[Misc] Avoid direct access of global mm_registry in compute_encoder_budget (#15621)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

vllm/v1/worker/gpu_model_runner.py

@@ -137,6 +137,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
         encoder_compute_budget, encoder_cache_size = compute_encoder_budget(
             model_config=model_config,
             scheduler_config=scheduler_config,
+            mm_registry=self.mm_registry,
         )
         self.max_num_encoder_input_tokens = encoder_compute_budget
         self.encoder_cache_size = encoder_cache_size
@@ -1439,9 +1440,8 @@ class GPUModelRunner(LoRAModelRunnerMixin):
             # NOTE: Currently model is profiled with a single non-text
             # modality with the max possible input tokens even when
             # it supports multiple.
-            max_tokens_by_modality_dict = (
-                MULTIMODAL_REGISTRY.
-                get_max_tokens_per_item_by_nonzero_modality(self.model_config))
+            max_tokens_by_modality_dict = self.mm_registry \
+                .get_max_tokens_per_item_by_nonzero_modality(self.model_config)
             dummy_data_modality, max_tokens_per_mm_item = max(
                 max_tokens_by_modality_dict.items(), key=lambda item: item[1])
 

vllm/v1/worker/tpu_model_runner.py

@@ -109,6 +109,7 @@ class TPUModelRunner:
         encoder_compute_budget, encoder_cache_size = compute_encoder_budget(
             model_config=model_config,
             scheduler_config=scheduler_config,
+            mm_registry=self.mm_registry,
         )
         self.max_num_encoder_input_tokens = encoder_compute_budget
         self.encoder_cache_size = encoder_cache_size