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 9, 07:30 UTC
[CI] Qwen3-ASR transcriptios tests (#33414)
Signed-off-by: NickLucche <nlucches@redhat.com>
Nicolò Lucchesi committed
Jan 30, 2026 at 17:17 UTC
8ece60768f63989db7f26ef81c7f68635f70e081
1 file changed
+9
-5
tests/entrypoints/openai/test_transcription_validation.py
+9
-5
@@ -20,7 +20,9 @@ MISTRAL_FORMAT_ARGS = [
20
21
22
@pytest.mark.asyncio
23
-@pytest.mark.parametrize("model_name", ["mistralai/Voxtral-Mini-3B-2507"])
23
+@pytest.mark.parametrize(
24
+ "model_name", ["mistralai/Voxtral-Mini-3B-2507", "Qwen/Qwen3-ASR-0.6B"]
25
+)
26
async def test_basic_audio(mary_had_lamb, model_name, rocm_aiter_fa_attention):
27
server_args = ["--enforce-eager"]
28
@@ -42,7 +44,7 @@ async def test_basic_audio(mary_had_lamb, model_name, rocm_aiter_fa_attention):
44
out = json.loads(transcription)
45
out_text = out["text"]
46
out_usage = out["usage"]
45
- assert "Mary had a little lamb," in out_text
47
+ assert "Mary had a little lamb" in out_text
48
assert out_usage["seconds"] == 16, out_usage["seconds"]
49
50
@@ -89,10 +91,12 @@ async def test_basic_audio_with_lora(mary_had_lamb, rocm_aiter_fa_attention):
91
92
93
@pytest.mark.asyncio
92
-async def test_basic_audio_gemma(foscolo, rocm_aiter_fa_attention):
94
+@pytest.mark.parametrize(
95
+ "model_name", ["google/gemma-3n-E2B-it", "Qwen/Qwen3-ASR-0.6B"]
96
+)
97
+async def test_basic_audio_foscolo(foscolo, rocm_aiter_fa_attention, model_name):
98
# Gemma accuracy on some of the audio samples we use is particularly bad,
99
# hence we use a different one here. WER is evaluated separately.
95
- model_name = "google/gemma-3n-E2B-it"
100
server_args = ["--enforce-eager"]
101
102
add_attention_backend(server_args, rocm_aiter_fa_attention)
@@ -109,4 +113,4 @@ async def test_basic_audio_gemma(foscolo, rocm_aiter_fa_attention):
113
temperature=0.0,
114
)
115
out = json.loads(transcription)["text"]
112
- assert "da cui vergine nacque Venere" in out
116
+ assert "ove il mio corpo fanciulletto giacque" in out