[CPU] Optimize small oc GEMM for Qwen3-next on CPU#12446
Merged
Fridge003 merged 14 commits intosgl-project:mainfrom Dec 4, 2025
Merged
[CPU] Optimize small oc GEMM for Qwen3-next on CPU#12446Fridge003 merged 14 commits intosgl-project:mainfrom
Fridge003 merged 14 commits intosgl-project:mainfrom
Conversation
mingfeima
requested changes
Nov 10, 2025
Collaborator
mingfeima
left a comment
There was a problem hiding this comment.
make as little change in the python level as possible,
make the changes happen in C++ level.
that's to say, we still use weight_packed_linear, but we use a different implementation for it under certain conditions.
Collaborator
|
@jianan-gu put more detailed descriptions in the PR message. |
Collaborator
|
one more thing, using F32 brgemm might not be a good idea from performance point of view, but this kernel doesn't contribute too much in the end to end profiling results, we can keep it as it is right now. |
Collaborator
Contributor
Author
Noted and added comments in codes. |
Contributor
Author
Contributor
Author
|
Test commit: 0b97584
|
Collaborator
|
Hi, could you please fix conflicts? thx! |
Contributor
Author
Thanks and have fixed. |
mingfeima
approved these changes
Dec 4, 2025
Collaborator
|
@jianan-gu rebase |
2 tasks
FlamingoPg
approved these changes
Dec 4, 2025
tom-jerr
pushed a commit
to tom-jerr/sglang
that referenced
this pull request
Dec 4, 2025
Co-authored-by: Zheng, Beilei <beilei.zheng@intel.com>
yingluosanqian
pushed a commit
to yingluosanqian/sglang
that referenced
this pull request
Dec 4, 2025
Co-authored-by: Zheng, Beilei <beilei.zheng@intel.com>
tonyluj
pushed a commit
to openanolis/sglang
that referenced
this pull request
Dec 5, 2025
Co-authored-by: Zheng, Beilei <beilei.zheng@intel.com>
tonyluj
pushed a commit
to openanolis/sglang
that referenced
this pull request
Dec 5, 2025
Co-authored-by: Zheng, Beilei <beilei.zheng@intel.com>
yuchengz816-bot
pushed a commit
to yuchengz816-bot/sglang
that referenced
this pull request
Dec 8, 2025
Co-authored-by: Zheng, Beilei <beilei.zheng@intel.com>
Kevin-XiongC
pushed a commit
to novitalabs/sglang
that referenced
this pull request
Dec 9, 2025
Co-authored-by: Zheng, Beilei <beilei.zheng@intel.com>
blzheng
added a commit
to blzheng/sglang
that referenced
this pull request
Jan 30, 2026
Co-authored-by: Zheng, Beilei <beilei.zheng@intel.com>
blzheng
added a commit
to blzheng/sglang
that referenced
this pull request
Feb 2, 2026
* [CPU] Optimize small oc GEMM for Qwen3-next on CPU (sgl-project#12446) Co-authored-by: Zheng, Beilei <beilei.zheng@intel.com> * port linear_gelu_linear kernel * apply linear_gelu_linear for TP=1 * fix numa memory bind * apply parallel partition patch --------- Co-authored-by: jianan-gu <jianan.gu@intel.com>
jianan-gu
added a commit
to jianan-gu/sglang
that referenced
this pull request
Feb 25, 2026
* [CPU] Optimize small oc GEMM for Qwen3-next on CPU (sgl-project#12446) Co-authored-by: Zheng, Beilei <beilei.zheng@intel.com> * port linear_gelu_linear kernel * apply linear_gelu_linear for TP=1 * fix numa memory bind * apply parallel partition patch --------- Co-authored-by: jianan-gu <jianan.gu@intel.com>
sywangyi
added a commit
to sywangyi/sglang
that referenced
this pull request
Feb 27, 2026
* port layernorm 3d * apply layernorm * support for bias * fix * intf fix * add support for CPU * fix tp=3/6 padding issue in encoder vision * fix tp=3/6 padding issue in qwen3-omni * refactor code * add mrope * change attention_mask shape to use flash attn * add kernel apply_rotary_pos_emb_cpu * replace nn.Linear with ReplicatedLinear * enable torch.compile * construct mask using query.dtype instead of bool on CPU * add fast path for sparse attention * fix double free segfault by wrong setting of BLOCK_M * improve extend kernel performance for long context length * update test_extend.py * update comment * fix topk softmax performance issue * port optimization for image preprocessor in Qwen2VLImageProcessorFast * apply optimization for image preprocessor * update docker file * optimize conv3d used in patch embedding * resolve conflict * apply optimized conv3d * apply optimization for flash_attn_varlen_func (sgl-project#19) * port optimization for flash_attn_varlen_func * apply flash_attn_varlen_func * remove contiguous before rope (sgl-project#20) * Revert "resolve conflict" This reverts commit 7622f6d. * fix after rebase * Update pyproject_cpu.toml * Update xeon.Dockerfile * minor fix after rebase * rope: add support for bf16 sincos (sgl-project#102) * format * Update xeon.Dockerfile * odd tp for cpu * Apply linear_gelu_linear and fix numa memory bind (sgl-project#22) * [CPU] Optimize small oc GEMM for Qwen3-next on CPU (sgl-project#12446) Co-authored-by: Zheng, Beilei <beilei.zheng@intel.com> * port linear_gelu_linear kernel * apply linear_gelu_linear for TP=1 * fix numa memory bind * apply parallel partition patch --------- Co-authored-by: jianan-gu <jianan.gu@intel.com> * Revert "Fix: test_vlm_offline_throughput output throughput (sgl-project#13279)" (sgl-project#101) This reverts commit 7ee3e36. * fix input dtype mismatch issue * apply optimized layernorm --------- Co-authored-by: Zheng, Beilei <beilei.zheng@intel.com> Co-authored-by: ZailiWang <zaili.wang@intel.com> Co-authored-by: mingfeima <mingfei.ma@intel.com> Co-authored-by: jianan-gu <jianan.gu@intel.com>
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 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 adds optimization for BF16 GEMMs with small oc shapes ( oc < 16), including:
weight_packed_linear， which is common for all GEMMs with oc < 16fused_linear_sigmoid_mul, which is specific for cases found in Qwen3-next (shared_expert_gate, where N = 1 and having post op sigmoid+mul)