[Bugfix] Enable chunked prefill for qwen3.5-0.8B ppl test (#51153)
Signed-off-by: Dino Music <Dino.Music@amd.com>
music-dino committed
Aug 5, 2026 at 15:21 UTC
a9b39d6dcbe9970a33b5d1ceb9848ba975167b1b
1 file changed
+3
tests/models/language/generation_ppl_test/test_qwen.py
+3
@@ -21,6 +21,9 @@ def test_ppl(hf_runner, vllm_runner, model_info: GenerateModelInfo):
21
vllm_extra_kwargs = {}
22
if model_info.name == "Qwen/Qwen3.5-0.8B":
23
vllm_extra_kwargs["language_model_only"] = True
24
+ # Qwen3.5 is a hybrid model; its default mamba cache mode ('align')
25
+ # for prefix caching requires chunked prefill.
26
+ vllm_extra_kwargs["enable_chunked_prefill"] = True
27
28
wikitext_ppl_test(
29
hf_runner, vllm_runner, model_info, vllm_extra_kwargs=vllm_extra_kwargs