[Bugfix] [CI] [ROCm] [CUDA] Fix voxtral test (#4380)

Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>

.buildkite/test-amd-merge.yml

@@ -179,17 +179,30 @@ steps:
             pytest -s -v tests/e2e/online_serving/test_qwen3_tts_base.py tests/e2e/offline_inference/test_qwen3_tts_base.py -m "advanced_model" --run-level "advanced_model"
           '
 
-    - label: "Voxtral-TTS E2E Test"
+    # It is the whisper ASR validation step that is slow, not the Voxtral TTS generation.
+    # whisper ASR runs on CPU.
+    - label: "Voxtral-TTS E2E Test (Online Serving)"
       agent_pool: mi325_1
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
       grade: Blocking
       commands:
         - |
-          timeout 20m bash -c '
-            export VLLM_LOGGING_LEVEL=DEBUG
+          timeout 40m bash -c '
             export VLLM_WORKER_MULTIPROC_METHOD=spawn
-            pytest -s -v tests/e2e/online_serving/test_voxtral_tts.py tests/e2e/offline_inference/test_voxtral_tts.py -m "advanced_model" --run-level "advanced_model"
+            pytest -s -v tests/e2e/online_serving/test_voxtral_tts.py -m "advanced_model" --run-level "advanced_model"
+          '
+
+    - label: "Voxtral-TTS E2E Test (Offline Inference)"
+      agent_pool: mi325_1
+      depends_on: amd-build
+      mirror_hardwares: [amdproduction]
+      grade: Blocking
+      commands:
+        - |
+          timeout 30m bash -c '
+            export VLLM_WORKER_MULTIPROC_METHOD=spawn
+            pytest -s -v tests/e2e/offline_inference/test_voxtral_tts.py -m "advanced_model" --run-level "advanced_model"
           '
 
 # TODO: Bagel test on ROCm is very unstable. @tjtanaa

.buildkite/test-amd-ready.yaml

@@ -185,20 +185,6 @@ steps:
         - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
         - timeout 30m pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py -m "core_model" --run-level "core_model"
 
-    - label: "Voxtral-TTS E2E Test"
-      agent_pool: mi325_1
-      depends_on: amd-build
-      mirror_hardwares: [amdproduction]
-      grade: Blocking
-      commands:
-        - |
-          timeout 20m bash -c '
-            export VLLM_LOGGING_LEVEL=DEBUG
-            export VLLM_WORKER_MULTIPROC_METHOD=spawn
-            pytest -s -v tests/e2e/online_serving/test_voxtral_tts.py -m "advanced_model" --run-level "advanced_model"
-            pytest -s -v tests/e2e/offline_inference/test_voxtral_tts.py -m "advanced_model" --run-level "advanced_model"
-          '
-
     - label: "CosyVoice3-TTS E2E Test"
       agent_pool: mi325_1
       depends_on: amd-build

tests/distributed/omni_connectors/test_shm_connector.py

@@ -41,6 +41,19 @@ class TestKeyBasedReadWrite:
         result = connector.get("s0", "s1", "no_such_key_xyz", metadata=None)
         assert result is None
 
+    def test_get_empty_shm_race_returns_none(self, connector, monkeypatch):
+        def raise_empty_file(*args, **kwargs):
+            raise ValueError("cannot mmap an empty file")
+
+        monkeypatch.setattr(
+            "vllm_omni.distributed.omni_connectors.connectors.shm_connector.shm_pkg.SharedMemory",
+            raise_empty_file,
+        )
+
+        result = connector.get("s0", "s1", "not_ready_yet", metadata=None)
+
+        assert result is None
+
     def test_rank_aware_keys_independent(self, connector):
         """Each TP rank writes/reads its own key — simulates homogeneous TP."""
         payloads = {}

vllm_omni/distributed/omni_connectors/connectors/shm_connector.py

@@ -116,6 +116,14 @@ class SharedMemoryConnector(OmniConnectorBase):
             return result
         except FileNotFoundError:
             return None
+        except ValueError as e:
+            # A receiver can observe a newly-created POSIX SHM object before
+            # the writer has finished sizing it. Treat that as "not ready yet"
+            # so async polling can retry without a traceback.
+            if "empty file" in str(e):
+                return None
+            logger.debug("_get_by_key: unexpected error reading SHM segment %s", get_key, exc_info=True)
+            return None
         except Exception:
             logger.debug("_get_by_key: unexpected error reading SHM segment %s", get_key, exc_info=True)
             return None

vllm_omni/entrypoints/openai/audio_utils_mixin.py

@@ -96,12 +96,16 @@ class AudioMixin:
             else:
                 waveform = torch.from_numpy(audio_tensor).unsqueeze(0)
 
-            # Match librosa.stft defaults: n_fft=2048, hop_length=n_fft//4
-            n_fft = 2048
+            # Use a speech-sized analysis window. The previous 2048-sample
+            # window is tuned for music and can smear short consonants after
+            # aggressive compression, which makes ASR transcript checks flaky.
+            n_fft = 768
             hop_length = n_fft // 4
+            window = torch.hann_window(n_fft, device=waveform.device, dtype=waveform.dtype)
             to_spec = torchaudio.transforms.Spectrogram(
                 n_fft=n_fft,
                 hop_length=hop_length,
+                window_fn=lambda *_args, **_kwargs: window,
                 power=None,
             )
             stretch = torchaudio.transforms.TimeStretch(
@@ -111,6 +115,7 @@ class AudioMixin:
             to_wave = torchaudio.transforms.InverseSpectrogram(
                 n_fft=n_fft,
                 hop_length=hop_length,
+                window_fn=lambda *_args, **_kwargs: window,
             )
 
             spec = to_spec(waveform)