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 14, 05:30 UTC
[CI] Adjust threshold for flaky ngram spec decoding test (#24528)
Signed-off-by: Nick Hill <nhill@redhat.com>
Nick Hill committed
Sep 9, 2025 at 21:07 UTC
83dd28aae4c3cd8442f5be42e0d77263b68c4236
1 file changed
+2
-2
tests/v1/e2e/test_spec_decode.py
+2
-2
@@ -117,9 +117,9 @@ def test_ngram_correctness(
117
print(f"ref_output: {ref_output.outputs[0].text}")
118
print(f"spec_output: {spec_output.outputs[0].text}")
119
120
- # Heuristic: expect at least 70% of the prompts to match exactly
120
+ # Heuristic: expect at least 68% of the prompts to match exactly
121
# Upon failure, inspect the outputs to check for inaccuracy.
122
- assert matches > int(0.7 * len(ref_outputs))
122
+ assert matches >= int(0.68 * len(ref_outputs))
123
del spec_llm
124
torch.cuda.empty_cache()
125
cleanup_dist_env_and_memory()