tweak and clarify

Signed-off-by: Roger Wang <[email protected]>

vllm/v1/worker/gpu_model_runner.py

@@ -647,11 +647,18 @@ def profile_run(self) -> None:
                 self.mm_registry.get_max_tokens_per_item_by_modality(
                     self.model_config).values())
 
-            max_num_mm_items = min(
+            max_num_mm_items_encoder_budget = min(
                 self.max_num_encoder_input_tokens,
                 self.encoder_cache_size) // max_tokens_per_mm_item
 
-            max_num_mm_items = min(self.max_num_reqs, max_num_mm_items)
+            max_mm_items_per_req = max(
+                self.mm_registry.get_mm_limits_per_prompt(
+                    self.model_config).values())
+            max_num_mm_items_decoder_budget = self.max_num_reqs * \
+                max_mm_items_per_req
+
+            max_num_mm_items = min(max_num_mm_items_encoder_budget,
+                                   max_num_mm_items_decoder_budget)
 
             # Dummy data definition in V0 may contain multiple multimodal items
             # (e.g, multiple images) for a single request, therefore here we