Minor remove one kernel for DeepSeek #6977
Conversation
There was a problem hiding this comment.
Hello @fzyzcjy, 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/gemini-code-assist here, providing a summary of this pull request to help everyone get up to speed quickly.
This pull request, titled "Minor remove one kernel for DeepSeek" by fzyzcjy, focuses on a specific adjustment within the DeepSeek v2 model implementation, specifically in the forward_deepep function located in python/sglang/srt/models/deepseek_v2.py. The primary change involves modifying how the self.routed_scaling_factor is applied to the final_hidden_states and combined with shared_output. The original code scaled final_hidden_states unconditionally and then added shared_output. This PR changes the logic to apply the scaling factor as part of an in-place addition (add_) when shared_output is present. If shared_output is None, the scaling is now applied directly in an else block. This modification likely aims to optimize the computation by potentially using a more efficient kernel for the combined scale-and-add operation or to correct the application of the scaling factor under different conditions.
Highlights
- DeepSeek Model Logic: Modifies the
forward_deepepfunction within the DeepSeek v2 model implementation (python/sglang/srt/models/deepseek_v2.py). - Scaling Factor Application: Adjusts how the
self.routed_scaling_factoris applied tofinal_hidden_states, particularly when combining withshared_output. - Optimization/Correction: Changes the addition logic to use an in-place
add_with an alpha parameter whenshared_outputis present, potentially optimizing the operation or ensuring correct behavior. - Conditional Logic: Moves the direct scaling of
final_hidden_statesinto anelseblock, applying it only whenshared_outputisNone.
Changelog
- python/sglang/srt/models/deepseek_v2.py
- Modified the
forward_deepepfunction to change the logic for applyingself.routed_scaling_factorand combiningfinal_hidden_stateswithshared_output. - Replaced the unconditional scaling and subsequent addition with a conditional structure (lines 424-429): if
shared_outputis not None, useshared_output.add_(final_hidden_states, alpha=self.routed_scaling_factor); otherwise, apply direct scalingfinal_hidden_states *= self.routed_scaling_factor.
- Modified the
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. ↩
There was a problem hiding this comment.
Code Review
This pull request refactors how the routed_scaling_factor is applied when shared_output is present in the forward_deepep method of deepseek_v2.py. The change appears to be mathematically equivalent to the previous logic and aims for potential memory efficiency by using an in-place operation.
The PR title "Minor remove one kernel for DeepSeek" is noted, though this specific code modification doesn't directly remove a kernel in the Python code itself. It might enable better compiler fusion or be part of a larger set of changes that achieved this.
The PR description is currently a template. It would be helpful to fill it out with the motivation and a summary of modifications for future reference and easier review.
Summary of Findings
- In-place modification of
shared_output: The primary change involves an in-place modification of theshared_outputtensor when it's not None. This is a valid optimization ifshared_outputis an intermediate tensor not used elsewhere in its original state. Confirmation of this assumption is requested to ensure no unintended side effects. A suggestion for a slightly more direct way to write the in-place update has also been provided. - PR Context: The PR description is currently a template. Filling it with details about the motivation and changes would improve clarity and maintainability.
Merge Readiness
The core logic change seems sound and aims for a valid optimization. However, confirmation regarding the in-place modification of shared_output is important to ensure correctness and avoid potential side effects. Addressing this point and potentially filling out the PR description would make the PR ready for merge. I am unable to approve the pull request, but I recommend that the author address the comment before merging.
| x = shared_output | ||
| x.add_(final_hidden_states, alpha=self.routed_scaling_factor) | ||
| final_hidden_states = x |
There was a problem hiding this comment.
The new logic for when shared_output is not None modifies shared_output in-place via the reference x before assigning it to final_hidden_states. This approach (shared_output_original + final_hidden_states_unscaled * scale) is mathematically equivalent to the previous logic ((final_hidden_states_unscaled * scale) + shared_output_original) and can be more memory-efficient by avoiding an intermediate tensor allocation.
Could you confirm that shared_output is an intermediate tensor that can be safely modified in-place here and is not used elsewhere in its original form after this operation?
If so, this is a good optimization. For slightly more directness, you could consider directly operating on shared_output without the intermediate x variable, as shown in the suggestion.
| x = shared_output | |
| x.add_(final_hidden_states, alpha=self.routed_scaling_factor) | |
| final_hidden_states = x | |
| # Update shared_output in-place, then assign to final_hidden_states. | |
| # This is equivalent to: | |
| # final_hidden_states = shared_output_original + final_hidden_states_unscaled * self.routed_scaling_factor | |
| # and can be more memory-efficient if shared_output is an intermediate tensor. | |
| shared_output.add_(final_hidden_states, alpha=self.routed_scaling_factor) | |
| final_hidden_states = shared_output |
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
Modifications
Checklist