vllm-project
diff --git a/‎docs/design/feature/async_chunk_design.md‎
Lines changed: 71 additions & 319 deletions b/‎docs/design/feature/async_chunk_design.md‎
Lines changed: 71 additions & 319 deletions
diff --git a/‎docs/source/architecture/async-chunk-architecture.png‎
50.6 KB b/‎docs/source/architecture/async-chunk-architecture.png‎
50.6 KB
diff --git a/‎docs/source/performance/qwen3-omni_tpot_performance.png‎
151 KB b/‎docs/source/performance/qwen3-omni_tpot_performance.png‎
151 KB
diff --git a/‎docs/source/performance/qwen3-omni_ttfp_performance.png‎
144 KB b/‎docs/source/performance/qwen3-omni_ttfp_performance.png‎
144 KB
diff --git a/‎docs/source/performance/qwen3-omni_ttft_performance.png‎
143 KB b/‎docs/source/performance/qwen3-omni_ttft_performance.png‎
143 KB
diff --git a/‎vllm_omni/benchmarks/patch/patch.py‎
Lines changed: 2 additions & 3 deletions b/‎vllm_omni/benchmarks/patch/patch.py‎
Lines changed: 2 additions & 3 deletions
@@ -150,9 +150,8 @@ async def async_request_openai_chat_omni_completions(
                                     generated_text += content or ""
                                 elif modality == "audio":
                                     if output.audio_ttfp == 0.0:
-                                        audio_first_timestamp = timestamp
-                                        output.audio_ttfp = timestamp - st
-                                    audio_generate_time = timestamp - audio_first_timestamp
+                                        output.audio_ttfp = timestamp - st                                    
+                                    audio_generate_time = timestamp - st
                                     if content != "":
                                         audio_bytes = base64.b64decode(content)
                                         seg = AudioSegment.from_file(io.BytesIO(audio_bytes))