diff --git a/tests/models/language/generation_ppl_test/test_qwen.py b/tests/models/language/generation_ppl_test/test_qwen.py index 60e69c3f87..bad30fa50d 100644 --- a/tests/models/language/generation_ppl_test/test_qwen.py +++ b/tests/models/language/generation_ppl_test/test_qwen.py @@ -21,6 +21,9 @@ def test_ppl(hf_runner, vllm_runner, model_info: GenerateModelInfo): vllm_extra_kwargs = {} if model_info.name == "Qwen/Qwen3.5-0.8B": vllm_extra_kwargs["language_model_only"] = True + # Qwen3.5 is a hybrid model; its default mamba cache mode ('align') + # for prefix caching requires chunked prefill. + vllm_extra_kwargs["enable_chunked_prefill"] = True wikitext_ppl_test( hf_runner, vllm_runner, model_info, vllm_extra_kwargs=vllm_extra_kwargs