[Core] Don't use cache during multi-modal profiling (#14336)
This commit is contained in:
@@ -331,7 +331,9 @@ class InputRegistry:
|
||||
|
||||
if mm_registry.has_processor(model_config):
|
||||
tokenizer = cached_tokenizer_from_config(model_config)
|
||||
processor = mm_registry.create_processor(model_config, tokenizer)
|
||||
processor = mm_registry.create_processor(model_config,
|
||||
tokenizer,
|
||||
disable_cache=True)
|
||||
profiler = MultiModalProfiler(processor)
|
||||
dummy_data = profiler.get_dummy_data(
|
||||
seq_len, is_encoder_data=is_encoder_data)
|
||||
|
||||
Reference in New Issue
Block a user