[Core][Frontend] Support Passing Multimodal Processor Kwargs (#8657)

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

vllm/utils.py

@@ -4,6 +4,7 @@ import contextlib
 import datetime
 import enum
 import gc
+import inspect
 import os
 import random
 import socket
@@ -1237,6 +1238,53 @@ async def _run_task_with_lock(task: Callable, lock: asyncio.Lock, *args,
         return await task(*args, **kwargs)
 
 
+def get_allowed_kwarg_only_overrides(
+    callable: Callable[..., object],
+    overrides: Optional[Dict[str, Any]],
+) -> Dict[str, Any]:
+    """
+    Given a callable which has one or more keyword only params and a dict
+    mapping param names to values, drop values that can be not be kwarg
+    expanded to overwrite one or more keyword-only args. This is used in a
+    few places to handle custom processor overrides for multimodal models,
+    e.g., for profiling when processor options provided by the user
+    may affect the number of mm tokens per instance.
+
+    Args:
+        callable: Callable which takes 0 or more keyword only arguments.
+        overrides: Potential overrides to be used when invoking the callable.
+
+    Returns:
+        Dictionary containing the kwargs to be leveraged which may be used
+        to overwrite one or more keyword only arguments when invoking the
+        callable.
+    """
+    if not overrides:
+        return {}
+
+    allowed_override_names = [
+        name for name, param in inspect.signature(callable).parameters.items()
+        if param.kind == inspect.Parameter.KEYWORD_ONLY
+    ]
+
+    # Drop any mm_processor_kwargs provided by the user that are
+    # not kwarg names accepted by the provided input processor.
+    filtered_overrides = {
+        kwarg_name: val
+        for kwarg_name, val in overrides.items()
+        if kwarg_name in allowed_override_names
+    }
+
+    # If anything is dropped, log a warning
+    dropped_keys = overrides.keys() - filtered_overrides.keys()
+    if dropped_keys:
+        logger.warning(
+            "The following intended overrides are not keyword-only args "
+            "and and will be dropped: %s", dropped_keys)
+
+    return filtered_overrides
+
+
 # Using dynamo with vLLM doesn't really work well with PyTorch versions < 2.4.0.
 # In particular, the FakeScalarType is not supported for earlier versions of
 # PyTorch which breaks dynamo for any ops registered using ScalarType.