A high-throughput and memory-efficient inference and serving engine for LLMs
[MM][CG] Simplify ViT CUDA graph interfaces (#44484)
Signed-off-by: shen-shanshan <467638484@qq.com>
Shanshan Shen committed
Jun 8, 2026 at 13:57 UTC
8fb0274415062912ec56c6629a04cbac60121fe3
2 files changed
+2
-8
vllm/model_executor/models/interfaces.py
+2
-2
@@ -1524,8 +1524,8 @@ class SupportsEncoderCudaGraph(Protocol):
1524
self,
1525
mm_kwargs: dict[str, Any],
1526
) -> str:
1527
- """Return the modality of the inputs."""
1528
- ...
1527
+ """Return the modality of the inputs (default: image-only)."""
1528
+ return "image"
1529
1530
def get_max_frames_per_video(
1531
self,
vllm/model_executor/models/step3_vl.py
-6
@@ -709,12 +709,6 @@ class Step3VLForConditionalGeneration(
709
out_hidden_size=self.config.hidden_size,
710
)
711
712
- def get_input_modality(
713
- self,
714
- mm_kwargs: dict[str, Any],
715
- ) -> str:
716
- return "image"
717
-
712
def get_encoder_cudagraph_budget_range(
713
self,
714
vllm_config: "VllmConfig",