[perf][sgl-kernel] extend cutlass_mla_decode to support num_head < 128 #6929
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
There was a problem hiding this comment.
Hello @Alcanderian, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!
Summary of Changes
Hello! Gemini-code-assist has reviewed this pull request. The main purpose of this PR is to enhance the cutlass_mla_decode kernel to properly support attention head counts (num_head) that are less than 128. Previously, this required workarounds. The changes involve modifying the underlying CUDA kernel to handle different page sizes more effectively, updating the Python wrapper to manage input tensors for smaller head counts by padding, adding a new benchmark script to measure performance across various configurations including different head counts, and extending the unit tests to ensure correctness for the newly supported range of head counts. According to the author's benchmarks, these changes result in performance improvements of up to 1.8x for cases where the page_size is 128.
Highlights
- Extended Head Count Support: The
cutlass_mla_decodekernel now supports a number of attention heads (num_head) less than or equal to 128, removing a previous limitation. - New Performance Benchmark: A new Triton-based benchmark script (
bench_cutlass_mla.py) has been added to measure the performance (GB/s) of thecutlass_mla_decodekernel across various batch sizes, sequence lengths, block sizes, and importantly, different numbers of heads (128, 64, 32, 16). - Kernel Adaptation for Page Size: The core CUDA kernel (
cutlass_mla_kernel.cu) has been updated to use a template parameter (IsPaged128) to conditionally adjust its behavior (specifically, thekIsCpAsyncsetting) based on whether the page size is 128. This allows for better optimization across different memory layouts. - Python Wrapper Input Handling: The Python interface (
attention.py) now checks the number of heads (H) and pads the input query tensor to 128 heads ifH < 128before calling the CUDA kernel. The output is then sliced back to the originalHheads. - Expanded Unit Tests: Unit tests (
test_cutlass_mla.py) have been updated to include parameterization for different numbers of heads (16, 32, 64, 128) to verify the correctness of the kernel across the extended support range. - Reported Performance Improvement: The author's benchmark results indicate performance gains, particularly for
page_size == 128configurations with varying head counts, showing up to a 1.8x improvement.
Changelog
Click here to see the changelog
- sgl-kernel/benchmark/bench_cutlass_mla.py
- Added a new Triton benchmark script for
cutlass_mla_decode(Lines 1-115). - Configured the benchmark to test different numbers of heads (128, 64, 32, 16) (Lines 21-31).
- Included logic to pad
block_numfor CUTLASS tile alignment based onblock_size(Lines 56-59). - Calculates and reports performance in GB/s (Lines 83-91).
- Added a new Triton benchmark script for
- sgl-kernel/csrc/attention/cutlass_mla_kernel.cu
- Added
IsPaged128template parameter toMlaSm100struct (Line 58). - Made
kIsCpAsyncconditional on!IsPaged128inFmhaKerneldefinition (Line 86). - Added
IsPaged128template parameter torunMlafunction signature (Line 168). - Instantiated
MlaSm100Typewith theIsPaged128parameter inrunMla(Line 177). - Added logic in
cutlass_mla_decodeto checkpage_sizeand callrunMlawith the appropriateIsPaged128value (Lines 198-215). - Fixed
MlaSm100Typeinstantiation incutlass_mla_get_workspace_sizeto useIsPaged128=true(Line 221).
- Added
- sgl-kernel/python/sgl_kernel/attention.py
- Changed the assertion for the number of heads (
H) from== 128to<= 128(Line 76). - Added logic to pad the input query tensor (
q_nope_and_q_pe) to 128 heads whenH < 128(Lines 77-80). - Changed the output tensor creation to always be size
(B_q, 128, D_latent)(Line 104). - Sliced the output tensor back to the original number of heads (
H) and made it contiguous before returning (Line 109).
- Changed the assertion for the number of heads (
- sgl-kernel/tests/test_cutlass_mla.py
- Added parameterization for
num_headswith values [16, 32, 64, 128] to thetest_cutlass_mla_decodefunction (Line 43). - Updated the
test_cutlass_mla_decodefunction signature to accept thenum_headsparameter (Lines 45-50). - Set the number of query heads (
h_q) to the parameterizednum_headsvalue (Line 57).
- Added parameterization for
Using Gemini Code Assist
The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.
Invoking Gemini
You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.
| Feature | Command | Description |
|---|---|---|
| Code Review | /gemini review |
Performs a code review for the current pull request in its current state. |
| Pull Request Summary | /gemini summary |
Provides a summary of the current pull request in its current state. |
| Comment | @gemini-code-assist | Responds in comments when explicitly tagged, both in issue comments and review comments. |
| Help | /gemini help |
Displays a list of available commands. |
Customization
To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.
Limitations & Feedback
Gemini Code Assist is currently in preview and may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments to provide feedback.
You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.
Footnotes
-
Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩
Alcanderian
commented
Jun 6, 2025
There was a problem hiding this comment.
Code Review
This pull request introduces valuable enhancements to cutlass_mla_decode by enabling support for num_head < 128 through Python-level padding and optimizing performance. The benchmark results are promising, and the addition of a benchmark script is valuable.
I have a few questions and suggestions to improve the robustness and clarity of these changes.
Summary of Findings
- Kernel Internals (CUDA): The rationale for
kIsCpAsync = !IsPaged128(synchronous copy for page_size=128, async otherwise) needs clarification for better understanding of the performance optimization. - Workspace Calculation (CUDA): It's important to confirm that hardcoding
IsPaged128 = trueincutlass_mla_get_workspace_sizeis safe and doesn't lead to incorrect workspace allocation ifkIsCpAsync(which depends onIsPaged128) affects workspace requirements. - Code Maintainability (Benchmark Script): The benchmark script uses magic numbers for dimensions (
d,dv) and a fragileif/elifchain for determining head counts, which could be improved for robustness and readability. - Code Clarity (Python Wrapper): The magic number
128for padding head dimensions inattention.pyshould ideally be a named constant to improve maintainability.
Merge Readiness
The pull request introduces significant improvements. However, due to the high-severity questions regarding the CUDA kernel's workspace size calculation and the rationale behind the kIsCpAsync logic, I recommend addressing these points before merging. The medium-severity suggestions for maintainability would also be good to consider. As a reviewer, I am not authorized to approve the pull request; please ensure further review and approval from other maintainers after addressing the feedback.
|
Is the |
2 tasks
yes |
Nice! I feel we can remove the limitation of page_size=128 for cutlass mla backend |
5 tasks
zhyncs
approved these changes
Jun 9, 2025
jianan-gu
pushed a commit
to jianan-gu/sglang
that referenced
this pull request
Jun 12, 2025
walker-ai
pushed a commit
to walker-ai/sglang
that referenced
this pull request
Jul 8, 2025
Merge branch 'sgl_20250610_sync_tag047 of git@code.alipay.com:Theta/SGLang.git into main https://code.alipay.com/Theta/SGLang/pull_requests/52 Reviewed-by: 剑川 <jianchuan.gys@antgroup.com> * [Bugfix] Fix slice operation when chunk size mismatch (sgl-project#6697) * [Bugfix] Fix ChatCompletion endpoint of mini_lb when stream is set (sgl-project#6703) * [CI] Fix setup of disaggregation with different tp (sgl-project#6706) * [PD] Remove Unnecessary Exception Handling for FastQueue.get() (sgl-project#6712) * Fuse routed_scaling_factor in DeepSeek (sgl-project#6710) * Overlap two kernels in DeepSeek with communication (sgl-project#6711) * Minor refactor two-batch overlap (sgl-project#6682) * Speed up when having padding tokens two-batch overlap (sgl-project#6668) * [Feature] Support Flashinfer fp8 blockwise GEMM kernel on Blackwell (sgl-project#6479) * Fix LoRA bench (sgl-project#6719) * temp * Fix PP for Qwen3 MoE (sgl-project#6709) * [feat] triton kernel for get_last_loc (sgl-project#6676) * [fix] more mem for draft_extend cuda_graph (sgl-project#6726) * [PD] bug fix: Update status if nixl receiver send a a dummy req. (sgl-project#6720) * Tune memory arguments on B200 (sgl-project#6718) * Add DeepSeek-R1-0528 function call chat template (sgl-project#6725) * refactor(tool call): Fix BaseFormatDetector tool_index issue and refactor `parse_streaming_increment` (sgl-project#6715) * Add draft extend CUDA graph for Triton backend (sgl-project#6705) * refactor apply_w8a8_block_fp8_linear in fp (sgl-project#6545) * [PD] Support completion endpoint (sgl-project#6729) * PD Rust LB (PO2) (sgl-project#6437) * Super tiny enable sole usage of expert distribution metrics and update doc (sgl-project#6680) * Support picking variants of EPLB algorithms (sgl-project#6728) * Support tuning DeepEP configs (sgl-project#6742) * [test] add ut and bm for get_last_loc (sgl-project#6746) * Fix mem_fraction_static for AMD CI (sgl-project#6748) * [fix][RL] Fix DeepSeekV3ForCausalLM.post_load_weights for multiple update weight (sgl-project#6265) * Improve EPLB logical to physical dispatch map (sgl-project#6727) * Update DeepSeek-R1-0528 function call chat template (sgl-project#6765) * [PD] Optimize time out logic and add env var doc for mooncake (sgl-project#6761) * Fix aiohttp 'Chunk too big' in bench_serving (sgl-project#6737) * Support sliding window in triton backend (sgl-project#6509) * Fix shared experts fusion error (sgl-project#6289) * Fix one bug in the grouped-gemm triton kernel (sgl-project#6772) * update llama4 chat template and pythonic parser (sgl-project#6679) * feat(tool call): Enhance Llama32Detector for improved JSON parsing in non-stream (sgl-project#6784) * Support token-level quantization for EP MoE (sgl-project#6782) * Temporarily lower mmlu threshold for triton sliding window backend (sgl-project#6785) * ci: relax test_function_call_required (sgl-project#6786) * Add intel_amx backend for Radix Attention for CPU (sgl-project#6408) * Fix incorrect LoRA weight loading for fused gate_up_proj (sgl-project#6734) * fix(PD-disaggregation): Can not get local ip (sgl-project#6792) * [FIX] mmmu bench serving result display error (sgl-project#6525) (sgl-project#6791) * Bump torch to 2.7.0 (sgl-project#6788) * chore: bump sgl-kernel v0.1.5 (sgl-project#6794) * Improve profiler and integrate profiler in bench_one_batch_server (sgl-project#6787) * chore: upgrade sgl-kernel v0.1.5 (sgl-project#6795) * [Minor] Always append newline after image token when parsing chat message (sgl-project#6797) * Update CI tests for Llama4 models (sgl-project#6421) * [Feat] Enable PDL automatically on Hopper architecture (sgl-project#5981) * chore: update blackwell docker (sgl-project#6800) * misc: cache is_hopper_arch (sgl-project#6799) * Remove contiguous before Flashinfer groupwise fp8 gemm (sgl-project#6804) * Correctly abort the failed grammar requests & Improve the handling of abort (sgl-project#6803) * [EP] Add cuda kernel for moe_ep_pre_reorder (sgl-project#6699) * Add draft extend CUDA graph for flashinfer backend (sgl-project#6805) * Refactor CustomOp to avoid confusing bugs (sgl-project#5382) * Tiny log prefill time (sgl-project#6780) * Tiny fix EPLB assertion about rebalancing period and recorder window size (sgl-project#6813) * Add simple utility to dump tensors for debugging (sgl-project#6815) * Fix profiles do not have consistent names (sgl-project#6811) * Speed up rebalancing when using non-static dispatch algorithms (sgl-project#6812) * [1/2] Add Kernel support for Cutlass based Fused FP4 MoE (sgl-project#6093) * [Router] Fix k8s Service Discovery (sgl-project#6766) * Add CPU optimized kernels for topk and rope fusions (sgl-project#6456) * fix new_page_count_next_decode (sgl-project#6671) * Fix wrong weight reference in dynamic EPLB (sgl-project#6818) * Minor add metrics to expert location updater (sgl-project#6816) * [Refactor] Rename `n_share_experts_fusion` as `num_fused_shared_experts` (sgl-project#6735) * [FEAT] Add transformers backend support (sgl-project#5929) * [fix] recover auto-dispatch for rmsnorm and rope (sgl-project#6745) * fix ep_moe_reorder kernel bugs (sgl-project#6858) * [Refactor] Multimodal data processing for VLM (sgl-project#6659) * Decoder-only Scoring API (sgl-project#6460) * feat: add dp-rank to KV events (sgl-project#6852) * Set `num_fused_shared_experts` as `num_shared_experts` when shared_experts fusion is not disabled (sgl-project#6736) * Fix one missing arg in DeepEP (sgl-project#6878) * Support LoRA in TestOpenAIVisionServer and fix fused kv_proj loading bug. (sgl-project#6861) * support 1 shot allreduce in 1-node and 2-node using mscclpp (sgl-project#6277) * Fix Qwen3MoE missing token padding optimization (sgl-project#6820) * Tiny update error hints (sgl-project#6846) * Support layerwise rebalancing experts (sgl-project#6851) * Tiny allow profiler API to auto create directory (sgl-project#6865) * Support Blackwell DeepEP docker images (sgl-project#6868) * [EP] Add cuda kernel for moe_ep_post_reorder (sgl-project#6837) * [theta]merge 0605 * oai: fix openAI client error with single request via batch api (sgl-project#6170) * [PD] Fix potential perf spike caused by tracker gc and optimize doc (sgl-project#6764) * Use deepgemm instead of triton for fused_qkv_a_proj_with_mqa (sgl-project#6890) * [CUTLASS-FP4-MOE] Introduce CutlassMoEParams class for easy initialization of Cutlass Grouped Gems Metadata (sgl-project#6887) * bugfix(OAI): Fix image_data processing for jinja chat templates (sgl-project#6877) * [CPU] enable CI for PRs, add Dockerfile and auto build task (sgl-project#6458) * AITER backend extension and workload optimizations (sgl-project#6838) * [theta]merge * [theta]merge * [Feature] Support Flashinfer fmha on Blackwell (sgl-project#6930) * Fix a bug in abort & Improve docstrings for abort (sgl-project#6931) * Tiny support customize DeepEP max dispatch tokens per rank (sgl-project#6934) * Sync the changes on cuda graph runners (sgl-project#6932) * [PD] Optimize transfer queue forward logic for dummy rank (sgl-project#6922) * [Refactor] image data process in bench_serving (sgl-project#6879) * [fix] logical_to_all_physical_map index 256 is out of bounds in EP parallel. (sgl-project#6767) * Add triton fused moe kernel config for E=257 on B200 (sgl-project#6939) * [sgl-kernel] update deepgemm (sgl-project#6942) * chore: bump sgl-kernel v0.1.6 (sgl-project#6943) * Minor compile fused topk (sgl-project#6944) * [Bugfix] pipeline parallelism and Eagle Qwen2 (sgl-project#6910) * Tiny re-introduce profile id logging (sgl-project#6912) * Add triton version as a fused_moe_triton config search key to avoid performace decrease in different Triton version (sgl-project#5955) * reduce torch.zeros overhead in moe align block size kernel (sgl-project#6369) * chore: upgrade sgl-kernel v0.1.6 (sgl-project#6945) * add fbgemm moe grouped gemm kernel benchmark (sgl-project#6924) * [Docker] Add docker file for SGL Router (sgl-project#6915) * Disabling mixed chunked prefill when eagle is enabled (sgl-project#6874) * Add canary for EPLB rebalancing (sgl-project#6895) * Refactor global_server_args_dict (sgl-project#6866) * Fuse routed scaling factor in topk_reduce kernel (sgl-project#6220) * Update server timeout time in AMD CI. (sgl-project#6953) * [misc] add is_cpu() (sgl-project#6950) * Add H20 fused MoE kernel tuning configs for DeepSeek-R1/V3 (sgl-project#6885) * Add a CUDA kernel for fusing mapping and weighted sum for MoE. (sgl-project#6916) * chore: bump sgl-kernel v0.1.6.post1 (sgl-project#6955) * chore: upgrade sgl-kernel v0.1.6.post1 (sgl-project#6957) * [DeepseekR1-FP4] Add Support for nvidia/DeepSeekR1-FP4 model (sgl-project#6853) * Revert "Fuse routed scaling factor in topk_reduce kernel (sgl-project#6220)" (sgl-project#6968) * [AMD] Add more tests to per-commit-amd (sgl-project#6926) * chore: bump sgl-kernel v0.1.7 (sgl-project#6963) * Slightly improve the sampler to skip unnecessary steps (sgl-project#6956) * rebase h20 fused_moe config (sgl-project#6966) * Fix CI and triton moe Configs (sgl-project#6974) * Remove unnecessary kernels of num_token_non_padded (sgl-project#6965) * Extend cuda graph capture bs for B200 (sgl-project#6937) * Fuse routed scaling factor in deepseek (sgl-project#6970) * Sync cuda graph runners (sgl-project#6976) * Fix draft extend ut stability with flush cache (sgl-project#6979) * Fix triton sliding window test case (sgl-project#6981) * Fix expert distribution dumping causes OOM (sgl-project#6967) * Minor remove one kernel for DeepSeek (sgl-project#6977) * [perf][sgl-kernel] extend cutlass_mla_decode to support num_head < 128 (sgl-project#6929) * Enable more unit tests for AMD CI. (sgl-project#6983) * Use torch.compile to fuse flash attention decode metadata preparation (sgl-project#6973) * Eliminate stream sync to speed up LoRA batch init (sgl-project#6960) * support qwen3 emebedding (sgl-project#6990) * Fix torch profiler bugs for bench_offline_throughput.py (sgl-project#6557) * chore: upgrade flashinfer v0.2.6.post1 jit (sgl-project#6958) * cleanup tmp dir (sgl-project#7007) * chore: update pr test xeon (sgl-project#7008) * Fix cutlass MLA gets almost zero accuracy (sgl-project#6998) * Update amd nightly models CI. (sgl-project#6992) * feat: add direct routing strategy to DP worker (sgl-project#6884) * Fallback to lower triton version for unfound fused moe configs (sgl-project#7013) * Fix torchvision version for Blackwell (sgl-project#7015) * Simplify prepare_extend_after_decode (sgl-project#6987) * Migrate to assertEqual (sgl-project#6741) * Fix torch version in blackwell dockerfile (sgl-project#7017) * chore: update pr test xeon (sgl-project#7018) * Update default settings for blackwell (sgl-project#7023) * Support both approximate and exact expert distribution collection (sgl-project#6964) * Add decode req pool (sgl-project#6980) * [theta]merge 0610 * [theta]merge 0610 * [CI] Add CI workflow for sgl-router docker build (sgl-project#7027) * Fix fused_moe triton configs (sgl-project#7029) * CPU: map changes from developing branch in sgl-kernel (sgl-project#6833) * chore: bump v0.4.7 (sgl-project#7038) * Update README.md (sgl-project#7040)
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Motivation
An inelegant workaround, thus the kernel have to be rewrited to support num_head != 128
Limitation:
TODO futher:
Try per batch spilt kv schedule
Tables below shows GB/s of cases
Modifications
Checklist