Add draft extend CUDA graph for Triton backend #6705
Conversation
There was a problem hiding this comment.
Hello @ispobock, 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 team,
Gemini here, providing a summary of this pull request. This PR, titled "Add draft extend CUDA graph for Triton backend", is a follow-up to #6606 and focuses on optimizing the 'draft extend' phase of speculative decoding when using the Triton attention backend. The core idea is to enable CUDA graph capture and replay for this specific operation, which can reduce launch overhead and potentially improve performance.
The description includes benchmark results comparing the main branch against this PR using the EAGLE speculative algorithm with the Triton attention backend. The tables show improvements in input and output throughput, as well as reduced mean, median, and p99 latencies (TTFT and TPOT), particularly as concurrency increases. This suggests the CUDA graph optimization is having a positive impact on performance for the draft extension step.
Highlights
- CUDA Graph for Draft Extend: The primary goal is to enable CUDA graph capture and replay for the 'draft extend' operation within the Triton attention backend, aiming to reduce kernel launch overhead.
- Triton Backend Implementation: Specific logic is added to the
TritonAttnBackendto handle the metadata setup required for 'draft extend' mode when capturing and replaying CUDA graphs, utilizing FlashInfer kernels. - EAGLE Worker Integration: The EAGLE speculative decoding worker is updated to utilize the
TritonAttnBackendfor the draft extension phase when the Triton attention backend is configured. - Performance Improvement: Benchmark results provided in the description indicate performance gains in throughput and latency when running with the Triton backend and speculative decoding enabled by this PR.
Changelog
- python/sglang/srt/layers/attention/triton_backend.py
- Added
speculative_num_stepsattribute to theTritonAttnBackendclass. - Implemented the
elif forward_mode.is_draft_extend():block ininit_forward_metadata_capture_cuda_graphto prepare necessary metadata (likeqo_indptr,kv_indptr,kv_indices) for CUDA graph capture in draft extend mode. - Implemented the
elif forward_mode.is_draft_extend():block ininit_forward_metadata_replay_cuda_graphto prepare metadata for CUDA graph replay in draft extend mode.
- Added
- python/sglang/srt/speculative/eagle_worker.py
- Imported
TritonAttnBackend. - Initialized
self.draft_extend_attn_backendwith an instance ofTritonAttnBackendwhen the attention backend is set to 'triton', enabling its use for the draft extension step.
- Imported
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.
Graph the kernels tight,
Reduce the overhead's blight,
Speed takes its swift flight.
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. ↩
There was a problem hiding this comment.
Code Review
This pull request aims to add CUDA graph support for the draft extend functionality in the Triton attention backend, following up on #6606.
The code changes themselves in triton_backend.py and eagle_worker.py seem to correctly set up the necessary metadata and backend initialization for this new graph path.
However, a significant concern is that the performance tables provided in the PR description for 'main' and 'This PR' are identical. This raises questions about whether the benchmark is effectively exercising the new CUDA graph path for draft extend, or if there's an issue with the performance measurement. Clarification on this would be highly appreciated, as demonstrating the impact (ideally, improvement) of these changes is key.
Overall, the introduced logic for graph capture and replay for draft_extend appears consistent with CUDA graph principles (static maximums for capture, dynamic values for replay).
Summary of Findings
- Performance Benchmark Verification: The performance tables in the PR description are identical for 'main' and 'This PR'. This needs verification to ensure the new CUDA graph path for draft extend is being exercised by the benchmark and to understand its performance impact.
- Clarity on
spec_infoUsage in Graph Capture: The CUDA graph capture path fordraft_extenddoes not usespec_info, unlike the non-capture path. While likely intentional for graph capture (using static maximums), a comment explaining this would enhance code clarity.
Merge Readiness
The core logic for adding CUDA graph support for draft extend in the Triton backend appears sound. However, the primary concern is the identical performance benchmarks provided in the PR description. It's crucial to verify that these changes are indeed being tested by the benchmark and to understand their actual performance implications.
Additionally, a minor clarification regarding spec_info usage in the graph capture path would be beneficial for maintainability.
Given the uncertainty around the performance impact, I recommend that these points, especially the benchmark verification, be addressed before merging. As an AI reviewer, I am not authorized to approve pull requests; please ensure further review and approval from the maintainers.
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)
Motivation
Follow up of #6606.
main:
This PR: