A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 7, 20:15 UTC
[glm-asr] change defaults dummy audio size (#35108)
Signed-off-by: Eustache Le Bihan <eulebihan@gmail.com>
eustlb committed
Feb 24, 2026 at 13:13 UTC
b3ad37c5db8240e79195f644a3148a989e8e9422
1 file changed
+10
-2
tests/models/multimodal/processing/test_common.py
+10
-2
@@ -333,10 +333,12 @@ def _test_processing_correctness(
333
334
rng = np.random.RandomState(0)
335
336
+ # GLM-ASR requires a minimum audio length of 70ms
337
+ min_audio_len = 512 if model_config.hf_config.model_type != "glmasr" else 1120
338
input_to_hit = {
339
"image": Image.new("RGB", size=(128, 128)),
340
"video": np.zeros((4, 128, 128, 3), dtype=np.uint8),
339
- "audio": (np.zeros((512,)), 16000),
341
+ "audio": (np.zeros((min_audio_len,)), 16000),
342
"vision_chunk": {"type": "image", "image": Image.new("RGB", size=(128, 128))},
343
}
344
input_factory = {
@@ -344,7 +346,13 @@ def _test_processing_correctness(
346
"video": partial(
347
random_video, rng, min_frames=2, max_frames=16, min_wh=128, max_wh=256
348
),
347
- "audio": partial(random_audio, rng, min_len=512, max_len=1024, sr=16000),
349
+ "audio": partial(
350
+ random_audio,
351
+ rng,
352
+ min_len=min_audio_len,
353
+ max_len=min_audio_len + 512,
354
+ sr=16000,
355
+ ),
356
"vision_chunk": partial(
357
random_vision_chunk, rng, min_wh=128, max_wh=256, min_frames=1, max_frames=1
358
),