Issue Station
Sign In
Tools hereToolbox

vllm - ✅(Solved) Fix [Bug]: Empty output when using FP8 + Tensor Parallel (2 GPUs) with Qwen3-8B [1 pull requests, 1 comments, 2 participants]

Official PRs (…)
ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

Helpful · Quick feedback

Loading…
GitHub stats
vllm-project/vllm#36583•Fetched 2026-04-08 00:36:13
View on GitHub
Comments
1
Participants
2
Timeline
3
Reactions
0
Author
Cqh123-HCqh123-H
Participants
Cqh123-HCqh123-Hflutistflutist
Timeline (top)
commented ×1cross-referenced ×1labeled ×1

Fix Action

Fixed

PR fix notes

PR #36621: [Bugfix] Fix FP8 online quantization premature trigger with TP sharded weights

Description (problem / solution / changelog)

Use >= instead of == for loaded numel check to guard against edge cases in copy_ tracking with TP > 1.

Closes #36583

Changed files

  • tests/quantization/test_fp8.py (modified, +144/-0)
  • vllm/model_executor/layers/quantization/fp8.py (modified, +12/-4)

Code Example

Your output of `python collect_env.py` here
RAW_BUFFERClick to expand / collapse

Your current environment

vllm_env.txt

<details> <summary>The output of <code>python collect_env.py</code></summary> 
Your output of `python collect_env.py` here
</details>

🐛 Describe the bug

Describe the bug

When serving vLLM with FP8 quantization and tensor parallelism (2 GPUs) using Qwen3-8B, the model returns empty output.

However:

  • Single GPU + FP8 works correctly
  • Two GPUs without FP8 also works correctly The issue only happens when FP8 quantization is enabled together with tensor parallelism.

Environment

  • vLLM version: 0.13.0
  • GPUs: 2 GPUs
  • CUDA: (please fill if needed)
  • Model: Qwen3-8B

Reproduction

Start server

CUDA_VISIBLE_DEVICES=0,6 vllm serve /new_nfs/models/Qwen3-8B/ \ -tp 2 \ --port 15973 \ --served-model-name "qwen3-8b" \ --gpu-memory-utilization 0.9 \ -q fp8 \ --enable-chunked-prefill \ --enforce-eager \ --enable-prefix-caching \ --kv-cache-dtype fp8_e4m3

Test script

`from openai import OpenAI

client = OpenAI( base_url="http://xxx:15973/v1", api_key="EMPTY", )

response = client.chat.completions.create( model="qwen3-8b", messages=[{"role": "user", "content": "Hello, world!"}], )

print(response) print(response.choices[0].message.content)`

Actual behavior

The model returns empty output:

ChatCompletion( choices=[ Choice( finish_reason='stop', message=ChatCompletionMessage( content='' ), stop_reason=151643 ) ], usage=CompletionUsage( completion_tokens=1, prompt_tokens=12, total_tokens=13 ) )

Only 1 token is generated and the response content is empty.

Additional observations

The issue appears only when FP8 quantization and tensor parallelism are used together.

ConfigurationResult
1 GPU + FP8✅ Works
2 GPUs (TP=2) without FP8✅ Works
2 GPUs (TP=2) + FP8❌ Empty output

If needed, I can provide additional logs or debugging information.

Before submitting a new issue...

  • Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

extent analysis

Fix Plan

To resolve the issue with empty output when using FP8 quantization and tensor parallelism, we need to adjust the configuration and potentially modify the code. Here are the steps:

  • Update vLLM configuration:
    • Set --gpu-memory-utilization to a lower value (e.g., 0.7) to reduce memory usage and alleviate potential memory constraints.
    • Disable --enable-chunked-prefill and --enable-prefix-caching to simplify the serving process and reduce memory usage.
  • Modify the test script:
    • Increase the max_tokens parameter in the chat.completions.create method to allow for longer responses (e.g., max_tokens=1024).
  • Code changes:
    • No code changes are required in this case, as the issue appears to be related to configuration and memory usage.

Example code snippet for the modified test script:

from openai import OpenAI

client = OpenAI(
    base_url="http://xxx:15973/v1",
    api_key="EMPTY",
)

response = client.chat.completions.create(
    model="qwen3-8b",
    messages=[{"role": "user", "content": "Hello, world!"}],
    max_tokens=1024,  # Increased max tokens
)

print(response)
print(response.choices[0].message.content)

Verification

To verify that the fix worked, run the modified test script and check the response content. If the issue is resolved, the response should contain a non-empty output.

Extra Tips

  • Monitor GPU memory usage and adjust the --gpu-memory-utilization value as needed to prevent memory constraints.
  • Consider disabling FP8 quantization or reducing the model size to alleviate memory usage and improve performance.
  • If issues persist, provide additional logs and debugging information to further investigate the problem.

Vote matrix · Quick signals

Works
Did the solution work? Tap to confirm.
Easy Fix
Was it a quick fix?
Time Saver
Did it save you time?
Blocking
Was it severely blocking?
Common Issue
Are others likely hitting this too?
Flaky / Intermittent
Is it intermittent?
Verified / Reproducible
Can you reproduce it reliably?
Loading…
#api#ssr#installation#tensor shape#autograd error#inference speed#output truncation#response parsing#generation error

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Back to top recommendations

TRENDING

vllm - ✅(Solved) Fix [Bug]: Empty output when using FP8 + Tensor Parallel (2 GPUs) with Qwen3-8B [1 pull requests, 1 comments, 2 participants]