[0.9.1][Perf]Remove NZ of kv_b_proj in Deepseek MLA. #1872
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Signed-off-by: whx-sjtu <2952154980@qq.com>
|
#1131 This PR does this |
NNUCJ
pushed a commit
to NNUCJ/vllm-ascend
that referenced
this pull request
Jul 21, 2025
This PR removes NZ transformation of weights of kv_b_proj. This is because we find that this matmul weight is not quantized and will fall back to ND calculation in runtime (because currently float bmm nz is not supported in torchair graph), which causes two redundant transData operations (trans weight from NZ back to ND). Removing these two operations will provide an optimization of about 40us per layer. Signed-off-by: whx-sjtu <2952154980@qq.com>
845473182
added a commit
to 845473182/vllm-ascend
that referenced
this pull request
Jul 23, 2025
* br_eplb_into_v091: (29 commits) add eplb design doc merge update in eplb branch dynamic eplb [0.9.1][Perf] Use fused ops npu_top_k_top_p (vllm-project#1920) [0.9.1][PD][Perf] Avoid performing cpu all_reduce in disaggregated-prefill scenario. (vllm-project#1644) [0.9.1][BugFix] Fix bug in path_decorator when engine v0 (vllm-project#1919) [0.9.1][Perf] apply npu_moe_gating_top_k_softmax for moe (vllm-project#1902) [0.9.1][bugfix] W4A8 does not currently support apply_mlp_decode (vllm-project#1910) [0.9.1][CI] Pin vllm version to v0.9.1 to make mypy check passed (vllm-project#1904) [0.9.1][Dist][Bugfix] Fix mc2 process group to resolve self.cpu_group is None (vllm-project#1831) [0.9.1][Perf]Remove NZ of kv_b_proj in Deepseek MLA. (vllm-project#1872) [0.9.1][bugfix] V0.9.1 fix rope accruracy bug for deepseek model (vllm-project#1887) [0.9.1] Fix wheel glibc version incompatibility (vllm-project#1808) [BUGFIX][v0.9.1] repair moe error when set multistream. (vllm-project#1882) [BUGFIX][v0.9.1] ep_group is not equal to word_size in some cases. (vllm-project#1862) [BUGFIX][v0.9.1] fix enable_multistream_moe bug when DBO is enabled (… (vllm-project#1827) [0.9.1]optmize rope in qwen2 (vllm-project#1782) [BugFix] Fix flashcomm_v1 when engine v0 (vllm-project#1859) [BugFix] Fix decorator patch (vllm-project#1858) [0.9.1][Fix] Fix DeepSeek OOM issue in extreme `--gpu-memory-utilization` scenario (vllm-project#1829) ...
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR removes NZ transformation of weights of kv_b_proj. This is because we find that this matmul weight is not quantized and will fall back to ND calculation in runtime (because currently float bmm nz is not supported in torchair graph), which causes two redundant transData operations (trans weight from NZ back to ND). Removing these two operations will provide an optimization of about 40us per layer.