26.0.0_core_r0.12.1

ascend-robot[fix]:solve the problem of chunk_bwd_dqkwg triton's time degradation.

d1f74d4a创建于 4月9日历史提交

文件	最后提交记录	最后更新时间
distributed	feat：【FSDP backend】Change expert parallel to support combined experts weights. Co-authored-by: JialiZheng<jializheng@huawei.com> # message auto-generated for no-merge-commit merge: !3147 merge master into master feat：【FSDP backend】Change expert parallel to support combined experts weights. Created-by: JialiZheng1 Commit-by: JialiZheng Merged-by: ascend-robot Description: 修改 EP 为支持专家合并权重 eager dispatcher 精度验证: http://image.huawei.com/tiny-lts/v1/images/mdstorm/729781f03639b9a5e430e986c51fc6ee_665x772.png fused dispatcher 精度验证: http://image.huawei.com/tiny-lts/v1/images/mdstorm/52a71f80ec498d1136da7bb83b3b585d_665x770.png See merge request: Ascend/MindSpeed!3147	4 个月前
memory	【feature】 adaptor swap optimizer for fsdp Co-authored-by: libaokui<libaokui@huawei.com> # message auto-generated for no-merge-commit merge: !2995 merge swap_optimizer into master 【feature】 adaptor swap optimizer for fsdp Created-by: libaokui Commit-by: libaokui Merged-by: ascend-robot Description: adaptor for swap optimizer for fsdp ## 自验文档 https://wiki.huawei.com/domains/73366/wiki/264623/WIKI202510318827632 See merge request: Ascend/MindSpeed!2995	5 个月前
ops	[fix]:solve the problem of chunk_bwd_dqkwg triton's time degradation. Co-authored-by: LinShua<707894133@qq.com> # message auto-generated for no-merge-commit merge: !3379 merge 26.0.0_core_r0.12.1_dqkwg_triton_time into 26.0.0_core_r0.12.1 [fix]:solve the problem of chunk_bwd_dqkwg triton's time degradation. Created-by: LinShua Commit-by: LinShua Merged-by: ascend-robot Description: What this PR does / why we need it? 解决chunk_bwd_dqkwg 算子定长性能优化带来的变长性能劣化问题: chunk_bwd_kernel_dqkwg 定长： 原逻辑：26158us 优化后：18855us 变长： 原逻辑：33450us 优化后：89597us 当前PR： 定长： 18769us 变长： 33509us Does this PR introduce any user-facing change? NA How was this patch tested? 见PR：[fix]:solve the UT of GDN triton. See merge request: Ascend/MindSpeed!3379	1 个月前
utils	【MindSpeed Lite】Support FSDP+EP+EFSDP Co-authored-by: JialiZheng<jializheng@huawei.com> # message auto-generated for no-merge-commit merge: !3073 merge master into master 【MindSpeed Lite】Support FSDP+EP+EFSDP Created-by: JialiZheng1 Commit-by: JialiZheng Merged-by: ascend-robot Description: support FSDP+EP+EFSDP precision verification: http://image.huawei.com/tiny-lts/v1/images/mdstorm/e9e4418f5cf63119827a5ecd7fedc66e_661x778.png See merge request: Ascend/MindSpeed!3073	5 个月前
__init__.py	Add MindSeed Lite Co-authored-by: JialiZheng<jializheng@huawei.com> # message auto-generated for no-merge-commit merge: !2978 merge lite into master Add MindSeed Lite Created-by: JialiZheng1 Commit-by: JialiZheng Merged-by: ascend-robot Description: Add MindSeed Lite See merge request: Ascend/MindSpeed!2978	6 个月前
mindspeed_lite.py	feat：【FSDP backend】Change expert parallel to support combined experts weights. Co-authored-by: JialiZheng<jializheng@huawei.com> # message auto-generated for no-merge-commit merge: !3147 merge master into master feat：【FSDP backend】Change expert parallel to support combined experts weights. Created-by: JialiZheng1 Commit-by: JialiZheng Merged-by: ascend-robot Description: 修改 EP 为支持专家合并权重 eager dispatcher 精度验证: http://image.huawei.com/tiny-lts/v1/images/mdstorm/729781f03639b9a5e430e986c51fc6ee_665x772.png fused dispatcher 精度验证: http://image.huawei.com/tiny-lts/v1/images/mdstorm/52a71f80ec498d1136da7bb83b3b585d_665x770.png See merge request: Ascend/MindSpeed!3147	4 个月前
mindspeed_lite_config.py	【MindSpeed Lite】Support FSDP+EP+EFSDP Co-authored-by: JialiZheng<jializheng@huawei.com> # message auto-generated for no-merge-commit merge: !3073 merge master into master 【MindSpeed Lite】Support FSDP+EP+EFSDP Created-by: JialiZheng1 Commit-by: JialiZheng Merged-by: ascend-robot Description: support FSDP+EP+EFSDP precision verification: http://image.huawei.com/tiny-lts/v1/images/mdstorm/e9e4418f5cf63119827a5ecd7fedc66e_661x778.png See merge request: Ascend/MindSpeed!3073	5 个月前