[Model] Fused mm preprocess normalisation on the Device (#50411)

Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>
Signed-off-by: wang.yuqi <noooop@126.com>
Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
This commit is contained in:
wang.yuqi
2026-08-06 03:11:14 -07:00
committed by GitHub
co-authored by Cyrus Leung
parent 9c22668436
commit 2fa490470d
13 changed files with 359 additions and 11 deletions
+50
View File
@@ -396,6 +396,7 @@ class ModelConfig:
video_pruning_method: InitVar[str | None] = None
mm_tensor_ipc: InitVar[MMTensorIPC] = None
mm_ipc_gpu_memory_gb: InitVar[float | None] = None
mm_device_do_normalize: InitVar[bool | None] = None
mm_processor_device: InitVar[MMProcessorDevice | None] = None
def compute_hash(self) -> str:
@@ -527,6 +528,7 @@ class ModelConfig:
video_pruning_method: str | None,
mm_tensor_ipc: MMTensorIPC,
mm_ipc_gpu_memory_gb: float | None,
mm_device_do_normalize: bool | None,
mm_processor_device: MMProcessorDevice | None,
) -> None:
# Keep set served_model_name before maybe_model_redirect(self.model)
@@ -793,6 +795,9 @@ class ModelConfig:
video_pruning_method=video_pruning_method,
mm_tensor_ipc=mm_tensor_ipc,
mm_ipc_gpu_memory_gb=mm_ipc_gpu_memory_gb,
mm_device_do_normalize=self._resolve_mm_device_do_normalize(
mm_device_do_normalize
),
)
mm_config_kwargs = {
@@ -927,6 +932,51 @@ class ModelConfig:
)
return self
def _resolve_mm_device_do_normalize(
self, mm_device_do_normalize: bool | None
) -> bool:
if mm_device_do_normalize is None:
if envs.VLLM_USE_RUST_FRONTEND:
logger.debug(
"VLLM_USE_RUST_FRONTEND is set. "
"Rust frontend does not currently support mm_device_do_normalize, "
"forcing mm_device_do_normalize = False."
)
mm_device_do_normalize = False
else:
mm_device_do_normalize = (
self._model_info.supports_mm_device_do_normalize
)
logger.debug(
"mm_device_do_normalize is %s by default.",
"enabled" if mm_device_do_normalize else "disabled",
)
else:
if mm_device_do_normalize and envs.VLLM_USE_RUST_FRONTEND:
logger.warning(
"VLLM_USE_RUST_FRONTEND is set. "
"Rust frontend does not currently support mm_device_do_normalize, "
"forcing mm_device_do_normalize = False."
)
mm_device_do_normalize = False
if (
mm_device_do_normalize
and not self._model_info.supports_mm_device_do_normalize
):
logger.warning(
"Model does not support mm_device_do_normalize, "
"forcing mm_device_do_normalize = False."
)
mm_device_do_normalize = False
logger.debug(
"mm_device_do_normalize is %s.",
"enabled" if mm_device_do_normalize else "disabled",
)
return mm_device_do_normalize
def _get_transformers_backend_cls(self) -> str:
"""Determine which Transformers modeling backend class will be used if
`model_impl` is set to `transformers` or `auto`."""