From 2a194ddd72a0cc5b6c404a694a64197d0c572f5b Mon Sep 17 00:00:00 2001 From: Woosuk Kwon Date: Mon, 9 Mar 2026 15:14:51 -0700 Subject: [PATCH] [Model Runner V2] Add model_state inputs to CUDA graph capture (#36544) Signed-off-by: Woosuk Kwon --- vllm/v1/worker/gpu/cudagraph_utils.py | 1 + 1 file changed, 1 insertion(+) diff --git a/vllm/v1/worker/gpu/cudagraph_utils.py b/vllm/v1/worker/gpu/cudagraph_utils.py index 2b3cee110..2ec3cb2a2 100644 --- a/vllm/v1/worker/gpu/cudagraph_utils.py +++ b/vllm/v1/worker/gpu/cudagraph_utils.py @@ -320,6 +320,7 @@ class ModelCudaGraphManager(CudaGraphManager): model_inputs = { "input_ids": input_buffers.input_ids[:num_tokens], "positions": input_buffers.positions[:num_tokens], + **model_state.prepare_dummy_inputs(num_reqs, num_tokens), } model_output = model(**model_inputs) if self.use_aux_hidden_state_outputs: