main

ascend-robotfix: set input_quantizer usage explicitly so a forward-only pass does not pollute the next training forward

b438145d创建于 11 小时前历史提交

文件	最后提交记录	最后更新时间
__init__.py	refactor(distributed): Refactor distributed tools and simplify linear parameters Co-authored-by: Muu<koimuu@163.com>	1 个月前
_common.py	refactor: Adapted to the latest TE linear layer. Co-authored-by: Muu<koimuu@163.com>	1 个月前
base.py	feat(FP4): Support W4A4-MXFP4 Co-authored-by: mingzhenwang<wangmingzhen4@huawei.com>	15 天前
fp8_padding.py	refactor(distributed): Refactor distributed tools and simplify linear parameters Co-authored-by: Muu<koimuu@163.com>	1 个月前
fp8_unpadding.py	feat(fp8): finalize FP8 core implementation and fixes Co-authored-by: Muu<koimuu@163.com>	2 个月前
grouped_linear.py	refactor(grouped_linear,gemm,fp8): overhaul grouped matmul with unified dispatch, NPU version check, and FP8 quantization cleanup Co-authored-by: Muu<koimuu@163.com>	17 天前
layernorm.py	linear fp8 refactor Co-authored-by: clc2025<chenlucong@huawei.com>	1 个月前
layernorm_linear.py	fix: set input_quantizer usage explicitly so a forward-only pass does not pollute the next training forward Co-authored-by: Bruce-rl-hw<okwsl201210@gmail.com> # message auto-generated for no-merge-commit merge: !95 merge fix/mxfp8-thd-columnwise into main fix: set input_quantizer usage explicitly so a forward-only pass does not pollute the next training forward Created-by: Bruce-rl-hw Commit-by: Bruce-rl-hw Merged-by: ascend-robot Description: Fix mxfp8 training in LayerNormLinear by setting the input quantizer's usage explicitly every forward (rowwise=True, columnwise=backward_needs_input), so a prior forward-only pass (e.g. an RL log-prob recompute) can't leave columnwise=False and make the backward weight-gradient GEMM crash with "Cannot access storage of UndefinedTensorImpl". Verified on qwen3-0.6B/Ascend 950 DT See merge request: Ascend/TransformerEngineNPU!95	11 小时前
layernorm_mlp.py	fix ut Co-authored-by: clc2025<chenlucong@huawei.com>	28 天前
linear.py	refactor(grouped_linear,gemm,fp8): overhaul grouped matmul with unified dispatch, NPU version check, and FP8 quantization cleanup Co-authored-by: Muu<koimuu@163.com>	17 天前
performance_grouped_linear_impl.py	refactor(quantized_tensor): replace QuantizedTensorView with tensor tuples Co-authored-by: Muu<koimuu@163.com>	12 天前
rmsnorm.py	feat: basic.rmsnorm precision and optimize ut of ops/act Co-authored-by: GuoHaifeng1999<guohaifeng12@huawei.com>	1 个月前