112 lines
2.7 KiB
Python
112 lines
2.7 KiB
Python
"""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_LR_FLOOR,
|
|
DEFAULT_LR_PATIENCE,
|
|
DEFAULT_LR_THRESHOLD,
|
|
DEFAULT_SCALE_OPTIMIZATION,
|
|
DEFAULT_SCALE_REFINEMENT_ROUNDS,
|
|
FP4_LEVELS,
|
|
FP4_LEVELS_SORTED,
|
|
FP4_LEVEL_MAP,
|
|
FP4_TABLE,
|
|
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,
|
|
optimize_block_scales,
|
|
optimize_block_scales_torch,
|
|
pack_nvfp4,
|
|
preserve_scale_tensor,
|
|
quant_profile,
|
|
quant_settings,
|
|
quantize_and_pack,
|
|
refine_codes_for_scale,
|
|
refine_codes_for_scale_torch,
|
|
report_reconstruction,
|
|
require_quant_settings,
|
|
resolve_optimizer,
|
|
scale_bake_notice,
|
|
ue4m3_bits,
|
|
ue4m3_bits_torch,
|
|
coordinate_descent_refine,
|
|
coordinate_descent_refine_torch,
|
|
sweep_per_tensor_scale,
|
|
sweep_per_tensor_scale_torch,
|
|
get_ue4m3_lut,
|
|
get_fp4_torch_constants,
|
|
)
|
|
|
|
__all__ = [
|
|
"CTQ_NUM_ITER",
|
|
"DEFAULT_CONVERT_TO_QUANT",
|
|
"DEFAULT_LLAMA_CPP",
|
|
"DEFAULT_EARLY_STOP_FLOOR_STALL",
|
|
"DEFAULT_TARGET_LOSS",
|
|
"DEFAULT_EARLY_STOP_STALL",
|
|
"DEFAULT_LR_FLOOR",
|
|
"DEFAULT_LR_PATIENCE",
|
|
"DEFAULT_LR_THRESHOLD",
|
|
"DEFAULT_SCALE_OPTIMIZATION",
|
|
"DEFAULT_SCALE_REFINEMENT_ROUNDS",
|
|
"FP4_LEVELS",
|
|
"FP4_LEVELS_SORTED",
|
|
"FP4_LEVEL_MAP",
|
|
"FP4_TABLE",
|
|
"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",
|
|
"optimize_block_scales",
|
|
"optimize_block_scales_torch",
|
|
"pack_nvfp4",
|
|
"preserve_scale_tensor",
|
|
"quant_profile",
|
|
"quant_settings",
|
|
"quantize_and_pack",
|
|
"refine_codes_for_scale",
|
|
"refine_codes_for_scale_torch",
|
|
"coordinate_descent_refine",
|
|
"coordinate_descent_refine_torch",
|
|
"sweep_per_tensor_scale",
|
|
"sweep_per_tensor_scale_torch",
|
|
"get_ue4m3_lut",
|
|
"get_fp4_torch_constants",
|
|
"report_reconstruction",
|
|
"require_quant_settings",
|
|
"resolve_optimizer",
|
|
"scale_bake_notice",
|
|
"ue4m3_bits",
|
|
"ue4m3_bits_torch",
|
|
]
|