Enforce valid max_num_batched_tokens when disable_chunked_mm_input=True (#16447)

Signed-off-by: mgoin <mgoin64@gmail.com>

vllm/engine/arg_utils.py

@@ -1030,7 +1030,7 @@ class EngineArgs:
             action=StoreBoolean,
             default=EngineArgs.disable_chunked_mm_input,
             nargs="?",
-            const="False",
+            const="True",
             help="Disable multimodal input chunking attention for V1. "
             "If set to true and chunked prefill is enabled, we do not want to"
             " partially schedule a multimodal item. This ensures that if a "

vllm/v1/core/encoder_cache_manager.py

@@ -133,6 +133,14 @@ def _compute_encoder_budget_multimodal(
     _, max_tokens_per_mm_item = max(max_tokens_by_modality_dict.items(),
                                     key=lambda item: item[1])
 
+    if (scheduler_config.disable_chunked_mm_input and max_tokens_per_mm_item
+            > scheduler_config.max_num_batched_tokens):
+        raise ValueError(
+            "Chunked MM input disabled but max_tokens_per_mm_item "
+            f"({max_tokens_per_mm_item}) is larger than max_num_batched_tokens"
+            f" ({scheduler_config.max_num_batched_tokens}). Please increase "
+            "max_num_batched_tokens.")
+
     encoder_compute_budget = max(scheduler_config.max_num_encoder_input_tokens,
                                  max_tokens_per_mm_item)
     encoder_cache_size = max(scheduler_config.encoder_cache_size,