[Multi Modal] Configurable MM Profiling (#25631)

Signed-off-by: wwl2755 <wangwenlong2755@gmail.com>
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>

tests/models/multimodal/processing/test_common.py

@@ -12,6 +12,8 @@ from mistral_common.protocol.instruct.request import ChatCompletionRequest
 from PIL import Image
 
 from vllm.config import ModelConfig
+from vllm.config.multimodal import (AudioDummyOptions, BaseDummyOptions,
+                                    ImageDummyOptions, VideoDummyOptions)
 from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalDataDict
 from vllm.multimodal.cache import MultiModalProcessorOnlyCache
 from vllm.multimodal.inputs import MultiModalInputs
@@ -112,12 +114,26 @@ def _test_processing_correctness(
 
     processing_info = factories.info(ctx)
     supported_mm_limits = processing_info.get_supported_mm_limits()
-    limit_mm_per_prompt = {
+    # Keep integer limits for local data generation
+    limit_mm_per_prompt_ints = {
         modality: 3 if limit is None else limit
         for modality, limit in supported_mm_limits.items()
     }
 
-    model_config.get_multimodal_config().limit_per_prompt = limit_mm_per_prompt
+    def _to_dummy_options(modality: str, count: int) -> BaseDummyOptions:
+        if modality == "video":
+            return VideoDummyOptions(count=count)
+        if modality == "image":
+            return ImageDummyOptions(count=count)
+        if modality == "audio":
+            return AudioDummyOptions(count=count)
+        return BaseDummyOptions(count=count)
+
+    # Assign normalized DummyOptions to the model config
+    model_config.get_multimodal_config().limit_per_prompt = {
+        modality: _to_dummy_options(modality, count)
+        for modality, count in limit_mm_per_prompt_ints.items()
+    }
 
     baseline_processor = factories.build_processor(ctx, cache=None)
     cached_processor = factories.build_processor(ctx, cache=cache)
@@ -150,7 +166,7 @@ def _test_processing_correctness(
             k:
             [(input_to_hit[k] if rng.rand() < hit_rate else input_factory[k]())
              for _ in range(rng.randint(limit + 1))]
-            for k, limit in limit_mm_per_prompt.items()
+            for k, limit in limit_mm_per_prompt_ints.items()
         }
 
         mm_counts = {k: len(vs) for k, vs in mm_data.items()}

tests/models/multimodal/processing/test_mllama4.py

@@ -17,23 +17,23 @@ def test_profiling(model_id: str, max_model_len: int):
     model_config_kwargs = {
         "max_model_len": max_model_len,
     }
+    mm_counts = {"image": 1}
     ctx = build_model_context(
         model_id,
         model_config_kwargs=model_config_kwargs,
-        limit_mm_per_prompt={"image": 1},
+        limit_mm_per_prompt=mm_counts,
     )
 
-    mm_config = ctx.get_mm_config()
     processor = MULTIMODAL_REGISTRY.create_processor(ctx.model_config)
     profiler = MultiModalProfiler(processor)
 
     decoder_dummy_data = profiler.get_decoder_dummy_data(
         max_model_len,
-        mm_counts=mm_config.limit_per_prompt,
+        mm_counts=mm_counts,
     )
     dummy_mm_data = processor.dummy_inputs.get_dummy_processor_inputs(
         max_model_len,
-        mm_counts=mm_config.limit_per_prompt,
+        mm_counts=mm_counts,
     )
 
     hf_config = ctx.get_hf_config(Llama4Config)
@@ -58,7 +58,7 @@ def test_profiling(model_id: str, max_model_len: int):
 
     profiled_tokens = profiler.get_mm_max_contiguous_tokens(
         max_model_len,
-        mm_counts=mm_config.limit_per_prompt,
+        mm_counts=mm_counts,
     )
 
     assert total_tokens == profiled_tokens["image"]

tests/models/multimodal/processing/test_tensor_schema.py

@@ -15,6 +15,8 @@ from mistral_common.protocol.instruct.request import ChatCompletionRequest
 from PIL import Image
 
 from vllm.config import ModelConfig, VllmConfig, set_current_vllm_config
+from vllm.config.multimodal import (AudioDummyOptions, BaseDummyOptions,
+                                    ImageDummyOptions, VideoDummyOptions)
 from vllm.distributed import (cleanup_dist_env_and_memory,
                               init_distributed_environment,
                               initialize_model_parallel)
@@ -236,7 +238,20 @@ def test_model_tensor_schema(model_arch: str, model_id: str):
         modality: 3 if limit is None else limit
         for modality, limit in supported_mm_limits.items()
     }
-    model_config.get_multimodal_config().limit_per_prompt = limit_mm_per_prompt
+
+    def _to_dummy_options(modality: str, count: int) -> BaseDummyOptions:
+        if modality == "video":
+            return VideoDummyOptions(count=count)
+        if modality == "image":
+            return ImageDummyOptions(count=count)
+        if modality == "audio":
+            return AudioDummyOptions(count=count)
+        return BaseDummyOptions(count=count)
+
+    model_config.get_multimodal_config().limit_per_prompt = {
+        modality: _to_dummy_options(modality, count)
+        for modality, count in limit_mm_per_prompt.items()
+    }
     processor = factories.build_processor(ctx, cache=None)
 
     with initialize_dummy_model(model_cls, model_config) as model: