[david6666666]

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

Purpose

issue #20805 #24116
pr #18343
EPLB Execution

• Parallelize the rearrangement algorithm (calculating new expert mapping, not the communication)
• Shuffle one layer at once and use multiple steps, to lower the impact on inter-token latency
• Investigate should we pre-allocate expert weight buffer used for transferring
• Take locality into consideration in expert weight transmission, e.g. prioritize transferring to GPUs on the same node
• Use cuda.Stream() asynchronously moves the weight to the buffer

They jointly provided the initial code
5 Key Pitfalls in EPLB Asynchronous Core Implementation

1. CUDA Stream must be bound to Device:
   After creating an asynchronous thread, the initialized torch.cuda.Stream must be explicitly bound to the target GPU device. Otherwise, mismatched devices will cause asynchronous task execution to fail.

2. Separate “move_to_buffer” and “move_to_workspace”:
   move_to_buffer: should be executed in asynchronous threads to hide data transfer latency. move_to_workspace: involves copying cached weights into the working area. Since this is an on-device operation (only ~300–400 µs), it must be executed synchronously. However, the performance impact is minimal.

3. Layer-level synchronization: wait for P2P completion:
   After each layer executes move to buffer, all GPUs must wait until their P2P (point-to-point) send/receive operations finish. This ensures that subsequent steps won’t fail due to missing or incomplete data.

4. All Reduce must be coupled with Rank synchronization to avoid deadlocks:
   While waiting for P2P operations, an All Reduce is required to ensure all Ranks (GPUs) reach the same condition. This All Reduce must be executed at every rank interaction, otherwise desynchronized progress across GPUs may cause deadlocks. To reduce overhead, All Reduce should only be performed during the Rearrange phase, not in other stages.

5. Update expert weights after each layer’s move_from_buffer:
   After move_from_buffer at each layer, expert weights must be updated. Otherwise, computations may use stale weights, causing mapping inconsistencies and leading to accuracy issues.

Test Plan

pytest tests/distributed/test_eplb_execute.py
pytest tests/distributed/test_eplb_spec_decode.py

Test Result

Benchmark TP=8,EP=8, 8XH800 80G dataset: MMLU-pro,
Due to resource limitations, we only measured up to EP8. A larger EP size with no-blocking EPLB would yield greater benefits:
benchmark test without eplb:

vllm serve Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 \
     --tensor-parallel-size 8 \
     --enable-expert-parallel \
     --gpu-memory-utilization 0.8 \
     --trust-remote-code

vllm serve Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 \
     --tensor-parallel-size 8 \
     --enable-expert-parallel \
     --enable-eplb \
     --eplb-config '{"window_size":1000,"step_interval":3000}' \
     --gpu-memory-utilization 0.8 \
     --trust-remote-code

vllm serve Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 \
     --tensor-parallel-size 8 \
     --enable-expert-parallel \
     --enable-eplb \
     --eplb-config '{"window_size":1000,"step_interval":3000,"use_async":"true"}' \
     --gpu-memory-utilization 0.8 \
     --trust-remote-code

accuracy:

python tests/evals/gsm8k/gsm8k_eval.py --port 8000

w/o EPLB:
Running GSM8K evaluation: 1319 questions, 5-shot
Evaluating: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1319/1319 [01:07<00:00, 19.67it/s]
Results:
Accuracy: 0.901
Invalid responses: 0.000    
Total latency: 67.070 s     
Questions per second: 19.666

w/ EPLB, main:
Running GSM8K evaluation: 1319 questions, 5-shot
Evaluating: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1319/1319 [01:22<00:00, 16.00it/s]
Results:
Accuracy: 0.902
Invalid responses: 0.000    
Total latency: 82.434 s     
Questions per second: 16.001

w/ EPLB, this PR:
Running GSM8K evaluation: 1319 questions, 5-shot
Evaluating: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1319/1319 [01:20<00:00, 16.46it/s]
Results:
Accuracy: 0.908
Invalid responses: 0.000    
Total latency: 80.133 s     
Questions per second: 16.460

async EPLB + MTP:
vllm serve Qwen/Qwen3-Next-80B-A3B-Instruct --max-model-len 2048  --trust-remote-code --port 8000 -dp 1 -tp 4 --enable-expert-parallel --gpu-memory-utilization 0.86 --enable-eplb  --eplb-config '{"window_size":200,"step_interval":600,"use_async":"true"}' --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":1}' 2>&1 | tee ../prof/test_async_mtp.log

Running GSM8K evaluation: 1319 questions, 5-shot
Evaluating: 100%|████████████████████████████████████████████████████████████████| 1319/1319 [01:34<00:00, 14.00it/s]

Results:
Accuracy: 0.864
Invalid responses: 0.000
Total latency: 94.256 s
Questions per second: 13.994

benchmark:

vllm bench serve  \
--backend vllm   \
--model Qwen/Qwen3-235B-A22B-Instruct-2507-FP8   \
--endpoint /v1/completions   \
--dataset-name custom   \
--dataset-path /root/.cache/cwq/datasets/mmlu_pro_test.jsonl   \
--max-concurrency 64 --num-prompt 1000

w/o EPLB:
============ Serving Benchmark Result ============
Successful requests:                     1000
Maximum request concurrency:             64
Benchmark duration (s):                  109.47
Total input tokens:                      69309
Total generated tokens:                  231400
Request throughput (req/s):              9.13
Output token throughput (tok/s):         2113.74
Peak output token throughput (tok/s):    2367.00
Peak concurrent requests:                112.00
Total Token throughput (tok/s):          2746.85
---------------Time to First Token----------------
Mean TTFT (ms):                          93.20
Median TTFT (ms):                        61.91
P99 TTFT (ms):                           310.75
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          29.11
Median TPOT (ms):                        29.18
P99 TPOT (ms):                           29.96
---------------Inter-token Latency----------------
Mean ITL (ms):                           29.10
Median ITL (ms):                         28.52
P99 ITL (ms):                            38.27
==================================================

w/ EPLB, main:
============ Serving Benchmark Result ============
Successful requests:                     1000
Maximum request concurrency:             64
Benchmark duration (s):                  131.87
Total input tokens:                      69309
Total generated tokens:                  231534
Request throughput (req/s):              7.58
Output token throughput (tok/s):         1755.72
Peak output token throughput (tok/s):    2068.00
Peak concurrent requests:                113.00
Total Token throughput (tok/s):          2281.29
---------------Time to First Token----------------
Mean TTFT (ms):                          97.56
Median TTFT (ms):                        69.89
P99 TTFT (ms):                           348.90
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          35.04
Median TPOT (ms):                        33.06
P99 TPOT (ms):                           63.84
---------------Inter-token Latency----------------
Mean ITL (ms):                           35.07
Median ITL (ms):                         32.17
P99 ITL (ms):                            45.01
==================================================

w/ EPLB, this PR:
============ Serving Benchmark Result ============
Successful requests:                     1000
Maximum request concurrency:             64
Benchmark duration (s):                  129.85
Total input tokens:                      69309
Total generated tokens:                  230828
Request throughput (req/s):              7.70
Output token throughput (tok/s):         1777.61
Peak output token throughput (tok/s):    2047.00
Peak concurrent requests:                114.00
Total Token throughput (tok/s):          2311.35
---------------Time to First Token----------------
Mean TTFT (ms):                          100.83
Median TTFT (ms):                        71.17
P99 TTFT (ms):                           350.46
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          34.47
Median TPOT (ms):                        33.27
P99 TPOT (ms):                           46.01
---------------Inter-token Latency----------------
Mean ITL (ms):                           34.51
Median ITL (ms):                         32.28
P99 ITL (ms):                            45.79
==================================================

profile:
non-blocking EPLB:

(Optional) Documentation Update

Co-authored-by: jiangkuaixue123 jiangxiaozhou111@163.com
Co-authored-by: SunChenxiang123 1291824390@qq.com

[gemini-code-assist]

Code Review

This pull request introduces an asynchronous, layer-by-layer execution optimization for EPLB, which is a great step towards reducing latency. The core idea is solid, but the implementation has several critical correctness issues, mainly related to the new asynchronous logic. I've identified bugs like undefined variables, incorrect method calls, and improper use of asyncio primitives. Addressing these will be crucial for the stability of this feature. I've also pointed out some maintainability improvements, such as translating comments and using the standard logger.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[hmellor]

Thank you for the optimisation.

Please wait for #20562 and add the new config to the new EPLBConfig class.

[david6666666]

Thank you for the optimisation.

Please wait for #20562 and add the new config to the new EPLBConfig class.

OK, after I prepare this PR and test result, I will rebase and add the new config to the new EPLBConfig class

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @david6666666.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[david6666666]

@hmellor @abmfy the implementation is ready, PTAL, thanks. wait for #20562 and add the new config to the new EPLBConfig class and rebase

[jiangkuaixue123]

Please check the above issues. Thank you.

[david6666666]

@simon-mo CI has passed, thanks for your support

[tlrmchlsmth]

We should add coverage for async eplb to our nightly tests -- could be a follow up PR, but we should add it to the following DSv2 lite test (and the Qwen test when eplb gets turned on)

vllm/.buildkite/test-pipeline.yaml

Lines 1334 to 1350 in 4de8786

	- label: DeepSeek V2-Lite Accuracy
	timeout_in_minutes: 60
	gpu: h100
	optional: true
	num_gpus: 4
	working_dir: "/vllm-workspace"
	commands:
	- bash .buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh 0.25 200 8010
	- label: Qwen3-30B-A3B-FP8-block Accuracy
	timeout_in_minutes: 60
	gpu: h100
	optional: true
	num_gpus: 4
	working_dir: "/vllm-workspace"
	commands:
	- bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep.sh 0.8 200 8020

[tlrmchlsmth]

It's not clear to me what this comment is describing

Suggested change

	assert num_physical_experts == ep_size * num_local_physical_experts
	# A buffer to hold the expert weights in one layer during the exchange.
	# NOTE: Currently we assume the same weights across different layers
	# have the same shape.
	is_unchanged, is_received_locally, experts_recv_loc = move_to_buffer(
	assert num_physical_experts == ep_size * num_local_physical_experts
	# A buffer to hold the expert weights in one layer during the exchange.
	# NOTE: Currently we assume the same weights across different layers
	# have the same shape.
	is_unchanged, is_received_locally, experts_recv_loc = move_to_buffer(

[david6666666]

useless comment, will removed with a follow up PR

[david6666666]

We should add coverage for async eplb to our nightly tests -- could be a follow up PR, but we should add it to the following DSv2 lite test (and the Qwen test when eplb gets turned on)

vllm/.buildkite/test-pipeline.yaml

Lines 1334 to 1350 in 4de8786

	- label: DeepSeek V2-Lite Accuracy
	timeout_in_minutes: 60
	gpu: h100
	optional: true
	num_gpus: 4
	working_dir: "/vllm-workspace"
	commands:
	- bash .buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh 0.25 200 8010
	- label: Qwen3-30B-A3B-FP8-block Accuracy
	timeout_in_minutes: 60
	gpu: h100
	optional: true
	num_gpus: 4
	working_dir: "/vllm-workspace"
	commands:
	- bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep.sh 0.8 200 8020

thx, I will open a follow up PR to coverage this