vllm-project/vllm-omni
[Bug]: Nightly CI, Qwen3-Omni-30B-A3B-Instruct, performance metrics regressed by more than 10% compared to the baseline in some scenarios
Fermée
#4 965 ouverte le 8 juil. 2026
bugci-failurehelp wantedhigh priority
Métriques du dépôt
- Stars
- (4 990 étoiles)
- Métriques de merge PR
- (Métriques PR en attente)
Description
Your current environment
CI env
Your code version
0.24.0
7ab50965e8d3e64507d0504d706368f7e8318cf5
🐛 Describe the bug
| Source | Model | Type | Config | Test | Metric | latest | baseline | vs baseline | Status | Days failing |
|---|---|---|---|---|---|---|---|---|---|---|
| Buildkite | Qwen/Qwen3-Omni-30B-A3B-Instruct | qwen3_omni | data=random, in=2500, out=900, c=8, n=32, profile=audio_metrics | qwen3_omni_chunk | mean_ttft_ms | 432.5542 | 271.9025 | -59.08% | fail | 3 days+ |
| Buildkite | Qwen/Qwen3-Omni-30B-A3B-Instruct | qwen3_omni | data=random, in=2500, out=900, c=8, n=32, profile=text_only_metrics | qwen3_omni_chunk | mean_ttft_ms | 458.9636 | 326.5864 | -40.53% | fail | 3 days+ |
| Buildkite | Qwen/Qwen3-Omni-30B-A3B-Instruct | qwen3_omni | data=random-mm, in=100, out=100, n=10, qbs=0.1, profile=text_only_metrics | qwen3_omni_chunk | mean_ttft_ms | 429.4516 | 318.048 | -35.03% | fail | 3 days+ |
| Buildkite | Qwen/Qwen3-Omni-30B-A3B-Instruct | qwen3_omni | data=random-mm, in=100, out=100, n=100, qbs=1.0, profile=text_only_metrics | qwen3_omni_chunk | mean_ttft_ms | 470.5099 | 375.5282 | -25.29% | fail | 3 days+ |
| Buildkite | Qwen/Qwen3-Omni-30B-A3B-Instruct | qwen3_omni | data=random-mm, in=100, out=100, n=40, qbs=0.5, profile=text_only_metrics | qwen3_omni_chunk | mean_ttft_ms | 232.5628 | 186.168 | -24.92% | fail | 3 days+ |
| Buildkite | Qwen/Qwen3-Omni-30B-A3B-Instruct | qwen3_omni | data=random, in=2500, out=900, c=32, n=128, profile=text_only_metrics | qwen3_omni_chunk | mean_ttft_ms | 1096.2507 | 961.3929 | -14.03% | fail | 3 days+ |
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.