已合并
refactor: remove deepseek4 redundant Triton paths and align CSA/MHC CLI #4871
丁子叉创建于 16 天前
refactor: remove deepseek4 redundant Triton paths and align CSA/MHC CLI #4871
已合并
Pull Request已成功合入, 合并人@ascend-robot
(感谢 丁子叉 的贡献)atomgit-bot
16 天前 评论:
16 天前 评论:
🤖 正在生成合并请求摘要,请稍候…


ascend-robot
16 天前 评论:
16 天前 评论:
atomgit-bot
16 天前 评论:
16 天前 评论:
🤖 AI 代码检视正在进行中,请稍候…


16 天前 添加了label:stat/needs-squash
此处折叠了255条消息 查看更多
guozhihua2
13 天前 评论:
13 天前 评论:
/approve


13 天前 添加了label:approvedlgtm
13 天前 合入了pull request
ascend-robot
13 天前 评论:
13 天前 评论:
Pull Request 已合并或已关闭。
If you want to solve this problem, you can click here to do it in the FAQs.


ascend-robot
13 天前 评论:
13 天前 评论:
Pull Request 已合并或已关闭。
If you want to solve this problem, you can click here to do it in the FAQs.


What this PR does / why we need it?
This PR simplifies the DeepSeek-V4 CSA and MHC implementations after the Ascend fused paths became stable.
Main changes:
pre_onlykernel required by the MHC head stage.opsdirectories.Does this PR introduce any user-facing change?
Yes. DeepSeek-V4 CLI configurations need to be updated as follows:
--use-ascend-mhc--use-fused-mhc--g2-window-size 128--sliding-window-size 128--rope-head-dim 64--qk-pos-emb-head-dim 64--norm-eps 1e-6--norm-epsilon 1e-6--position-embedding-type g2--position-embedding-type deepseek4--use-g2-indexer-loss--indexer-loss-coeff 1.0; use0for inference--use-triton-sfa--use-sparse-flash-attn--use-g2-attention--use-triton-sinkhorn--use-triton-rmsnorm-without-weight--mhc-recompute--use-fused-rotary-pos-emb--use-rotary-position-embeddings--use-triton-mhcpre_onlykernel--use-fused-lightning-indexerand--use-fused-lightning-indexer-lossRecommended DeepSeek-V4 position embedding configuration:
Do not combine it with:
How was this patch tested?
bash -nagainst all affected DeepSeek-V4 shell scripts.git diff --check../gitleaksbinary was unavailable.try/exceptlogic.