| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
feat: support W4A8-QAT Co-authored-by: xusiyang<xusiyang2@huawei.com> # message auto-generated for no-merge-commit merge: !3664 merge master into master feat: support W4A8-QAT Created-by: weixin_44492126 Commit-by: xusiyang Merged-by: ascend-robot Description: ## What this PR does / why we need it? 本PR 新增了对 W4A8-QAT（权重MXFP4/激活 MXFP8 量化感知训练，仅限 MoE 场景），并量化粒度分别支持：32(标准MXFP4量化粒度)/128(deepseek-V4采用量化粒度） ## Does this PR introduce any user-facing change? Yes. 用户可以通过在训练命令中添加以下参数来启用W4A8功能，默认量化粒度32： --qat-scheme w4a8-moe-only --qat-quant-block-size 128 ## How was this patch tested? 添加 --qat-scheme w4a8-moe-only 参数启动量化粒度为32的W4A8-QAT训练，添加--qat-quant-block-size 128启动量化粒度为128的W4A8-QAT训练 实现方案和实验结果如下：https://wiki.huawei.com/domains/171785/wiki/358154/WIKI2026071511870516 昇腾 950DT上测试通过：基于减层DeepSeek-V4，分别训练100step对比BF16 loss误差：量化粒度为32时loss误差在0.4%，量化粒度为128时loss误差在0.79% See merge request: Ascend/MindSpeed!3664 | 1 个月前 | |
fix: add open component Copyright Co-authored-by: GuoHaifeng1999<guohaifeng12@huawei.com> # message auto-generated for no-merge-commit merge: !3611 merge master into master fix: add open component Copyright Created-by: GuoHaifeng1999 Commit-by: GuoHaifeng1999 Merged-by: ascend-robot Description: ## What this PR does / why we need it? add open component Copyright ## Does this PR introduce any user-facing change? nothing ## How was this patch tested? don't needs See merge request: Ascend/MindSpeed!3611 | 2 个月前 | |
feat: add mxfp8-32x32 for FSDP Co-authored-by: kyle_zhangchi<zhangchi158@huawei.com> # message auto-generated for no-merge-commit merge: !3534 merge feat_mxfp-32x32_for_fsdp into master feat: add mxfp8-32x32 for FSDP Created-by: kyle_zhangchi Commit-by: kyle_zhangchi Merged-by: ascend-robot Description: ## What this PR does / why we need it? 在fsdp框架下新增mxfp8-32x32量化算子，降低权重显存占用 ## Does this PR introduce any user-facing change? --model.quant_recipe_name新增mxfp8-32x32选项 ## How was this patch tested? 验证文档 https://wiki.huawei.com/domains/170864/wiki/367830/WIKI2026060611372464 RFC：https://gitcode.com/Ascend/MindSpeed/issues/175 See merge request: Ascend/MindSpeed!3534 | 2 个月前 | |
ut | 27 天前 | |
Distributed test framework | 2 年前 | |
fix【问题】use_custom_fsdp特性和dualpipev不兼容 Co-authored-by: lizhi<lizhi166@huawei.com> # message auto-generated for no-merge-commit merge: !3080 merge bug-master into master fix【问题】use_custom_fsdp特性和dualpipev不兼容 Created-by: liz_hw Commit-by: lizhi Merged-by: ascend-robot Description: fix【问题】use_custom_fsdp特性和dualpipev不兼容 See merge request: Ascend/MindSpeed!3080 | 9 个月前 | |
feature(ut): uses third-party verification for mc2 UT. Co-authored-by: Muu<koimuu@163.com> # message auto-generated for no-merge-commit merge: !3247 merge feature_mc2 into master feature(ut): uses third-party verification for mc2 UT. Created-by: Muuyo Commit-by: Muu Merged-by: ascend-robot Description: 采用三方验证 输入值域 [-5,5] 验证文档: https://wiki.huawei.com/domains/76578/wiki/233229/WIKI2026013110011445 See merge request: Ascend/MindSpeed!3247 | 7 个月前 |