[perf] Improve Prefill Performance by Removing Redundant Padding and Optimizing Alltoall Communication #949
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
Signed-off-by: SlightwindSec <[email protected]>
…oject#945) Adjust inputbatch to be compatible with latest vllm, as kvcache group feature has been redo in vllm-project/vllm#18593 --------- Signed-off-by: MengqingCao <[email protected]>
### What this PR does / why we need it? This is a continuing work of vllm-project#716. This PR add workflow to build and release wheel, and also release source to PYPI. We have 3 conditions to trigger the workflow: 1. PR to `main` and `*-dev` 2. push to `main` and `*-dev` 3. push tag with name of `v*` Release to PYPI will only be done under condition 3. Under condition 1 and 2, it will generate .tar.gz and build .whl, upload to github artifacts but will not release. update: Will build .whl and upload to github artifacts with scheduled task. ### Does this PR introduce _any_ user-facing change? No. ### How was this patch tested? All triggered conditions are well tested with my fork repo. --------- Signed-off-by: Shuqiao Li <[email protected]> Signed-off-by: Yikun Jiang <[email protected]> Co-authored-by: Yikun Jiang <[email protected]>
… FusedMoEParallelConfig when using vLLM 0.9.0 (vllm-project#961) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? This PR fix accuracy issues incurred by codes that adapt to `FusedMoEParallelConfig` in vLLM 0.9.0 version. The `tp_size` used to split weights are wrongly passed. The root cause is that vLLM community and vLLM-Ascend are using different methods to decide whether to use Expert Parallel. vLLM: vLLM use a flag `enable_expert_parallel` to indicate whether to use EP and use the following codes to decide `ep_size`: ``` use_ep = (dp_size_ * tp_size_ > 1 and vllm_parallel_config.enable_expert_parallel) dp_size = dp_size_ dp_rank = get_dp_group().rank_in_group if dp_size > 1 else 0 tp_size, tp_rank = flatten_tp_across_dp(dp_rank) if not use_ep: return FusedMoEParallelConfig(tp_size=tp_size, tp_rank=tp_rank, dp_size=dp_size, dp_rank=dp_rank, ep_size=1, ep_rank=0, use_ep=False) # DP + EP / TP + EP / DP + TP + EP assert use_ep # In EP, each device owns a set of experts fully. There is no tensor # parallel update tp_size, tp_rank, ep_size and ep_rank to reflect that. ep_size = tp_size ep_rank = tp_rank return FusedMoEParallelConfig(tp_size=1, tp_rank=0, dp_size=dp_size, dp_rank=dp_rank, ep_size=ep_size, ep_rank=ep_rank, use_ep=True) ``` vLLM-Ascend: vLLM-Ascend uses `etp` to specify Tensor Parallel in MoE. ``` self.ep_size = get_ep_group().world_size self.tp_size = get_etp_group().world_size self.dp_size = (dp_size if dp_size is not None else get_dp_group().world_size) ``` So there will be conflicts if we simply combine these codes together. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> Signed-off-by: angazenn <[email protected]> Co-authored-by: angazenn <[email protected]>
added 3 commits
May 27, 2025 23:07
Signed-off-by: SlightwindSec <[email protected]>
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.
What this PR does / why we need it?
This PR improves Prefill performance by making two key optimizations:
all_to_all_singlecall followed by threeall_to_allcalls. This has been refactored to use threeall_to_all_singlecalls instead, with a fixed communication buffer to eliminate an extra communication step. This change not only simplifies the communication pattern but also leverages the better performance ofall_to_all_single.While there might be minor precision trade-offs, the choice of the coefficient
2is an empirically sound value that maintains accuracy even when expert ID distribution is imbalanced.In testing with DeepSeek-V3, the model was able to handle 3584-token inputs with significantly improved Prefill throughput and no regression in dialog quality.
Does this PR introduce any user-facing change?
No, this PR does not introduce any user-facing changes.
How was this patch tested?