[Core] Shared memory based object store for Multimodal data caching and IPC (#20452)

Signed-off-by: donglu <donglu@cohere.com>

vllm/config/__init__.py

@@ -262,6 +262,7 @@ def is_init_field(cls: ConfigType, name: str) -> bool:
 TokenizerMode = Literal["auto", "slow", "mistral", "custom"]
 ModelDType = Literal["auto", "half", "float16", "bfloat16", "float", "float32"]
 MMEncoderTPMode = Literal["weights", "data"]
+MMCacheType = Literal["shm", "lru"]
 
 
 class LogprobsMode(enum.Enum):
@@ -450,6 +451,13 @@ class ModelConfig:
     `mm_processor_cache_gb * (api_server_count + data_parallel_size)`.
 
     Set to `0` to disable this cache completely (not recommended)."""
+    mm_processor_cache_type: MMCacheType = "lru"
+    """Type of cache to use for the multi-modal preprocessor/mapper. If `shm`,
+    use shared memory FIFO cache. If `lru`, use mirrored LRU cache."""
+    mm_shm_cache_max_object_size_mb: int = 128
+    """Size limit (in MiB) for each object stored in the multi-modal processor
+    shared memory cache. Only effective when `mm_processor_cache_type` is
+    `"shm"`."""
     mm_encoder_tp_mode: MMEncoderTPMode = "weights"
     """Indicates how to optimize multi-modal encoder inference using
     tensor parallelism (TP).
@@ -881,6 +889,9 @@ class ModelConfig:
                 media_io_kwargs=self.media_io_kwargs,
                 mm_processor_kwargs=self.mm_processor_kwargs,
                 mm_processor_cache_gb=self.mm_processor_cache_gb,
+                mm_processor_cache_type=self.mm_processor_cache_type,
+                mm_shm_cache_max_object_size_mb=self.
+                mm_shm_cache_max_object_size_mb,
                 mm_encoder_tp_mode=self.mm_encoder_tp_mode,
                 interleave_mm_strings=self.interleave_mm_strings,
                 skip_mm_profiling=self.skip_mm_profiling,
@@ -2448,6 +2459,15 @@ class MultiModalConfig:
     Set to `0` to disable this cache completely (not recommended).
     """
 
+    mm_processor_cache_type: MMCacheType = "lru"
+    """Type of cache to use for the multi-modal preprocessor/mapper. If `shm`,
+    use shared memory FIFO cache. If `lru`, use mirrored LRU cache."""
+
+    mm_shm_cache_max_object_size_mb: int = 128
+    """Size limit (in MiB) for each object stored in the multi-modal processor
+    shared memory cache. Only effective when `mm_processor_cache_type` is
+    `"shm"`."""
+
     mm_encoder_tp_mode: MMEncoderTPMode = "weights"
     """
     Indicates how to optimize multi-modal encoder inference using