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 4, 10:00 UTC
Limit concurrency on `test_transcription_api_correctness.py` (#41478)
Signed-off-by: Ekagra Ranjan <3116519+ekagra-ranjan@users.noreply.github.com>
Ekagra Ranjan committed
May 1, 2026 at 23:19 UTC
3e49479c4b766a601804f0c6f5f1c9a3def5ad0c
1 file changed
+4
tests/entrypoints/openai/correctness/test_transcription_api_correctness.py
+4
@@ -26,6 +26,9 @@ from vllm.tokenizers import get_tokenizer
26
from ....models.registry import HF_EXAMPLE_MODELS
27
from ....utils import RemoteOpenAIServer
28
29
+# Tuned to prevent OOM on 18GB GPUs in transcription correctness tests.
30
+MAX_SEQS_FOR_TRANSCRIPTION_TEST = 32
31
+
32
33
def to_bytes(y, sr):
34
buffer = io.BytesIO()
@@ -184,6 +187,7 @@ def test_wer_correctness(
187
server_args = [
188
"--enforce-eager",
189
f"--tokenizer_mode={model_info.tokenizer_mode}",
190
+ f"--max_num_seqs={MAX_SEQS_FOR_TRANSCRIPTION_TEST}",
191
]
192
if model_info.trust_remote_code:
193
server_args.append("--trust-remote-code")