[VLM] Refactor `MultiModalConfig` initialization and profiling (#7530)

vllm/worker/cpu_worker.py

@@ -7,8 +7,8 @@ import torch.distributed
 import vllm.envs as envs
 from vllm.attention import get_attn_backend
 from vllm.config import (CacheConfig, DeviceConfig, LoadConfig, LoRAConfig,
-                         ModelConfig, MultiModalConfig, ParallelConfig,
-                         PromptAdapterConfig, SchedulerConfig)
+                         ModelConfig, ParallelConfig, PromptAdapterConfig,
+                         SchedulerConfig)
 from vllm.distributed import (ensure_model_parallel_initialized,
                               init_distributed_environment)
 from vllm.logger import init_logger
@@ -132,7 +132,6 @@ class CPUWorker(LoraNotSupportedWorkerBase, LocalOrDistributedWorkerBase):
         rank: int,
         distributed_init_method: str,
         lora_config: Optional[LoRAConfig] = None,
-        multimodal_config: Optional[MultiModalConfig] = None,
         kv_cache_dtype: Optional[str] = "auto",
         prompt_adapter_config: Optional[PromptAdapterConfig] = None,
         is_driver_worker: bool = False,
@@ -148,7 +147,6 @@ class CPUWorker(LoraNotSupportedWorkerBase, LocalOrDistributedWorkerBase):
         self.distributed_init_method = distributed_init_method
         self.lora_config = lora_config
         self.prompt_adapter_config = prompt_adapter_config
-        self.multimodal_config = multimodal_config
         self.is_driver_worker = is_driver_worker
         if self.is_driver_worker:
             assert self.rank == 0, "The driver worker must have rank 0."
@@ -173,7 +171,6 @@ class CPUWorker(LoraNotSupportedWorkerBase, LocalOrDistributedWorkerBase):
             cache_config,
             load_config=self.load_config,
             lora_config=self.lora_config,
-            multimodal_config=self.multimodal_config,
             kv_cache_dtype=kv_cache_dtype,
             prompt_adapter_config=self.prompt_adapter_config,
             is_driver_worker=is_driver_worker)

vllm/worker/embedding_model_runner.py

@@ -4,8 +4,8 @@ from typing import Any, Dict, List, Optional, Tuple, Type
 import torch
 
 from vllm.config import (CacheConfig, DeviceConfig, LoadConfig, LoRAConfig,
-                         ModelConfig, MultiModalConfig, ObservabilityConfig,
-                         ParallelConfig, PromptAdapterConfig, SchedulerConfig)
+                         ModelConfig, ObservabilityConfig, ParallelConfig,
+                         PromptAdapterConfig, SchedulerConfig)
 from vllm.logger import init_logger
 from vllm.model_executor.pooling_metadata import PoolingMetadata
 from vllm.multimodal import MultiModalInputs
@@ -44,7 +44,6 @@ class EmbeddingModelRunner(
         kv_cache_dtype: Optional[str] = "auto",
         is_driver_worker: bool = False,
         prompt_adapter_config: Optional[PromptAdapterConfig] = None,
-        multimodal_config: Optional[MultiModalConfig] = None,
         observability_config: Optional[ObservabilityConfig] = None,
     ):
         super().__init__(model_config,
@@ -57,7 +56,6 @@ class EmbeddingModelRunner(
                          kv_cache_dtype=kv_cache_dtype,
                          is_driver_worker=is_driver_worker,
                          prompt_adapter_config=prompt_adapter_config,
-                         multimodal_config=multimodal_config,
                          observability_config=observability_config)
 
     @torch.inference_mode()

vllm/worker/enc_dec_model_runner.py

@@ -10,8 +10,8 @@ from vllm.attention.selector import (_Backend, get_env_variable_attn_backend,
                                      get_global_forced_attn_backend,
                                      global_force_attn_backend)
 from vllm.config import (CacheConfig, DeviceConfig, LoadConfig, LoRAConfig,
-                         ModelConfig, MultiModalConfig, ObservabilityConfig,
-                         ParallelConfig, PromptAdapterConfig, SchedulerConfig)
+                         ModelConfig, ObservabilityConfig, ParallelConfig,
+                         PromptAdapterConfig, SchedulerConfig)
 from vllm.inputs import INPUT_REGISTRY, InputRegistry
 from vllm.logger import init_logger
 from vllm.model_executor import SamplingMetadata
@@ -82,7 +82,6 @@ class EncoderDecoderModelRunner(GPUModelRunnerBase[EncoderDecoderModelInput]):
         kv_cache_dtype: Optional[str] = "auto",
         is_driver_worker: bool = False,
         prompt_adapter_config: Optional[PromptAdapterConfig] = None,
-        multimodal_config: Optional[MultiModalConfig] = None,
         observability_config: Optional[ObservabilityConfig] = None,
         input_registry: InputRegistry = INPUT_REGISTRY,
         mm_registry: MultiModalRegistry = MULTIMODAL_REGISTRY,
@@ -90,7 +89,7 @@ class EncoderDecoderModelRunner(GPUModelRunnerBase[EncoderDecoderModelInput]):
         '''
         EncoderDecoderModelRunner constructor.
 
-        `lora_config`, `multimodal_config`, and prompt_adapter_config are
+        `lora_config` and `prompt_adapter_config` are
         unused (since these features are not yet supported for encoder/decoder
         models) but these arguments are present here for compatibility with 
         the base-class constructor.
@@ -273,14 +272,8 @@ class EncoderDecoderModelRunner(GPUModelRunnerBase[EncoderDecoderModelInput]):
         # number of tokens equal to max_num_batched_tokens.
         seqs: List[SequenceGroupMetadata] = []
 
-        model_config = self.model_config
-        mm_config = self.multimodal_config
-
-        input_registry = self.input_registry
-        mm_registry = self.mm_registry
-        mm_registry.init_mm_limits_per_prompt(model_config, mm_config)
-
-        max_mm_tokens = mm_registry.get_max_multimodal_tokens(model_config)
+        max_mm_tokens = self.mm_registry.get_max_multimodal_tokens(
+            self.model_config)
         if max_mm_tokens > 0:
             raise NotImplementedError(
                 "Multi-modal encoder-decoder models are not supported yet")
@@ -291,8 +284,10 @@ class EncoderDecoderModelRunner(GPUModelRunnerBase[EncoderDecoderModelInput]):
                        (group_id < max_num_batched_tokens % max_num_seqs))
             batch_size += seq_len
 
-            seq_data, _ = input_registry \
-                .dummy_data_for_profiling(model_config, seq_len, mm_registry)
+            seq_data, _ = self.input_registry \
+                .dummy_data_for_profiling(self.model_config,
+                                          seq_len,
+                                          self.mm_registry)
 
             # Having more tokens is over-conservative but otherwise fine
             assert len(seq_data.prompt_token_ids) >= seq_len, (

vllm/worker/model_runner.py

@@ -27,8 +27,8 @@ except ImportError:
 import vllm.envs as envs
 from vllm.attention import AttentionMetadata, get_attn_backend
 from vllm.config import (CacheConfig, DeviceConfig, LoadConfig, LoRAConfig,
-                         ModelConfig, MultiModalConfig, ObservabilityConfig,
-                         ParallelConfig, PromptAdapterConfig, SchedulerConfig)
+                         ModelConfig, ObservabilityConfig, ParallelConfig,
+                         PromptAdapterConfig, SchedulerConfig)
 from vllm.distributed import get_pp_group
 from vllm.distributed.parallel_state import graph_capture
 from vllm.inputs import INPUT_REGISTRY, InputRegistry
@@ -804,7 +804,6 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
         kv_cache_dtype: Optional[str] = "auto",
         is_driver_worker: bool = False,
         prompt_adapter_config: Optional[PromptAdapterConfig] = None,
-        multimodal_config: Optional[MultiModalConfig] = None,
         return_hidden_states: bool = False,
         observability_config: Optional[ObservabilityConfig] = None,
         input_registry: InputRegistry = INPUT_REGISTRY,
@@ -819,7 +818,6 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
         self.load_config = load_config
         self.is_driver_worker = is_driver_worker
         self.prompt_adapter_config = prompt_adapter_config
-        self.multimodal_config = multimodal_config
         self.return_hidden_states = return_hidden_states
         self.observability_config = observability_config
 
@@ -866,6 +864,7 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
         self.mm_registry = mm_registry
         self.multi_modal_input_mapper = mm_registry \
             .create_input_mapper(model_config)
+        self.mm_registry.init_mm_limits_per_prompt(self.model_config)
 
         # Lazy initialization
         self.model: nn.Module  # Set after load_model
@@ -893,7 +892,6 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
                                    device_config=self.device_config,
                                    load_config=self.load_config,
                                    lora_config=self.lora_config,
-                                   multimodal_config=self.multimodal_config,
                                    parallel_config=self.parallel_config,
                                    scheduler_config=self.scheduler_config,
                                    cache_config=self.cache_config)
@@ -1056,14 +1054,9 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
         # To exercise the worst scenario for GPU memory consumption,
         # the number of seqs (batch_size) is chosen to maximize the number
         # of images processed.
-        model_config = self.model_config
-        mm_config = self.multimodal_config
 
-        input_registry = self.input_registry
-        mm_registry = self.mm_registry
-        mm_registry.init_mm_limits_per_prompt(model_config, mm_config)
-
-        max_mm_tokens = mm_registry.get_max_multimodal_tokens(model_config)
+        max_mm_tokens = self.mm_registry.get_max_multimodal_tokens(
+            self.model_config)
         if max_mm_tokens > 0:
             max_num_seqs_orig = max_num_seqs
             max_num_seqs = min(max_num_seqs,
@@ -1082,8 +1075,10 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
                        (group_id < max_num_batched_tokens % max_num_seqs))
             batch_size += seq_len
 
-            seq_data, dummy_multi_modal_data = input_registry \
-                .dummy_data_for_profiling(model_config, seq_len, mm_registry)
+            seq_data, dummy_multi_modal_data = self.input_registry \
+                .dummy_data_for_profiling(self.model_config,
+                                          seq_len,
+                                          self.mm_registry)
 
             seq = SequenceGroupMetadata(
                 request_id=str(group_id),

vllm/worker/tpu_model_runner.py

@@ -11,7 +11,7 @@ import torch_xla.runtime as xr
 
 from vllm.attention import AttentionMetadata, get_attn_backend
 from vllm.config import (CacheConfig, DeviceConfig, LoadConfig, ModelConfig,
-                         MultiModalConfig, ParallelConfig, SchedulerConfig)
+                         ParallelConfig, SchedulerConfig)
 from vllm.logger import init_logger
 from vllm.model_executor.model_loader import get_model
 from vllm.model_executor.sampling_metadata import SamplingMetadata
@@ -89,7 +89,6 @@ class TPUModelRunner(ModelRunnerBase[ModelInputForTPU]):
         device_config: DeviceConfig,
         cache_config: CacheConfig,
         load_config: LoadConfig,
-        multimodal_config: Optional[MultiModalConfig] = None,
         is_driver_worker: bool = False,
     ):
         self.model_config = model_config
@@ -98,7 +97,6 @@ class TPUModelRunner(ModelRunnerBase[ModelInputForTPU]):
         self.device_config = device_config
         self.cache_config = cache_config
         self.load_config = load_config
-        self.multimodal_config = multimodal_config
         self.is_driver_worker = is_driver_worker
 
         self.block_size = self.cache_config.block_size
@@ -142,7 +140,6 @@ class TPUModelRunner(ModelRunnerBase[ModelInputForTPU]):
                 parallel_config=self.parallel_config,
                 cache_config=self.cache_config,
                 scheduler_config=self.scheduler_config,
-                multimodal_config=self.multimodal_config,
                 lora_config=None,
             )
         model = model.eval()

vllm/worker/tpu_worker.py

@@ -7,7 +7,7 @@ import torch_xla.runtime as xr
 
 import vllm.envs as envs
 from vllm.config import (CacheConfig, DeviceConfig, LoadConfig, ModelConfig,
-                         MultiModalConfig, ParallelConfig, SchedulerConfig)
+                         ParallelConfig, SchedulerConfig)
 from vllm.distributed import (ensure_model_parallel_initialized,
                               init_distributed_environment)
 from vllm.logger import init_logger
@@ -31,7 +31,6 @@ class TPUWorker(LoraNotSupportedWorkerBase, LocalOrDistributedWorkerBase):
         device_config: DeviceConfig,
         cache_config: CacheConfig,
         load_config: LoadConfig,
-        multimodal_config: Optional[MultiModalConfig],
         local_rank: int,
         rank: int,
         distributed_init_method: str,
@@ -44,7 +43,6 @@ class TPUWorker(LoraNotSupportedWorkerBase, LocalOrDistributedWorkerBase):
         self.device_config = device_config
         self.cache_config = cache_config
         self.load_config = load_config
-        self.multimodal_config = multimodal_config
         self.local_rank = local_rank
         self.rank = rank
         self.distributed_init_method = distributed_init_method
@@ -64,7 +62,6 @@ class TPUWorker(LoraNotSupportedWorkerBase, LocalOrDistributedWorkerBase):
             device_config,
             cache_config,
             load_config,
-            multimodal_config,
             is_driver_worker=is_driver_worker)
 
     def init_device(self) -> None:

vllm/worker/utils.py

@@ -39,7 +39,7 @@ def assert_enc_dec_mr_supported_scenario(
         raise NotImplementedError(
             STR_NOT_IMPL_ENC_DEC_ERR_STRS['STR_NOT_IMPL_ENC_DEC_PP'])
 
-    if enc_dec_mr.multimodal_config is not None:
+    if enc_dec_mr.model_config.multimodal_config is not None:
         raise NotImplementedError(
             STR_NOT_IMPL_ENC_DEC_ERR_STRS['STR_NOT_IMPL_ENC_DEC_MM'])
 

vllm/worker/worker.py

@@ -7,8 +7,8 @@ import torch
 import torch.distributed
 
 from vllm.config import (CacheConfig, DeviceConfig, LoadConfig, LoRAConfig,
-                         ModelConfig, MultiModalConfig, ObservabilityConfig,
-                         ParallelConfig, PromptAdapterConfig, SchedulerConfig,
+                         ModelConfig, ObservabilityConfig, ParallelConfig,
+                         PromptAdapterConfig, SchedulerConfig,
                          SpeculativeConfig)
 from vllm.distributed import (ensure_model_parallel_initialized,
                               init_distributed_environment,
@@ -46,7 +46,6 @@ class Worker(LocalOrDistributedWorkerBase):
         rank: int,
         distributed_init_method: str,
         lora_config: Optional[LoRAConfig] = None,
-        multimodal_config: Optional[MultiModalConfig] = None,
         speculative_config: Optional[SpeculativeConfig] = None,
         prompt_adapter_config: Optional[PromptAdapterConfig] = None,
         is_driver_worker: bool = False,
@@ -73,7 +72,6 @@ class Worker(LocalOrDistributedWorkerBase):
             # note: lazy import to avoid importing torch before initializing
             from vllm.utils import init_cached_hf_modules
             init_cached_hf_modules()
-        self.multimodal_config = multimodal_config
         self.observability_config = observability_config
 
         # Return hidden states from target model if the draft model is an
@@ -103,7 +101,6 @@ class Worker(LocalOrDistributedWorkerBase):
             kv_cache_dtype=self.cache_config.cache_dtype,
             is_driver_worker=is_driver_worker,
             prompt_adapter_config=prompt_adapter_config,
-            multimodal_config=multimodal_config,
             observability_config=observability_config,
             **speculative_args,
         )

vllm/worker/xpu_model_runner.py

@@ -125,6 +125,7 @@ class XPUModelRunner(ModelRunnerBase[ModelInputForXPU]):
         self.mm_registry = mm_registry
         self.multi_modal_input_mapper = mm_registry \
             .create_input_mapper(model_config)
+        self.mm_registry.init_mm_limits_per_prompt(self.model_config)
 
         # Lazy initialization.
         self.model: nn.Module  # Set after init_Model
@@ -166,14 +167,8 @@ class XPUModelRunner(ModelRunnerBase[ModelInputForXPU]):
         # To exercise the worst scenario for GPU memory consumption,
         # the number of seqs (batch_size) is chosen to maximize the number
         # of images processed.
-        model_config = self.model_config
-        mm_config = self.multimodal_config
-
-        input_registry = self.input_registry
-        mm_registry = self.mm_registry
-        mm_registry.init_mm_limits_per_prompt(model_config, mm_config)
-
-        max_mm_tokens = mm_registry.get_max_multimodal_tokens(model_config)
+        max_mm_tokens = self.mm_registry.get_max_multimodal_tokens(
+            self.model_config)
         if max_mm_tokens > 0:
             max_num_seqs_orig = max_num_seqs
             max_num_seqs = min(max_num_seqs,
@@ -190,8 +185,10 @@ class XPUModelRunner(ModelRunnerBase[ModelInputForXPU]):
             seq_len = (max_num_batched_tokens // max_num_seqs +
                        (group_id < max_num_batched_tokens % max_num_seqs))
 
-            seq_data, dummy_multi_modal_data = input_registry \
-                .dummy_data_for_profiling(model_config, seq_len, mm_registry)
+            seq_data, dummy_multi_modal_data = self.input_registry \
+                .dummy_data_for_profiling(self.model_config,
+                                          seq_len,
+                                          self.mm_registry)
 
             seq = SequenceGroupMetadata(
                 request_id=str(group_id),