[V1][VLM] Proper memory profiling for image language models (#11210)

Signed-off-by: Roger Wang <[email protected]>
Co-authored-by: ywang96 <[email protected]>

Author: ywang96

vllm/config.py

@@ -1280,6 +1280,14 @@ class SchedulerConfig:
 
     is_multimodal_model: bool = False
 
+    # FIXME(woosuk & ywang96): Below are placeholder values. We need to
+    # calculate the actual values from the configurations.
+    # Multimodal encoder run compute budget, only used in V1
+    max_num_encoder_input_tokens = 16384
+
+    # Multimodal encoder cache size, only used in V1
+    encoder_cache_size = 16384
+
     # Whether to perform preemption by swapping or
     # recomputation. If not specified, we determine the mode as follows:
     # We use recomputation by default since it incurs lower overhead than

vllm/model_executor/models/pixtral.py

@@ -245,6 +245,11 @@ def get_multimodal_embeddings(self, **kwargs) -> Optional[NestedTensors]:
             # Do not split, return as tensor of shape [1, fs, hs]
             return image_embeds.unsqueeze(0)
 
+        # If the last split index is the last index in image_tokens, we
+        # ignore it to avoid empty split tensor
+        if split_indices[-1] == len(image_tokens):
+            split_indices = split_indices[:-1]
+
         image_embeds = image_embeds.tensor_split(split_indices.cpu())
         return image_embeds
 

vllm/multimodal/registry.py

@@ -200,6 +200,23 @@ def register_max_image_tokens(
         """
         return self.register_max_multimodal_tokens("image", max_mm_tokens)
 
+    def get_max_tokens_per_item_by_modality(
+        self,
+        model_config: "ModelConfig",
+    ) -> Mapping[str, int]:
+        """
+        Get the maximum number of tokens per data item from each modality
+        for profiling the memory usage of a model.
+
+        Note:
+            This is currently directly used only in V1.
+        """
+
+        return {
+            key: plugin.get_max_multimodal_tokens(model_config)
+            for key, plugin in self._plugins.items()
+        }
+
     def get_max_tokens_by_modality(
         self,
         model_config: "ModelConfig",
@@ -216,9 +233,9 @@ def get_max_tokens_by_modality(
         limits_per_plugin = self._limits_by_model[model_config]
 
         return {
-            key: (limits_per_plugin[key] *
-                  plugin.get_max_multimodal_tokens(model_config))
-            for key, plugin in self._plugins.items()
+            key: limits_per_plugin[key] * max_tokens_per_mm_item
+            for key, max_tokens_per_mm_item in
+            self.get_max_tokens_per_item_by_modality(model_config).items()
         }
 
     def get_max_multimodal_tokens(self, model_config: "ModelConfig") -> int:

vllm/v1/core/scheduler.py

@@ -73,14 +73,13 @@ def __init__(
         # NOTE(woosuk): Here, "encoder" includes the vision encoder (and
         # projector if needed). Currently, we assume that the encoder also
         # has the Transformer architecture (e.g., ViT).
-        # FIXME(woosuk): Below are placeholder values. We need to calculate the
-        # actual values from the configurations.
-        self.max_num_encoder_input_tokens = 16384
+        self.max_num_encoder_input_tokens = self.scheduler_config.max_num_encoder_input_tokens  #noqa: E501
         # NOTE(woosuk): For the models without encoder (e.g., text-only models),
         # the encoder cache will not be initialized and used, regardless of
         # the cache size. This is because the memory space for the encoder cache
         # is preallocated in the profiling run.
-        self.encoder_cache_manager = EncoderCacheManager(cache_size=16384)
+        self.encoder_cache_manager = EncoderCacheManager(
+            cache_size=self.scheduler_config.encoder_cache_size)
 
     def schedule(self) -> "SchedulerOutput":
         # NOTE(woosuk) on the scheduling algorithm:

vllm/v1/engine/mm_input_mapper.py

@@ -54,6 +54,7 @@ def cache_hit_ratio(self, steps):
             logger.debug("MMInputMapper: cache_hit_ratio = %.2f ",
                          self.mm_cache_hits / self.mm_cache_total)
 
+    # TODO: Support modalities beyond image.
     def process_inputs(
         self,
         mm_data: MultiModalDataDict,

vllm/v1/worker/gpu_model_runner.py

@@ -10,15 +10,16 @@
 from vllm.config import CompilationLevel, VllmConfig
 from vllm.distributed.parallel_state import graph_capture
 from vllm.forward_context import set_forward_context
-from vllm.inputs import INPUT_REGISTRY, InputRegistry
+from vllm.inputs import INPUT_REGISTRY
 from vllm.logger import init_logger
 from vllm.model_executor.model_loader import get_model
-from vllm.multimodal import MultiModalKwargs
+from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalKwargs
 from vllm.sampling_params import SamplingType
 from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, DeviceMemoryProfiler,
                         LayerBlockType, cdiv, is_pin_memory_available)
 from vllm.v1.attention.backends.flash_attn import (FlashAttentionBackend,
                                                    FlashAttentionMetadata)
+from vllm.v1.engine.mm_input_mapper import MMInputMapperClient
 from vllm.v1.outputs import ModelRunnerOutput
 from vllm.v1.sample.metadata import SamplingMetadata
 from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
@@ -35,7 +36,6 @@ def __init__(
         self,
         vllm_config: VllmConfig,
         device: torch.device,
-        input_registry: InputRegistry = INPUT_REGISTRY,
     ):
         self.vllm_config = vllm_config
         self.model_config = vllm_config.model_config
@@ -77,7 +77,12 @@ def __init__(
         self.hidden_size = model_config.get_hidden_size()
 
         # Multi-modal data support
-        self.input_registry = input_registry
+        self.input_registry = INPUT_REGISTRY
+        self.mm_registry = MULTIMODAL_REGISTRY
+        # NOTE: mm_input_mapper is only used for memory profiling.
+        self.mm_input_mapper = MMInputMapperClient(self.model_config)
+        self.max_num_encoder_input_tokens = self.scheduler_config.max_num_encoder_input_tokens  # noqa: E501
+        self.encoder_cache_size = self.scheduler_config.encoder_cache_size
 
         # Lazy initialization
         # self.model: nn.Module  # Set after load_model
@@ -599,8 +604,6 @@ def _dummy_run(
         return hidden_states
 
     def profile_run(self) -> None:
-        # TODO(woosuk): Profile the max memory usage of the encoder and
-        # the encoder cache.
         # use an empty tensor instead of `None`` to force Dynamo to pass
         # it by reference, rather by specializing on the value `None`.
         # the `dtype` argument does not matter, and we use `float32` as
@@ -612,6 +615,57 @@ def profile_run(self) -> None:
             torch.tensor([], dtype=torch.float32, device=self.device)
             for _ in range(self.num_attn_layers)
         ]
+
+        # Profile with multimodal encoder & encoder cache.
+        # TODO (ywang96): generalize this beyond image modality since
+        # mm_input_mapper only supports image inputs.
+        if self.is_multimodal_model:
+
+            # Create dummy batch of multimodal inputs.
+            dummy_request_data = self.input_registry.dummy_data_for_profiling(
+                model_config=self.model_config,
+                seq_len=self.max_num_tokens,
+                mm_registry=self.mm_registry,
+            )
+            dummy_mm_data = dummy_request_data.multi_modal_data
+            dummy_mm_kwargs, _ = self.mm_input_mapper.process_inputs(
+                mm_data=dummy_mm_data,
+                mm_hashes=None,
+                mm_processor_kwargs=None,
+                precomputed_mm_inputs=None)
+
+            # NOTE: Currently model is profiled with a single non-text
+            # modality even when it supports multiple.
+            max_tokens_per_mm_item = max(
+                self.mm_registry.get_max_tokens_per_item_by_modality(
+                    self.model_config).values())
+
+            max_num_mm_items = min(
+                self.max_num_encoder_input_tokens,
+                self.encoder_cache_size) // max_tokens_per_mm_item
+
+            # Dummy data definition in V0 may contain multiple multimodal items
+            # (e.g, multiple images) for a single request, therefore here we
+            # always replicate first item by max_num_mm_items times since in V1
+            # they are scheduled to be processed separately.
+            batched_dummy_mm_inputs = MultiModalKwargs.batch(
+                [dummy_mm_kwargs[0]] * max_num_mm_items)
+            batched_dummy_mm_inputs = MultiModalKwargs.as_kwargs(
+                batched_dummy_mm_inputs, device=self.device)
+
+            # Run multimodal encoder.
+            dummy_encoder_outputs = self.model.get_multimodal_embeddings(
+                **batched_dummy_mm_inputs)
+            assert len(dummy_encoder_outputs) == max_num_mm_items, (
+                "Expected dimension 0 of encoder outputs to match the number "
+                f"of multimodal data items: {max_num_mm_items}, got "
+                f"{len(dummy_encoder_outputs)=} instead. This is most likely "
+                "due to the 'get_multimodal_embeddings' method of the model "
+                "not implemented correctly.")
+
+            # Cache the dummy encoder outputs.
+            self.encoder_cache["tmp"] = dict(enumerate(dummy_encoder_outputs))
+
         # Trigger compilation for general shape.
         hidden_states = self._dummy_run(self.model, self.max_num_tokens,
                                         dummy_kv_caches)
@@ -620,6 +674,7 @@ def profile_run(self) -> None:
         # TODO(woosuk): Consider the memory usage of the sampler.
         torch.cuda.synchronize()
         del hidden_states, logits
+        self.encoder_cache.clear()
         gc.collect()
 
     def capture_model(self) -> None: