[MM][CG] Support ViT full CUDA graph for DeepSeek-VL2 (#46005)

Signed-off-by: littlecircle0730 <littlecircle0730@gmail.com>
Signed-off-by: littlecircle0730 <43994952+littlecircle0730@users.noreply.github.com>
Co-authored-by: Shanshan Shen <467638484@qq.com>
This commit is contained in:
littlecircle0730
2026-10-15 21:38:52 +08:00
committed by GitHub
co-authored by Shanshan Shen
parent 0951fc183c
commit 71536bd5f8
5 changed files with 482 additions and 4 deletions
@@ -70,6 +70,10 @@ def gemma3_chat_template(content: str) -> str:
return f"<bos><start_of_turn>user\n{content}<end_of_turn>\n<start_of_turn>model\n"
def deepseek_vl2_chat_template(content: str) -> str:
return f"<|User|>: {content}\n\n<|Assistant|>:"
def ernie45_vl_chat_template(content: str) -> str:
return (
f"<|begin_of_sentence|>User: {content}"
@@ -356,6 +360,15 @@ MODEL_CONFIGS: dict[str, VitCudagraphTestConfig] = {
},
skip=True, # TODO: Re-enable this once OOM issues are resolved on CI.
),
"deepseek_vl2": VitCudagraphTestConfig(
model="deepseek-ai/deepseek-vl2-tiny",
modalities=["image"],
image_prompt=deepseek_vl2_chat_template("<image>\nWhat is in this image?"),
vllm_runner_kwargs={
"hf_overrides": {"architectures": ["DeepseekVLV2ForCausalLM"]},
},
marks=[pytest.mark.core_model],
),
"gemma4": VitCudagraphTestConfig(
model="google/gemma-4-E2B-it",
image_prompt=(