Improve EPLB logical to physical dispatch map #6727
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 or gemini-code-assist here, providing a summary of this pull request to help everyone get up to speed.
This pull request, authored by fzyzcjy, focuses on optimizing the Expert Parallelism (EP) setup within the system. The primary goal, as stated in the description, is to implement a faster logical-to-physical expert dispatch map. This is expected to yield performance improvements, specifically by speeding up GPU time during the forward pass and potentially optimizing CPU time when executing rebalancing operations.
Highlights
- Performance Optimization: The core motivation for this change is to improve the performance of the Expert Parallelism setup, targeting faster GPU execution and optimized CPU rebalancing.
- Rewrite of Dispatch Logic: The function responsible for computing the logical-to-rank dispatch physical map (
compute_logical_to_rank_dispatch_physical_map) has been completely rewritten to implement a new strategy. - Prioritizing Local Experts: The new dispatch logic attempts to prioritize assigning a logical expert to a physical expert that resides on the same GPU rank, which can help reduce inter-GPU communication overhead.
- Improved Expert Assignment Strategy: The updated logic first assigns a local expert if available and then uses a new helper function (
_fair_choices) with controlled randomness to fill any remaining dispatch slots from the pool of all possible physical experts for that logical expert.
Changelog
- python/sglang/srt/managers/expert_location.py
- Imported the
randommodule (L16). - Removed the
logical_to_all_physical_map_num_validparameter from the call tocompute_logical_to_rank_dispatch_physical_mapwithin the_init_rawfunction (L202). - Completely rewrote the
compute_logical_to_rank_dispatch_physical_mapfunction to implement the new dispatch logic, including removing thelogical_to_all_physical_map_num_validparameter from its signature and renamingbase_seedtoseed(starts L294). - Added three new helper functions:
_logical_to_all_physical_raw,_compute_gpu_id_of_physical_expert, and_fair_choicesto support the new dispatch calculation (starts L346).
- Imported 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.
Experts spread afar,
Map them fast to GPU,
Forward pass flies now.
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 the compute_logical_to_rank_dispatch_physical_map function, changing how logical experts are mapped to physical experts for each GPU rank. The new approach prioritizes local physical experts and then uses a fair selection mechanism for remaining assignments. This is intended to improve GPU performance during forward passes and optimize CPU time for future rebalancing operations.
The introduction of helper functions like _logical_to_all_physical_raw, _compute_gpu_id_of_physical_expert, and _fair_choices helps in breaking down the logic. However, there are a few areas regarding correctness, performance, and behavior that would benefit from further attention.
Summary of Findings
- Potential Runtime Error: A
ZeroDivisionErrorcould occur in the_fair_choicesfunction if it's called with an empty list of candidates when assignments are still needed. This needs to be addressed to prevent crashes for certain input configurations. - Performance Considerations: The refactored map computation logic uses more Python-level iteration and data conversions, which might impact its CPU performance. While a TODO for optimization exists, clarification on current performance expectations would be helpful.
- Behavioral Change in Expert Selection: The selection of a local physical expert (if multiple are available on the same GPU) has changed from potentially random to deterministic (picking the first). Confirmation of this intended change is sought.
- Missing Type Hint (Low Severity - Not Commented): In
python/sglang/srt/managers/expert_location.py, the function_logical_to_all_physical_rawis missing a type hint for its first parameterlogical_to_all_physical_map. It should betorch.Tensor.
Merge Readiness
The pull request introduces a significant refactoring to improve the expert dispatch map. While the new logic appears sound in its approach, there are a few critical and medium severity concerns related to potential runtime errors, performance implications of the new implementation, and a behavioral change in expert selection that should be addressed or clarified before merging. I am not authorized to approve pull requests, so please ensure these points are resolved and the code is reviewed by other maintainers.
| output_partial[output_partial == -1] = torch.tensor( | ||
| _fair_choices(candidate_physical_expert_ids, k=num_remain, r=r), | ||
| dtype=dtype, | ||
| ) |
There was a problem hiding this comment.
There's a potential ZeroDivisionError here if _fair_choices is called with an empty candidate_physical_expert_ids list and num_remain > 0. This would happen if a logical expert has no physical mappings, but we still need to assign it to num_remain GPU slots.
While the assertion on line 340 (torch.all(logical_to_rank_dispatch_physical_map != -1)) aims to ensure all slots are filled, the ZeroDivisionError in _fair_choices (specifically in divmod(k, len(arr))) would occur first.
Could we make _fair_choices more robust by handling the len(arr) == 0 case explicitly? For example, it could raise a ValueError if k > 0 and arr is empty.
Alternatively, should an assertion be added before this call to ensure candidate_physical_expert_ids is not empty if num_remain > 0? This would imply that every logical expert must have at least one physical candidate if it needs to be assigned.
| output_partial[output_partial == -1] = torch.tensor( | |
| _fair_choices(candidate_physical_expert_ids, k=num_remain, r=r), | |
| dtype=dtype, | |
| ) | |
| def _fair_choices(arr: List, k: int, r: random.Random) -> List: | |
| if not arr: | |
| raise ValueError("Cannot choose from an empty list") | |
| quotient, remainder = divmod(k, len(arr)) | |
| ans = arr * quotient + r.sample(arr, k=remainder) | |
| r.shuffle(ans) | |
| return ans |
| for layer_id in range(num_layers): | ||
| for logical_expert_id in range(num_logical_experts): | ||
| candidate_physical_expert_ids = _logical_to_all_physical_raw( | ||
| logical_to_all_physical_map, layer_id, logical_expert_id | ||
| ) | ||
| output_partial = logical_to_rank_dispatch_physical_map[ | ||
| :, layer_id, logical_expert_id | ||
| ] | ||
|
|
||
| for gpu_id in range(num_gpus): | ||
| same_gpu_physical_expert_ids = [ | ||
| physical_expert_id | ||
| for physical_expert_id in candidate_physical_expert_ids | ||
| if _compute_gpu_id_of_physical_expert( | ||
| physical_expert_id, num_local_physical_experts | ||
| ) | ||
| == gpu_id | ||
| ] | ||
| if len(same_gpu_physical_expert_ids) > 0: | ||
| output_partial[gpu_id] = same_gpu_physical_expert_ids[0] | ||
|
|
||
| num_remain = torch.sum(output_partial == -1).item() | ||
| output_partial[output_partial == -1] = torch.tensor( | ||
| _fair_choices(candidate_physical_expert_ids, k=num_remain, r=r), | ||
| dtype=dtype, | ||
| ) |
There was a problem hiding this comment.
The new implementation of compute_logical_to_rank_dispatch_physical_map involves Python loops (over layers, logical experts, and GPUs), list comprehensions, and conversions between PyTorch tensors and Python lists/scalars within these loops (e.g., _logical_to_all_physical_raw using .tolist(), and torch.tensor() for _fair_choices result).
This contrasts with the previous, more vectorized PyTorch implementation. While the PR aims to optimize CPU time for rebalancing later and improve GPU forward pass time (due to a better map), I'm wondering about the CPU performance of this specific function for computing the map itself. The TODO comment on line 293 (TODO optimize performance (rewrite and/or run in separate process with overlap)) suggests this is a known concern.
Could you elaborate on the expected performance characteristics of this new map computation logic, especially compared to the old one? Is the current CPU performance acceptable for the typical scenarios where this map is computed (e.g., initialization, rebalancing)?
| if len(same_gpu_physical_expert_ids) > 0: | ||
| output_partial[gpu_id] = same_gpu_physical_expert_ids[0] |
There was a problem hiding this comment.
In this part of the logic:
if len(same_gpu_physical_expert_ids) > 0:
output_partial[gpu_id] = same_gpu_physical_expert_ids[0]If multiple physical experts on the same GPU (gpu_id) can serve the current logical_expert_id, the code now deterministically picks the first one (same_gpu_physical_expert_ids[0]).
The previous implementation had an initial random selection phase. Was this change to a deterministic selection (picking the first available) an intentional design choice? If so, what's the reasoning behind it?
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
This will speed up gpu time for each forward pass; Will optimize cpu time when executing rebalancing later
Modifications
Checklist