PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
fix UT check for int zp (#745)
Weiwei1f79db9bd11d
diff --git a/auto_round/export/export_to_autoround/export_to_fp.py b/auto_round/export/export_to_autoround/export_to_fp.pyindex a3c724ad9..a845d2396 100644--- a/auto_round/export/export_to_autoround/export_to_fp.py+++ b/auto_round/export/export_to_autoround/export_to_fp.py@@ -29,7 +29,6 @@ from auto_round.utils import (check_start_with_block_name,check_to_quantized,filter_quantization_config,- get_block_names,get_module,is_mx_fp,is_nv_fp,diff --git a/auto_round/export/export_to_autoround/qlinear_fp.py b/auto_round/export/export_to_autoround/qlinear_fp.pyindex bd70a43ba..af188fda1 100644--- a/auto_round/export/export_to_autoround/qlinear_fp.py+++ b/auto_round/export/export_to_autoround/qlinear_fp.py@@ -38,7 +38,7 @@ import transformersfrom auto_round.data_type.mxfp import FP32_EXPONENT_BIAS, FP32_MIN_NORMALfrom auto_round.data_type.nvfp import cast_to_fp4, get_reciprocalfrom auto_round.data_type.utils import reshape_pad_tensor_by_group_size, revert_tensor_by_pad-from auto_round.utils import is_mx_fp, is_nv_fp, is_standard_fp+from auto_round.utils import is_mx_fp, is_nv_fp# from auto_round.utils import get_weight_compress_dtypelogger = getLogger(__name__)@@ -228,31 +228,3 @@ class QuantLinear(nn.Module):packed = (indices[:, 0] | (indices[:, 1] << 4)).to(torch.uint8)return packed.reshape(m, n // 2)-- # def get_fp_scale(self, scale_e8m0):- # E8M0_EXPONENT_BIAS = 127- # E8M0_EXPONENT_NAN_VAL = 255-- # scale_e8m0 = scale_e8m0.view(torch.uint8)- # s_offset = scale_e8m0.to(torch.int16) - E8M0_EXPONENT_BIAS- # # TODO(later): it would be nice if there was a way to do the 2^x operation- # # in PyTorch without creating a tensor of twos- # two = torch.full(s_offset.size(), 2.0, device=scale_e8m0.device)- # # pow(two, s_offset) can be out of range of floating point formats.- # # TODO(later): handle this for float16 if we decide to support float16- # # scales.- # s_fp = torch.pow(two, s_offset)-- # # If a block exponent was 255, set values of that block to NaN- # s_fp = torch.where(scale_e8m0 != E8M0_EXPONENT_NAN_VAL, s_fp, float("nan"))-- # return s_fp-- # def get_compressed_weight(self, tensor, bits, data_type):- # data_type = str(data_type)- # compress_dtype = None- # if self.is_mx:- # if bits == 4:-- # if self.is_nv:- # pass ## TODOdiff --git a/auto_round/export/export_to_itrex/export.py b/auto_round/export/export_to_itrex/export.pyindex ef64af77b..6fb33b010 100644--- a/auto_round/export/export_to_itrex/export.py+++ b/auto_round/export/export_to_itrex/export.py@@ -213,6 +213,8 @@ def pack_model(m = get_module(compressed_model, k)fp_weight = m.weight.datascale, zp = m.scale, m.zp+ if isinstance(zp, int | float):+ zp = torch.full_like(scale, zp)convert_dtype = scale_dtypeif not isinstance(scale, torch.Tensor):scale = torch.tensor(scale, dtype=convert_dtype)diff --git a/test/test_cpu/test_woq_linear.py b/test/test_cpu/test_woq_linear.pyindex fcdfeb27f..e077c7a21 100644--- a/test/test_cpu/test_woq_linear.py+++ b/test/test_cpu/test_woq_linear.py@@ -35,6 +35,8 @@ class TestWeightOnlyLinear:origin_shape = weight.shapeweight = weight.reshape(-1, group_size)qdq, scale, zp = quant_tensor_sym(weight, -1)+ if isinstance(zp, int | float):+ zp = torch.full_like(scale, zp)int_weight = qdq.div(scale).add(zp).clamp(0, 2 ** (bits) - 1).to(torch.int32).reshape(origin_shape)scale = scale.reshape(origin_shape[0], -1)if isinstance(zp, torch.Tensor):