o *i @sddlmZmZddlZddlmZmZddlmZm Z m Z ddl m Z ddl mZdgZ ddeed eed eed efd d Z  ddejjd eedeefddZ  ddejjd eedeed eefddZdS))ListOptionalN)CompressionFormatSparsityStructure)QuantizationArgsQuantizationStrategyQuantizationType)is_module_quantized)logger,infer_and_set_per_module_quantization_format input_args weight_argssparsity_structurereturncCst|tjk}|duo|du}|jdkr&|jtjjkr&|jdkr#tj Stj S|rL|jdvo3|jtj jk}|s9tj S|rI|j tjjtjjfvrItjStjS|jtjjkr[|jdkr[tjS|jtj jkretjStj S)a Using the weight and input quantization args as well as an optional sparsity structure, determine the compression format that should be applied to a given module :param input_args: input quantization parameters :param weight_args: weight quantization parameters :param sparsity_structure: optional (global) modle sparsity structure :return CompresssionFormat for the module N )rr)rTWO_FOURnum_bitstyperFLOATvalue group_sizermxfp4_pack_quantizednvfp4_pack_quantizedINTnaive_quantizedstrategyrCHANNELGROUP marlin_24pack_quantizedfloat_quantized int_quantized)r r ris_24_structureis_weight_only is_valid_packr']/home/ubuntu/veenaModal/venv/lib/python3.10/site-packages/compressed_tensors/config/format.py_get_quant_compression_formats2     r)modulequantization_formatcCs|jj}|jj}|durdSt|||}|dur(||jkr"td||j_dS|jjdur>|jj|jkrsJ   9 /