[MoE][Dist] Fix Qwen MoE accuracy bug in DP scenario #1856

MengqingCao · 2025-07-17T10:35:45Z

What this PR does / why we need it?

Fix Qwen MoE accuracy bug in DP scenario.

Now the implentment of FusedMoE in vLLM use All2AllManager to manager different all2all algorithm branch. And the default branch use Multicast in dispatch phase and all_reduce in combine phase, which are not implented in vLLM-Ascend. This leading to invoking into a default implentment in base_communicator, with empty dispatch and combine operations, thus causing the accuracy issue on it.

This pr is a temporary workaround, refacting all2all in vLLM-Ascend could be a better way.

Does this PR introduce any user-facing change?

How was this patch tested?

vLLM version: v0.10.0
vLLM main: vllm-project/vllm@ad57f23

codecov · 2025-07-17T11:01:59Z

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 76.67%. Comparing base (72eceff) to head (d768bc2).
⚠️ Report is 643 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1856      +/-   ##
==========================================
+ Coverage   74.41%   76.67%   +2.26%     
==========================================
  Files         100      107       +7     
  Lines       11208    11968     +760     
==========================================
+ Hits         8340     9177     +837     
+ Misses       2868     2791      -77

Flag	Coverage Δ
unittests	`76.67% <ø> (+2.26%)`	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

vllm_ascend/distributed/communicator.py

Potabk · 2025-07-18T02:25:20Z

What is confusing is that this patch indeed can solve the accuracy problem of online scenes, but it destroys the functional availability of offline scenes.

Potabk · 2025-07-18T02:30:10Z

online:

run dp2 on single node

#!/bin/sh

# this obtained through ifconfig
# nic_name is the network interface name corresponding to local_ip
nic_name="enp67s0f5"
local_ip="192.168.0.183"

export HCCL_IF_IP=$local_ip
export GLOO_SOCKET_IFNAME=$nic_name
export TP_SOCKET_IFNAME=$nic_name
export HCCL_SOCKET_IFNAME=$nic_name
export OMP_PROC_BIND=false
export OMP_NUM_THREADS=100
export VLLM_USE_V1=1
export HCCL_BUFFSIZE=1024

vllm serve /root/.cache/Qwen3-30B-A3B \
--host 0.0.0.0 \
--port 8004 \
--data-parallel-size 2 \
--data-parallel-size-local 2 \
--data-parallel-address $local_ip \
--data-parallel-rpc-port 13389 \
--seed 1024 \
--served-model-name qwen \
--enable-expert-parallel \
--max-num-seqs 16 \
--max-model-len 32768 \
--max-num-batched-tokens 4096 \
--trust-remote-code \
--no-enable-prefix-caching \
--gpu-memory-utilization 0.9 \
--additional-config '{"ascend_scheduler_config":{"enabled":true},"torchair_graph_config":{"enabled":false}}'

result:
server:

INFO:     127.0.0.1:38768 - "POST /v1/completions HTTP/1.1" 200 OK

client

curl http://127.0.0.1:8004/v1/completions \mpletions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "qwen",
        "prompt": "The future of AI is",
        "max_tokens": 50,
        "temperature": 0
    }'
{"id":"cmpl-5ac0743caa7f4c67aca6582781d07769","object":"text_completion","created":1752805401,"model":"qwen","choices":[{"index":0,"text":" not just about the technology itself, but about how it is used to solve real-world problems. As AI continues to evolve, it will become more integrated into our daily lives, from healthcare and education to transportation and entertainment. The key to unlocking the full","logprobs":null,"finish_reason":"length","stop_reason":null,"prompt_logprobs":null}],"service_tier":null,"system_fingerprint":null,"usage":{"prompt_tokens":5,"total_tokens":55,"completion_tokens":50,"prompt_tokens_details":null},"kv_transfer_params":null}

offline mode:

run offline_data_parallel_script.py

python examples/offline_data_parallel.py \
                --model="/root/.cache/Qwen3-30B-A3B" \
                --dp-size=2 \
                --tp-size=2 \
                --enable-expert-parallel

result:
functional failed

(EngineCore_0 pid=3232)   File "/vllm-workspace/vllm/vllm/v1/engine/core.py", line 596, in run_engine_core
(EngineCore_0 pid=3232)     raise e
(EngineCore_0 pid=3232)   File "/vllm-workspace/vllm/vllm/v1/engine/core.py", line 585, in run_engine_core
(EngineCore_0 pid=3232)     engine_core.run_busy_loop()
(EngineCore_0 pid=3232)   File "/vllm-workspace/vllm/vllm/v1/engine/core.py", line 944, in run_busy_loop
(EngineCore_0 pid=3232)     executed = self._process_engine_step()
(EngineCore_0 pid=3232)   File "/vllm-workspace/vllm/vllm/v1/engine/core.py", line 637, in _process_engine_step
(EngineCore_0 pid=3232)     outputs, model_executed = self.step_fn()
(EngineCore_0 pid=3232)   File "/vllm-workspace/vllm/vllm/v1/engine/core.py", line 241, in step
(EngineCore_0 pid=3232)     model_output = self.execute_model(scheduler_output)
(EngineCore_0 pid=3232)   File "/vllm-workspace/vllm/vllm/v1/engine/core.py", line 227, in execute_model
(EngineCore_0 pid=3232)     raise err
(EngineCore_0 pid=3232)   File "/vllm-workspace/vllm/vllm/v1/engine/core.py", line 218, in execute_model
(EngineCore_0 pid=3232)     return self.model_executor.execute_model(scheduler_output)
(EngineCore_0 pid=3232)   File "/vllm-workspace/vllm/vllm/v1/executor/multiproc_executor.py", line 172, in execute_model
(EngineCore_0 pid=3232)     (output, ) = self.collective_rpc(
(EngineCore_0 pid=3232)   File "/vllm-workspace/vllm/vllm/v1/executor/multiproc_executor.py", line 247, in collective_rpc
(EngineCore_0 pid=3232)     raise TimeoutError(f"RPC call to {method} timed out.") from e
(EngineCore_0 pid=3232) TimeoutError: RPC call to execute_model timed out.

Signed-off-by: MengqingCao <cmq0113@163.com>

MengqingCao · 2025-07-31T01:43:06Z

will add ut on dispatch and combine after #1897

wangxiyuan · 2025-08-01T01:02:12Z

This PR can fix all the MOE model which use common_fused_moe. There is also a PR to make all model to use AscendFusedMoe instead, while I think this PR is good to go first to unblock v0.10.0rc1 release

wangxiyuan · 2025-08-01T01:02:43Z

tests/e2e/long_term/accuracy/accuracy_multicard.py

+            "chatgpt is",
+        ] * 10
+
+        # 并发发送


use english

wangxiyuan · 2025-08-01T01:02:59Z

tests/e2e/long_term/accuracy/accuracy_multicard.py

+                prompt, result = future.result()
+                print(f"> Prompt: {prompt}\nResult: {result}\n")
+
+        # resp = requests.post(COMPLETIONS_URL, json=payload, timeout=30)


remove the uesless code directly

I will update the dp accuracy test using gsm8k, as the outputs between different dp groups don't seem to be exactly the same, although they all look reasonable. thus this will be removed latter

done now, plz take a look again, thanks!

Signed-off-by: MengqingCao <cmq0113@163.com>

MengqingCao · 2025-08-01T02:00:08Z

tests/e2e/long_term/accuracy/accuracy_multicard.py

+    p.join()
+    result = result_queue.get()
+    print(result)
+    assert (EXPECTED_VALUE[model] - RTOL < result < EXPECTED_VALUE[model] + RTOL), \


We use the same EXPECTED_VALUE as that of llm without dp to make sure the accuracy of dp is correct

wangxiyuan · 2025-08-04T02:24:09Z

Let's fixed the bug first. The AscendFuedMOE way can be done in the future.

### What this PR does / why we need it? Fix Qwen MoE accuracy bug in DP scenario. Now the implentment of `FusedMoE` in vLLM use `All2AllManager` to manager different all2all algorithm branch. And the default branch use `Multicast` in `dispatch` phase and `all_reduce` in `combine` phase, which are not implented in vLLM-Ascend. This leading to invoking into a default implentment in `base_communicator`, with empty `dispatch` and `combine` operations, thus causing the accuracy issue on it. This pr is a temporary workaround, refacting all2all in vLLM-Ascend could be a better way. - vLLM version: v0.10.0 - vLLM main: vllm-project/vllm@ad57f23 --------- Signed-off-by: MengqingCao <cmq0113@163.com>

@MengqingCao

I would like to nominate Mengqing Cao (@MengqingCao https://github.com/MengqingCao) as a maintainer, starting with my +1. ## Reason Review Quality‌: She has completed [120+ reviews](https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+is%3Aclosed+commenter%3Amengqingcao+-author%3Amengqingcao) since Feb. 2025, include [#review-3077842852](#2088 (review)), [comment-2990074116](#1032 (comment)), [comment-2921063723](#1013 (comment)) high quality review. Sustained and Quality Contributions: She has Deep understanding of ‌vLLM‌ and ‌vLLM Ascend‌ codebases and solid contributions include The vLLM contributions and help vLLM Ascend release is the main reason I nominated her: - vLLM: Things worth mentioning that she completed [28+ PR contributions](https://github.com/vllm-project/vllm/pulls?q=is%3Apr+author%3AMengqingCao+is%3Amerged+) in vllm-project/vllm, especially for vLLM platform module to improve vLLM mult hardware support. She is one of the important co-authors of [vllm#8054](vllm-project/vllm#8054) and hardware plugin RFC, this makes vllm-ascend plugin possible. Community Involvement: She is also very active and involved in [60+ issues](https://github.com/vllm-project/vllm-ascend/issues?q=is%3Aissue%20state%3Aclosed%20-author%3AMengqingCao%20commenter%3AMengqingCao). So I think she's a great addition to the vLLM Ascend Maintainer team. - ✅**Review Quality‌:** She has completed 120+ reviews since Feb. 2025. https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+is%3Aclosed+commenter%3Amengqingcao+-author%3Amengqingcao, include #2088 (review), #1446 (comment), #1032 (comment), #1013 (comment) quality review. - ✅**Sustained Contributions:** 99+ PR merged in vllm-project/vllm-ascend https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+author%3AMengqingCao+is%3Amerged - ✅**Quality Contribution‌:** She is one of the important co-authors of vllm-project/vllm#8054 , this makes vllm-ascend plugin possible. Things worth mentioning that she complete 28+ PR contributions in vllm-project/vllm, especially for vLLM platform module to improve vLLM mult hardware support: https://github.com/vllm-project/vllm/pulls?q=is%3Apr+author%3AMengqingCao+is%3Amerged+. At 2025 Q2, She also lead the [[RFC]: E2E CI test for key features](#413) and [[RFC]: Unit test coverage improvement](#1298) to help vllm ascend improve the coverage. Her main contributions focus on the adaptation of parallel strategies and communicator, such as #1800, #1856. These contributions are sufficient to prove she has “Deep understanding of ‌vLLM‌ and ‌vLLM Ascend‌ codebases” - ✅**Community Involvement‌:** Involved in 63+ issue reviewer https://github.com/vllm-project/vllm-ascend/issues?q=is%3Aissue%20state%3Aclosed%20-author%3AMengqingCao%20commenter%3AMengqingCao She led the v0.10.1 release as release manager - vLLM version: v0.10.0 - vLLM main: vllm-project/vllm@78dba40 Signed-off-by: Jade Zheng <zheng.shoujian@outlook.com>

@MengqingCao

I would like to nominate Mengqing Cao (@MengqingCao https://github.com/MengqingCao) as a maintainer, starting with my +1. ## Reason Review Quality‌: She has completed [120+ reviews](https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+is%3Aclosed+commenter%3Amengqingcao+-author%3Amengqingcao) since Feb. 2025, include [#review-3077842852](vllm-project#2088 (review)), [comment-2990074116](vllm-project#1032 (comment)), [comment-2921063723](vllm-project#1013 (comment)) high quality review. Sustained and Quality Contributions: She has Deep understanding of ‌vLLM‌ and ‌vLLM Ascend‌ codebases and solid contributions include The vLLM contributions and help vLLM Ascend release is the main reason I nominated her: - vLLM: Things worth mentioning that she completed [28+ PR contributions](https://github.com/vllm-project/vllm/pulls?q=is%3Apr+author%3AMengqingCao+is%3Amerged+) in vllm-project/vllm, especially for vLLM platform module to improve vLLM mult hardware support. She is one of the important co-authors of [vllm#8054](vllm-project/vllm#8054) and hardware plugin RFC, this makes vllm-ascend plugin possible. Community Involvement: She is also very active and involved in [60+ issues](https://github.com/vllm-project/vllm-ascend/issues?q=is%3Aissue%20state%3Aclosed%20-author%3AMengqingCao%20commenter%3AMengqingCao). So I think she's a great addition to the vLLM Ascend Maintainer team. - ✅**Review Quality‌:** She has completed 120+ reviews since Feb. 2025. https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+is%3Aclosed+commenter%3Amengqingcao+-author%3Amengqingcao, include vllm-project#2088 (review), vllm-project#1446 (comment), vllm-project#1032 (comment), vllm-project#1013 (comment) quality review. - ✅**Sustained Contributions:** 99+ PR merged in vllm-project/vllm-ascend https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+author%3AMengqingCao+is%3Amerged - ✅**Quality Contribution‌:** She is one of the important co-authors of vllm-project/vllm#8054 , this makes vllm-ascend plugin possible. Things worth mentioning that she complete 28+ PR contributions in vllm-project/vllm, especially for vLLM platform module to improve vLLM mult hardware support: https://github.com/vllm-project/vllm/pulls?q=is%3Apr+author%3AMengqingCao+is%3Amerged+. At 2025 Q2, She also lead the [[RFC]: E2E CI test for key features](vllm-project#413) and [[RFC]: Unit test coverage improvement](vllm-project#1298) to help vllm ascend improve the coverage. Her main contributions focus on the adaptation of parallel strategies and communicator, such as vllm-project#1800, vllm-project#1856. These contributions are sufficient to prove she has “Deep understanding of ‌vLLM‌ and ‌vLLM Ascend‌ codebases” - ✅**Community Involvement‌:** Involved in 63+ issue reviewer https://github.com/vllm-project/vllm-ascend/issues?q=is%3Aissue%20state%3Aclosed%20-author%3AMengqingCao%20commenter%3AMengqingCao She led the v0.10.1 release as release manager - vLLM version: v0.10.0 - vLLM main: vllm-project/vllm@78dba40 Signed-off-by: Jade Zheng <zheng.shoujian@outlook.com>

### What this PR does / why we need it? Fix Qwen MoE accuracy bug in DP scenario. Now the implentment of `FusedMoE` in vLLM use `All2AllManager` to manager different all2all algorithm branch. And the default branch use `Multicast` in `dispatch` phase and `all_reduce` in `combine` phase, which are not implented in vLLM-Ascend. This leading to invoking into a default implentment in `base_communicator`, with empty `dispatch` and `combine` operations, thus causing the accuracy issue on it. This pr is a temporary workaround, refacting all2all in vLLM-Ascend could be a better way. - vLLM version: v0.10.0 - vLLM main: vllm-project/vllm@ad57f23 --------- Signed-off-by: MengqingCao <cmq0113@163.com>

@MengqingCao

I would like to nominate Mengqing Cao (@MengqingCao https://github.com/MengqingCao) as a maintainer, starting with my +1. ## Reason Review Quality‌: She has completed [120+ reviews](https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+is%3Aclosed+commenter%3Amengqingcao+-author%3Amengqingcao) since Feb. 2025, include [#review-3077842852](vllm-project#2088 (review)), [comment-2990074116](vllm-project#1032 (comment)), [comment-2921063723](vllm-project#1013 (comment)) high quality review. Sustained and Quality Contributions: She has Deep understanding of ‌vLLM‌ and ‌vLLM Ascend‌ codebases and solid contributions include The vLLM contributions and help vLLM Ascend release is the main reason I nominated her: - vLLM: Things worth mentioning that she completed [28+ PR contributions](https://github.com/vllm-project/vllm/pulls?q=is%3Apr+author%3AMengqingCao+is%3Amerged+) in vllm-project/vllm, especially for vLLM platform module to improve vLLM mult hardware support. She is one of the important co-authors of [vllm#8054](vllm-project/vllm#8054) and hardware plugin RFC, this makes vllm-ascend plugin possible. Community Involvement: She is also very active and involved in [60+ issues](https://github.com/vllm-project/vllm-ascend/issues?q=is%3Aissue%20state%3Aclosed%20-author%3AMengqingCao%20commenter%3AMengqingCao). So I think she's a great addition to the vLLM Ascend Maintainer team. - ✅**Review Quality‌:** She has completed 120+ reviews since Feb. 2025. https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+is%3Aclosed+commenter%3Amengqingcao+-author%3Amengqingcao, include vllm-project#2088 (review), vllm-project#1446 (comment), vllm-project#1032 (comment), vllm-project#1013 (comment) quality review. - ✅**Sustained Contributions:** 99+ PR merged in vllm-project/vllm-ascend https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+author%3AMengqingCao+is%3Amerged - ✅**Quality Contribution‌:** She is one of the important co-authors of vllm-project/vllm#8054 , this makes vllm-ascend plugin possible. Things worth mentioning that she complete 28+ PR contributions in vllm-project/vllm, especially for vLLM platform module to improve vLLM mult hardware support: https://github.com/vllm-project/vllm/pulls?q=is%3Apr+author%3AMengqingCao+is%3Amerged+. At 2025 Q2, She also lead the [[RFC]: E2E CI test for key features](vllm-project#413) and [[RFC]: Unit test coverage improvement](vllm-project#1298) to help vllm ascend improve the coverage. Her main contributions focus on the adaptation of parallel strategies and communicator, such as vllm-project#1800, vllm-project#1856. These contributions are sufficient to prove she has “Deep understanding of ‌vLLM‌ and ‌vLLM Ascend‌ codebases” - ✅**Community Involvement‌:** Involved in 63+ issue reviewer https://github.com/vllm-project/vllm-ascend/issues?q=is%3Aissue%20state%3Aclosed%20-author%3AMengqingCao%20commenter%3AMengqingCao She led the v0.10.1 release as release manager - vLLM version: v0.10.0 - vLLM main: vllm-project/vllm@78dba40 Signed-off-by: Jade Zheng <zheng.shoujian@outlook.com>

### What this PR does / why we need it? Fix Qwen MoE accuracy bug in DP scenario. Now the implentment of `FusedMoE` in vLLM use `All2AllManager` to manager different all2all algorithm branch. And the default branch use `Multicast` in `dispatch` phase and `all_reduce` in `combine` phase, which are not implented in vLLM-Ascend. This leading to invoking into a default implentment in `base_communicator`, with empty `dispatch` and `combine` operations, thus causing the accuracy issue on it. This pr is a temporary workaround, refacting all2all in vLLM-Ascend could be a better way. - vLLM version: v0.10.0 - vLLM main: vllm-project/vllm@ad57f23 --------- Signed-off-by: MengqingCao <cmq0113@163.com>

@MengqingCao

I would like to nominate Mengqing Cao (@MengqingCao https://github.com/MengqingCao) as a maintainer, starting with my +1. ## Reason Review Quality‌: She has completed [120+ reviews](https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+is%3Aclosed+commenter%3Amengqingcao+-author%3Amengqingcao) since Feb. 2025, include [#review-3077842852](vllm-project#2088 (review)), [comment-2990074116](vllm-project#1032 (comment)), [comment-2921063723](vllm-project#1013 (comment)) high quality review. Sustained and Quality Contributions: She has Deep understanding of ‌vLLM‌ and ‌vLLM Ascend‌ codebases and solid contributions include The vLLM contributions and help vLLM Ascend release is the main reason I nominated her: - vLLM: Things worth mentioning that she completed [28+ PR contributions](https://github.com/vllm-project/vllm/pulls?q=is%3Apr+author%3AMengqingCao+is%3Amerged+) in vllm-project/vllm, especially for vLLM platform module to improve vLLM mult hardware support. She is one of the important co-authors of [vllm#8054](vllm-project/vllm#8054) and hardware plugin RFC, this makes vllm-ascend plugin possible. Community Involvement: She is also very active and involved in [60+ issues](https://github.com/vllm-project/vllm-ascend/issues?q=is%3Aissue%20state%3Aclosed%20-author%3AMengqingCao%20commenter%3AMengqingCao). So I think she's a great addition to the vLLM Ascend Maintainer team. - ✅**Review Quality‌:** She has completed 120+ reviews since Feb. 2025. https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+is%3Aclosed+commenter%3Amengqingcao+-author%3Amengqingcao, include vllm-project#2088 (review), vllm-project#1446 (comment), vllm-project#1032 (comment), vllm-project#1013 (comment) quality review. - ✅**Sustained Contributions:** 99+ PR merged in vllm-project/vllm-ascend https://github.com/vllm-project/vllm-ascend/pulls?q=is%3Apr+author%3AMengqingCao+is%3Amerged - ✅**Quality Contribution‌:** She is one of the important co-authors of vllm-project/vllm#8054 , this makes vllm-ascend plugin possible. Things worth mentioning that she complete 28+ PR contributions in vllm-project/vllm, especially for vLLM platform module to improve vLLM mult hardware support: https://github.com/vllm-project/vllm/pulls?q=is%3Apr+author%3AMengqingCao+is%3Amerged+. At 2025 Q2, She also lead the [[RFC]: E2E CI test for key features](vllm-project#413) and [[RFC]: Unit test coverage improvement](vllm-project#1298) to help vllm ascend improve the coverage. Her main contributions focus on the adaptation of parallel strategies and communicator, such as vllm-project#1800, vllm-project#1856. These contributions are sufficient to prove she has “Deep understanding of ‌vLLM‌ and ‌vLLM Ascend‌ codebases” - ✅**Community Involvement‌:** Involved in 63+ issue reviewer https://github.com/vllm-project/vllm-ascend/issues?q=is%3Aissue%20state%3Aclosed%20-author%3AMengqingCao%20commenter%3AMengqingCao She led the v0.10.1 release as release manager - vLLM version: v0.10.0 - vLLM main: vllm-project/vllm@78dba40 Signed-off-by: Jade Zheng <zheng.shoujian@outlook.com>

MengqingCao force-pushed the dp_moe branch from e4626e4 to 61d0b15 Compare July 17, 2025 10:39

wangxiyuan reviewed Jul 18, 2025

View reviewed changes

vllm_ascend/distributed/communicator.py Outdated Show resolved Hide resolved

MengqingCao force-pushed the dp_moe branch from 61d0b15 to 755bd0f Compare July 18, 2025 01:58

whx-sjtu mentioned this pull request Jul 30, 2025

[BugFix] Fix accuracy problem in dp situation. #2118

Closed

[MoE][Dist] Fix Qwen MoE accuracy bug in DP scenario

ff33d48

Signed-off-by: MengqingCao <cmq0113@163.com>

MengqingCao force-pushed the dp_moe branch from 755bd0f to ff33d48 Compare July 31, 2025 01:34

MengqingCao marked this pull request as ready for review July 31, 2025 01:35

github-actions bot added the module:tests label Jul 31, 2025

ApsarasX approved these changes Jul 31, 2025

View reviewed changes

wangxiyuan reviewed Aug 1, 2025

View reviewed changes

add test for dp + moe

d768bc2

Signed-off-by: MengqingCao <cmq0113@163.com>

MengqingCao force-pushed the dp_moe branch from b9e03b8 to d768bc2 Compare August 1, 2025 01:56

MengqingCao added accuracy-test enable all accuracy test for PR ready-for-test start test by label for PR labels Aug 1, 2025

MengqingCao commented Aug 1, 2025

View reviewed changes

MengqingCao changed the title ~~[MoE][Dist] Fix Qwen MoE accuracy bug in DP senario~~ [MoE][Dist] Fix Qwen MoE accuracy bug in DP scenario Aug 1, 2025

wangxiyuan approved these changes Aug 4, 2025

View reviewed changes

wangxiyuan merged commit af04ee9 into vllm-project:main Aug 4, 2025
39 checks passed

jianzs mentioned this pull request Aug 19, 2025

Nominate Mengqing Cao as vllm-ascend maintainer #2433

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[MoE][Dist] Fix Qwen MoE accuracy bug in DP scenario #1856

[MoE][Dist] Fix Qwen MoE accuracy bug in DP scenario #1856

MengqingCao commented Jul 17, 2025 •

edited

Loading

Uh oh!

codecov bot commented Jul 17, 2025 •

edited

Loading

Uh oh!

Uh oh!

Potabk commented Jul 18, 2025

Uh oh!

Potabk commented Jul 18, 2025

Uh oh!

MengqingCao commented Jul 31, 2025

Uh oh!

wangxiyuan commented Aug 1, 2025

Uh oh!

wangxiyuan Aug 1, 2025

Uh oh!

wangxiyuan Aug 1, 2025

Uh oh!

MengqingCao Aug 1, 2025

Uh oh!

MengqingCao Aug 1, 2025

Uh oh!

MengqingCao Aug 1, 2025 •

edited

Loading

Uh oh!

wangxiyuan commented Aug 4, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

[MoE][Dist] Fix Qwen MoE accuracy bug in DP scenario #1856

[MoE][Dist] Fix Qwen MoE accuracy bug in DP scenario #1856

Conversation

MengqingCao commented Jul 17, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What this PR does / why we need it?

Does this PR introduce any user-facing change?

How was this patch tested?

Uh oh!

codecov bot commented Jul 17, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

Uh oh!

Potabk commented Jul 18, 2025

Uh oh!

Potabk commented Jul 18, 2025

online:

offline mode:

Uh oh!

MengqingCao commented Jul 31, 2025

Uh oh!

wangxiyuan commented Aug 1, 2025

Uh oh!

wangxiyuan Aug 1, 2025

Choose a reason for hiding this comment

Uh oh!

wangxiyuan Aug 1, 2025

Choose a reason for hiding this comment

Uh oh!

MengqingCao Aug 1, 2025

Choose a reason for hiding this comment

Uh oh!

MengqingCao Aug 1, 2025

Choose a reason for hiding this comment

Uh oh!

MengqingCao Aug 1, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

wangxiyuan commented Aug 4, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

MengqingCao commented Jul 17, 2025 •

edited

Loading

codecov bot commented Jul 17, 2025 •

edited

Loading

MengqingCao Aug 1, 2025 •

edited

Loading