[V1][Core] Autotune encoder cache budget (#11895)

Signed-off-by: Roger Wang <ywang@roblox.com>

vllm/config.py

@@ -1387,13 +1387,15 @@ 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
+    # NOTE: The following multimodal encoder budget will be initialized to
+    # max_num_batched_tokens and overridden in case max multimodal embedding
+    # size is larger.
+    # TODO (ywang96): Make these configurable.
+    # Multimodal encoder compute budget, only used in V1
+    max_num_encoder_input_tokens: int = field(default=None)  # type: ignore
 
     # Multimodal encoder cache size, only used in V1
-    encoder_cache_size = 16384
+    encoder_cache_size: int = field(default=None)  # type: ignore
 
     # Whether to perform preemption by swapping or
     # recomputation. If not specified, we determine the mode as follows:
@@ -1467,6 +1469,9 @@ class SchedulerConfig:
                     _MULTIMODAL_MODEL_MAX_NUM_BATCHED_TOKENS,
                 )
 
+        self.max_num_encoder_input_tokens = self.max_num_batched_tokens
+        self.encoder_cache_size = self.max_num_batched_tokens
+
         if self.enable_chunked_prefill:
             logger.info(
                 "Chunked prefill is enabled with max_num_batched_tokens=%d.",

vllm/multimodal/registry.py

@@ -252,11 +252,8 @@ class MultiModalRegistry:
         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.
+        Get the maximum number of tokens per data item from each modality based 
+        on underlying model configuration.
         """
         if self.has_processor(model_config):
             tokenizer = cached_get_tokenizer(
@@ -272,6 +269,28 @@ class MultiModalRegistry:
             for key, plugin in self._plugins.items()
         }
 
+    def get_max_tokens_per_item_by_nonzero_modality(
+        self,
+        model_config: "ModelConfig",
+    ) -> Mapping[str, int]:
+        """
+        Get the maximum number of tokens per data item from each modality based
+        on underlying model configuration, excluding modalities that user 
+        explicitly disabled via `limit_mm_per_prompt`.
+
+        Note:
+            This is currently directly used only in V1 for profiling the memory 
+            usage of a model.
+        """
+        limits_per_plugin = self._limits_by_model[model_config]
+
+        return {
+            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()
+            if limits_per_plugin[key] > 0
+        }
+
     def get_max_tokens_by_modality(
         self,
         model_config: "ModelConfig",

vllm/v1/core/encoder_cache_manager.py

-from typing import Dict, List, Set, Tuple
+from typing import TYPE_CHECKING, Dict, List, Set, Tuple
 
+from vllm.logger import init_logger
+from vllm.multimodal import MULTIMODAL_REGISTRY
 from vllm.v1.request import Request
 
+if TYPE_CHECKING:
+    from vllm.config import ModelConfig, SchedulerConfig
+
+logger = init_logger(__name__)
+
 
 class EncoderCacheManager:
 
@@ -46,3 +53,72 @@ class EncoderCacheManager:
         freed = self.freed
         self.freed = []
         return freed
+
+
+def compute_encoder_budget(
+    model_config: "ModelConfig",
+    scheduler_config: "SchedulerConfig",
+) -> Tuple[int, int]:
+    """Compute the encoder cache budget based on the model and scheduler 
+    configurations.
+
+    Args:
+        model_config: Model configuration.
+        scheduler_config: Scheduler configuration.
+
+    Returns:
+        - Compute budget for encoder execution, in unit of number of tokens 
+            in the input sequence.
+        - Space budget for encoder cache size, in unit of number of tokens 
+            in the input sequence.
+    """
+
+    if not model_config.is_multimodal_model:
+        return 0, 0
+
+    # TODO: handle encoder-decoder models once we support them.
+    (
+        encoder_compute_budget,
+        encoder_cache_size,
+    ) = _compute_encoder_budget_multimodal(model_config, scheduler_config)
+
+    return encoder_compute_budget, encoder_cache_size
+
+
+def _compute_encoder_budget_multimodal(
+    model_config: "ModelConfig",
+    scheduler_config: "SchedulerConfig",
+) -> Tuple[int, int]:
+    """Compute the encoder cache budget based on the model and scheduler 
+    configurations for a multimodal model.
+
+    Args:
+        model_config: Model configuration.
+        scheduler_config: Scheduler configuration.
+
+    Returns:
+        - Compute budget for encoder execution, in unit of number of tokens 
+            in the input sequence.
+        - Space budget for encoder cache size, in unit of number of tokens 
+            in the input sequence.
+    """
+
+    max_tokens_by_modality_dict = MULTIMODAL_REGISTRY.get_max_tokens_per_item_by_nonzero_modality(  # noqa: E501
+        model_config)
+
+    if not max_tokens_by_modality_dict:
+        logger.warning(
+            "All non-text modalities supported by the model have been "
+            "explicitly disabled via limit_mm_per_prompt. Encoder cache will "
+            "not be initialized.")
+        return 0, 0
+
+    _, max_tokens_per_mm_item = max(max_tokens_by_modality_dict.items(),
+                                    key=lambda item: item[1])
+
+    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,
+                             max_tokens_per_mm_item)
+
+    return encoder_compute_budget, encoder_cache_size

vllm/v1/core/scheduler.py

@@ -3,10 +3,11 @@ from dataclasses import dataclass
 from typing import (TYPE_CHECKING, Deque, Dict, Iterable, List, Optional, Set,
                     Tuple, Union)
 
-from vllm.config import CacheConfig, LoRAConfig, SchedulerConfig
+from vllm.config import CacheConfig, LoRAConfig, ModelConfig, SchedulerConfig
 from vllm.logger import init_logger
 from vllm.sampling_params import SamplingParams
-from vllm.v1.core.encoder_cache_manager import EncoderCacheManager
+from vllm.v1.core.encoder_cache_manager import (EncoderCacheManager,
+                                                compute_encoder_budget)
 from vllm.v1.core.kv_cache_manager import KVCacheManager
 from vllm.v1.engine import EngineCoreOutput, EngineCoreOutputs
 from vllm.v1.metrics.stats import SchedulerStats
@@ -25,6 +26,7 @@ class Scheduler:
     def __init__(
         self,
         scheduler_config: SchedulerConfig,
+        model_config: ModelConfig,
         cache_config: CacheConfig,
         lora_config: Optional[LoRAConfig],
     ) -> None:
@@ -69,16 +71,24 @@ class Scheduler:
         self.running_reqs_data: Dict[str, RunningRequestData] = {}
 
         # Encoder-related.
+        # Calculate encoder cache size if applicable
+        # NOTE: For now we use the same budget for both compute and space.
+        # This can be changed when we make encoder cache for embedding caching
+        # across requests.
+        encoder_compute_budget, encoder_cache_size = compute_encoder_budget(
+            model_config=model_config,
+            scheduler_config=scheduler_config,
+        )
+
         # 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).
-        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.max_num_encoder_input_tokens = encoder_compute_budget
+        # NOTE: For the models without encoder (e.g., text-only models),
+        # the encoder cache will not be initialized because cache size is 0
+        # for these models.
         self.encoder_cache_manager = EncoderCacheManager(
-            cache_size=self.scheduler_config.encoder_cache_size)
+            cache_size=encoder_cache_size)
 
     def schedule(self) -> "SchedulerOutput":
         # NOTE(woosuk) on the scheduling algorithm:

vllm/v1/engine/core.py

@@ -54,9 +54,12 @@ class EngineCore:
         vllm_config.cache_config.num_cpu_blocks = num_cpu_blocks
 
         # Setup scheduler.
-        self.scheduler = Scheduler(vllm_config.scheduler_config,
-                                   vllm_config.cache_config,
-                                   vllm_config.lora_config)
+        self.scheduler = Scheduler(
+            scheduler_config=vllm_config.scheduler_config,
+            model_config=vllm_config.model_config,
+            cache_config=vllm_config.cache_config,
+            lora_config=vllm_config.lora_config,
+        )
 
         self.mm_input_mapper_server = MMInputMapperServer(
             vllm_config.model_config)

vllm/v1/worker/gpu_model_runner.py

@@ -20,6 +20,7 @@ from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, DeviceMemoryProfiler,
                         is_pin_memory_available)
 from vllm.v1.attention.backends.flash_attn import (FlashAttentionBackend,
                                                    FlashAttentionMetadata)
+from vllm.v1.core.encoder_cache_manager import compute_encoder_budget
 from vllm.v1.engine.mm_input_mapper import MMInputMapperClient
 from vllm.v1.outputs import ModelRunnerOutput
 from vllm.v1.sample.metadata import SamplingMetadata
@@ -88,8 +89,12 @@ class GPUModelRunner:
         self.mm_input_mapper_profiling = MMInputMapperClient(self.model_config)
         self.mm_input_mapper_profiling.use_cache = False
 
-        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
+        encoder_compute_budget, encoder_cache_size = compute_encoder_budget(
+            model_config=model_config,
+            scheduler_config=scheduler_config,
+        )
+        self.max_num_encoder_input_tokens = encoder_compute_budget
+        self.encoder_cache_size = encoder_cache_size
 
         # Lazy initialization
         # self.model: nn.Module  # Set after load_model
@@ -721,44 +726,30 @@ class GPUModelRunner:
         ]
 
         # Profile with multimodal encoder & encoder cache.
-        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
+        # TODO: handle encoder-decoder models once we support them.
+        if (self.is_multimodal_model and self.max_num_encoder_input_tokens > 0
+                and self.encoder_cache_size > 0):
 
             # NOTE: Currently model is profiled with a single non-text
             # modality with the max possible input tokens even when
             # it supports multiple.
-            max_tokens_by_modality_dict = self.mm_registry.get_max_tokens_per_item_by_modality(  # noqa: E501
+            max_tokens_by_modality_dict = MULTIMODAL_REGISTRY.get_max_tokens_per_item_by_nonzero_modality(  # noqa: E501
                 self.model_config)
-
             dummy_data_modality, max_tokens_per_mm_item = max(
                 max_tokens_by_modality_dict.items(), key=lambda item: item[1])
 
             # Check how many items of this modality can be supported by
-            # the encoder cache budget.
-            encoder_cache_budget = min(self.max_num_encoder_input_tokens,
+            # the encoder budget.
+            encoder_budget = min(self.max_num_encoder_input_tokens,
                                  self.encoder_cache_size)
-            max_num_mm_items_encoder_budget = encoder_cache_budget // \
-                max_tokens_per_mm_item
 
-            # TODO: Allow users to set encoder_cache_budget in case this
-            # happens.
-            assert max_num_mm_items_encoder_budget > 0, (
-                f"Encoder cache budget={encoder_cache_budget} is too small to "
-                f"support the maximum possible size of multimodal embeddings"
-                f"={max_tokens_per_mm_item}.")
+            max_num_mm_items_encoder_budget = cdiv(encoder_budget,
+                                                   max_tokens_per_mm_item)
 
             # Check how many items of this modality can be supported by
             # the decoder budget.
-            max_mm_items_per_req = max(
-                self.mm_registry.get_mm_limits_per_prompt(
-                    self.model_config).values())
+            max_mm_items_per_req = self.mm_registry.get_mm_limits_per_prompt(
+                self.model_config)[dummy_data_modality]
 
             # NOTE: We do not consider max_num_batched_tokens on purpose
             # because the multimodal embeddings can be generated in advance
@@ -769,6 +760,19 @@ class GPUModelRunner:
             max_num_mm_items = min(max_num_mm_items_encoder_budget,
                                    max_num_mm_items_decoder_budget)
 
+            logger.info(
+                "Encoder cache will be initialized with a budget of %s tokens,"
+                " and profiled with %s %s items of the maximum feature size.",
+                encoder_budget, max_num_mm_items, dummy_data_modality)
+
+            # 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 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