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 Failure] fix models/language/pooling/test_auto_prefix_cache_support.py (#24636)
Signed-off-by: wang.yuqi <noooop@126.com>
wang.yuqi committed
Sep 11, 2025 at 18:31 UTC
25bb9e8c65424e3bf24d2eab259743f9a97b7a3c
1 file changed
+4
vllm/config/__init__.py
+4
@@ -3558,6 +3558,10 @@ class VllmConfig:
3558
disable_chunked_prefill_reasons.append(
3559
"Only \"last\" pooling supports chunked "
3560
"prefill and prefix caching; disabling both.")
3561
+ if not getattr(self.model_config.hf_config, "is_causal", True):
3562
+ disable_chunked_prefill_reasons.append(
3563
+ "Only models using causal attention supports chunked "
3564
+ "prefill and prefix caching; disabling both.")
3565
elif self.model_config.is_encoder_decoder:
3566
self.scheduler_config.max_num_encoder_input_tokens = \
3567
MULTIMODAL_REGISTRY.get_encdec_max_encoder_len(self.model_config)