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 9, 07:30 UTC
Fix encoder-decoder model disabling mm processor cache (#33236)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Harry Mellor committed
Jan 30, 2026 at 16:30 UTC
67239c4c4235ae668971625304390ed2078b4dd0
2 files changed
+4
-3
tests/models/multimodal/processing/test_common.py
+3
-2
@@ -241,14 +241,15 @@ def _test_processing_correctness(
241
revision=model_info.revision,
242
trust_remote_code=model_info.trust_remote_code,
243
hf_overrides=model_info.hf_overrides,
244
- # Ensure that the cache can fit all of the data
245
- mm_processor_cache_gb=2048,
244
skip_tokenizer_init=model_info.require_embed_inputs,
245
enable_prompt_embeds=model_info.require_embed_inputs,
246
enable_mm_embeds=model_info.require_embed_inputs,
247
enforce_eager=model_info.enforce_eager,
248
dtype=model_info.dtype,
249
)
250
+ # Ensure that the cache can fit all of the data
251
+ # (set after because ModelConfig would set it to 0 for encoder-decoder models)
252
+ model_config.multimodal_config.mm_processor_cache_gb = 2048
253
254
model_cls = MULTIMODAL_REGISTRY._get_model_cls(model_config)
255
factories = model_cls._processor_factory
vllm/config/model.py
+1
-1
@@ -572,7 +572,7 @@ class ModelConfig:
572
self.max_model_len = self.get_and_verify_max_len(self.max_model_len)
573
574
if self.is_encoder_decoder:
575
- self.mm_processor_cache_gb = 0
575
+ mm_processor_cache_gb = 0
576
logger.info("Encoder-decoder model detected, disabling mm processor cache.")
577
578
# Init multimodal config if needed