VLLM_ENCODER_CACHE_SIZE控制encoder_cache_size大小

vllm/envs.py

@@ -157,6 +157,7 @@ if TYPE_CHECKING:
     VLLM_MXFP4_USE_MARLIN: bool | None = None
     VLLM_DEEPEPLL_NVFP4_DISPATCH: bool = False
     VLLM_V1_USE_OUTLINES_CACHE: bool = False
+    VLLM_ENCODER_CACHE_SIZE: int | None = None
     VLLM_TPU_BUCKET_PADDING_GAP: int = 0
     VLLM_TPU_MOST_MODEL_LEN: int | None = None
     VLLM_TPU_USING_PATHWAYS: bool = False
@@ -1925,6 +1926,8 @@ environment_variables: dict[str, Callable[[], Any]] = {
     "VLLM_USE_MOE_W16A16_TRITON":
         lambda: (os.environ.get("VLLM_USE_MOE_W16A16_TRITON", "0").lower() in
                  ("true", "1")),
+    "VLLM_ENCODER_CACHE_SIZE":
+        lambda: maybe_convert_int(os.environ.get("VLLM_ENCODER_CACHE_SIZE", None)),
     #If set to 1/True, enable the V1 fast token-id copy path in InputBatch.
     "VLLM_V1_FAST_TOKEN_ID_COPY":
         lambda: (os.environ.get("VLLM_V1_FAST_TOKEN_ID_COPY", "False").lower() in

vllm/v1/core/encoder_cache_manager.py

@@ -5,6 +5,7 @@ from collections import OrderedDict
 from collections.abc import Mapping
 from typing import TYPE_CHECKING
 
+import vllm.envs as envs
 from vllm.logger import init_logger
 from vllm.multimodal import MultiModalRegistry
 from vllm.v1.request import Request
@@ -15,6 +16,16 @@ if TYPE_CHECKING:
 logger = init_logger(__name__)
 
 
+def _get_encoder_cache_size_override() -> int | None:
+    encoder_cache_size_override = envs.VLLM_ENCODER_CACHE_SIZE
+    if encoder_cache_size_override is not None:
+        logger.info_once(
+            "Using VLLM_ENCODER_CACHE_SIZE=%d to override encoder cache size.",
+            encoder_cache_size_override,
+        )
+    return encoder_cache_size_override
+
+
 class EncoderCacheManager:
     """Manages caching of encoder outputs for multimodal models in vLLM V1.
 
@@ -342,6 +353,10 @@ def compute_mm_encoder_budget(
     encoder_compute_budget = max(
         scheduler_config.max_num_encoder_input_tokens, max_tokens_per_mm_item
     )
+    encoder_cache_size_override = _get_encoder_cache_size_override()
+    if encoder_cache_size_override is not None:
+        encoder_cache_size = encoder_cache_size_override
+    else:
         encoder_cache_size = max(
             scheduler_config.encoder_cache_size, max_tokens_per_mm_item
         )