFAULT_SCALE_OPTIMIZATION",
+    "DEFAULT_SCALE_REFINEMENT_ROUNDS",
+    "QUANT_VERSION",
+    "ShapeView",
+    "TensorQuantizer",
+    "add_quant_arguments",
+    "bind_exclude",
+    "ctq_skips_key",
+    "eligible_weight",
+    "ensure_cuda_interpreter",
+    "import_gguf",
+    "load_gguf_modules",
+    "load_learned_converter",
+    "looks_like_spark",
+    "check_scale_policy",
+    "must_bake_nvfp4_scale",
+    "mtp_layer_tensor",
+    "bake_packed_nvfp4",
+    "pack_nvfp4",
+    "preserve_scale_tensor",
+    "quant_profile",
+    "quant_settings",
+    "quantize_and_pack",
+    "report_reconstruction",
+    "require_quant_settings",
+    "resolve_optimizer",
+    "scale_bake_notice",
+    "ue4m3_bits",
+]