VhxddlmZmZddlZddlmZddlmcmZddl m Z dgZ Gddeje Z y))AnyOptionalN)ReferenceQuantizedModuleLinearceZdZdZdZ ddedededeejdeejd ee e e fd dffd Zd e fd Zd ej d ej fdZedej(d e e e fd dfdZxZS)raNA reference quantized linear module that fits into the FX Graph Mode Quantization workflow activation will be floating point Tensor, we will store floating point weight as well in the module, but in forward we'll quantize and dequantize the weight before running the floating point functional linear operator. TN in_features out_featuresbias_devicedtypeweight_qparamsreturncPt|||||||j||y)N)super__init___init_weight_qparams)selfr r r r r r __class__s ^/home/dcms/DCMS/lib/python3.12/site-packages/torch/ao/nn/quantized/reference/modules/linear.pyrzLinear.__init__s* lE65I !!.&9cy)NzQuantizedLinear(Reference))rs r _get_namezLinear._get_name$s+rxch|j}tj|||j}|S)aR we have: w(float) -- quant - dequant x(float) ------------- F.linear --- In the full model, we will see w(float) -- quant - *dequant x -- quant --- *dequant -- *F.linear --- *quant - dequant and the backend should be able to fuse the ops with `*` into a quantized linear ) get_weightFlinearbias)rrweight_quant_dequantresults rforwardzLinear.forward's- $0!1499= r float_linearct|j|j|jdu|jj |jj |}tjj|jj|_|jr@s3 + *8RYY08r