[Sync] Upstream 20250910 commit e408272

[tjtanaavllm]

Purpose

Sync with upstream to get latest changes and to fix the Compressed Tensor FP8 weight loading accuracy issue.

Upgrade vLLM version to 0.10.2rc2.dev+ge408272

Test Plan

Validated on model of interests

1. Qwen/Qwen3-235B-A22B-FP8

MODEL=Qwen/Qwen3-235B-A22B-FP8


VLLM_USE_V1=1
VLLM_ROCM_USE_AITER=1 \
vllm serve $MODEL \
--tensor-parallel-size 8 \
--max-num-batched-tokens 32768 \
--disable-log-requests \
--kv-cache-dtype fp8 \
--compilation-config '{"cudagraph_mode": "FULL_AND_PIECEWISE"}' \
--trust-remote-code \
--enable_expert_parallel \
--port 6789 \
> server-Qwen_Qwen3-235B-A22B-FP8-aiter-v1-fp8-cudagraph_FULL.log 2>&1

2. Qwen/Qwen2.5-VL-72B-Instruct

VLLM_RPC_TIMEOUT=1800000 \
VLLM_USE_V1=1 \
VLLM_ROCM_USE_AITER=1 \
VLLM_ROCM_USE_AITER_MHA=1 \
VLLM_USE_TRITON_FLASH_ATTN=0 \
SAFETENSORS_FAST_GPU=1 \
vllm serve Qwen/Qwen2.5-VL-72B-Instruct \
 -tp 8 \
 --compilation-config '{"cudagraph_mode": "FULL_AND_PIECEWISE"}' \
 --trust_remote_code \
 --mm-processor-kwargs '{"max_pixels":802816,"min_pixels":3136}' \
 --limit-mm-per-prompt='{"image": 64}' \
 --mm-encoder-tp-mode "data" \
> server_Qwen_Qwen2.5-VL-72B-Instruct-aiter-v1-tp8-dp8-cudagraph_FULL_AND_PIECEWISE.log 2>&1

3. Qwen/Qwen2.5-VL-3B-Instruct

VLLM_RPC_TIMEOUT=1800000 \
VLLM_USE_V1=1 \
VLLM_ROCM_USE_AITER=1 \
VLLM_ROCM_USE_AITER_MHA=1 \
VLLM_USE_TRITON_FLASH_ATTN=0 \
SAFETENSORS_FAST_GPU=1 \
vllm serve Qwen/Qwen2.5-VL-3B-Instruct \
 -tp 1 \
 --compilation-config '{"cudagraph_mode": "FULL_AND_PIECEWISE"}' \
 --trust_remote_code \
 --mm-processor-kwargs '{"max_pixels":802816,"min_pixels":3136}' \
 --limit-mm-per-prompt='{"image": 64}' \
> server_Qwen_Qwen2.5-VL-3B-Instruct-aiter-v1-tp1-cudagraph_FULL_AND_PIECEWISE.log 2>&1

Test Result

1. Qwen/Qwen3-235B-A22B-FP8

local-completions (model=Qwen/Qwen3-235B-A22B-FP8,base_url=http://127.0.0.1:6789/v1/completions), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: 100
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.8423|±  |0.0100|
|     |       |strict-match    |     5|exact_match|↑  |0.8150|±  |0.0107|

2. Qwen/Qwen2.5-VL-72B-Instruct

For detailed information on this command, run:
  run.py eval_vllm --model_name Qwen/Qwen2.5-VL-72B-Instruct --url http://0.0.0.0:8000 --output_dir ./chartqa --eval_name chartqa - --help
================================================================================
Metrics:
{
    "explicit_prompt_relaxed_correctness": 0.8684,
    "anywhere_in_answer_relaxed_correctness": 0.8852
}
================================================================================

3. Qwen/Qwen2.5-VL-3B-Instruct

For detailed information on this command, run:
  run.py eval_vllm --model_name Qwen/Qwen2.5-VL-3B-Instruct --url http://0.0.0.0:8000 --output_dir ./chartqa --eval_name chartqa - --help
================================================================================
Metrics:
{
    "explicit_prompt_relaxed_correctness": 0.8104,
    "anywhere_in_answer_relaxed_correctness": 0.8144
}
================================================================================

---

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.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[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.

🚀