main

ascend-robotfeat(FP8): Support W4A8 recipe

6134454a创建于 5 天前历史提交

文件	最后提交记录	最后更新时间
__init__.py	feat(FP8): Support W4A8 recipe Co-authored-by: mingzhenwang<wangmingzhen4@huawei.com> # message auto-generated for no-merge-commit merge: !73 merge w4a8-hybrid-linear into main feat(FP8): Support W4A8 recipe Created-by: mingzhenwang Commit-by: mingzhenwang Merged-by: ascend-robot Description: 1. deepseekv4 QAT方案，前向W4A8，反向W8A8 MXFP8 2. 区别点1，deepseekv4只在moe layer部分采用W4A8，TE作为低精基础能力，linear和grouped_linear均采用W4A8 3. 区别点2，deepseekv4采用量化为MXFP4，反量化回MXFP8，当前采用MXFP8@MXFP4伪量化算子代替该流程 See merge request: Ascend/TransformerEngineNPU!73	5 天前
float8_blockwise_tensor_storage.py	refactor(quantized_tensor): replace QuantizedTensorView with tensor tuples Co-authored-by: Muu<koimuu@163.com>	19 天前
float8_tensor_storage.py	refactor(fp8): fill fp8 vacancy Co-authored-by: Muu<koimuu@163.com>	2 个月前
mxfp4_tensor_storage.py	refactor(quantized_tensor): replace QuantizedTensorView with tensor tuples Co-authored-by: Muu<koimuu@163.com>	19 天前
mxfp8_tensor_storage.py	refactor(quantized_tensor): replace QuantizedTensorView with tensor tuples Co-authored-by: Muu<koimuu@163.com>	19 天前
w4a8_tensor_storage.py	feat(FP8): Support W4A8 recipe Co-authored-by: mingzhenwang<wangmingzhen4@huawei.com> # message auto-generated for no-merge-commit merge: !73 merge w4a8-hybrid-linear into main feat(FP8): Support W4A8 recipe Created-by: mingzhenwang Commit-by: mingzhenwang Merged-by: ascend-robot Description: 1. deepseekv4 QAT方案，前向W4A8，反向W8A8 MXFP8 2. 区别点1，deepseekv4只在moe layer部分采用W4A8，TE作为低精基础能力，linear和grouped_linear均采用W4A8 3. 区别点2，deepseekv4采用量化为MXFP4，反量化回MXFP8，当前采用MXFP8@MXFP4伪量化算子代替该流程 See merge request: Ascend/TransformerEngineNPU!73	5 天前