[MM][CG] Support ViT full CUDA graph for Ernie-4.5-VL image inference (#45254)

Signed-off-by: Qiuyang Yue <yueqiuyang1389@gmail.com>
This commit is contained in:
Qiuyang Yue
2026-08-06 05:04:19 +00:00
committed by GitHub
parent 821717118f
commit 777b01d1a8
4 changed files with 288 additions and 12 deletions
@@ -66,6 +66,14 @@ 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 ernie45_vl_chat_template(content: str) -> str:
return (
f"<|begin_of_sentence|>User: {content}"
"Picture 1:<|IMAGE_START|><|image@placeholder|><|IMAGE_END|>\n"
"Assistant: <think></think>"
)
MODEL_CONFIGS: dict[str, VitCudagraphTestConfig] = {
"gemma3": VitCudagraphTestConfig(
model="google/gemma-3-4b-it",
@@ -208,6 +216,25 @@ MODEL_CONFIGS: dict[str, VitCudagraphTestConfig] = {
),
},
),
"ernie45_vl": VitCudagraphTestConfig(
model="baidu/ERNIE-4.5-VL-28B-A3B-PT",
# Image only: Ernie's resampler applies a temporal conv for video
# that changes the output token count, so video uses the eager path.
modalities=["image"],
image_prompt=ernie45_vl_chat_template("What is in this image?"),
# Shrink to 1 text + 1 vision layer with random weights so the test
# runs on any CI GPU and skips the ~56 GiB weight download. The test
# only validates encoder CG capture/replay, not output quality.
vllm_runner_kwargs={
"load_format": "dummy",
"trust_remote_code": True,
"revision": "refs/pr/17",
"hf_overrides": partial(
dummy_hf_overrides,
model_arch="Ernie4_5_VLMoeForConditionalGeneration",
),
},
),
"glm4_1v": VitCudagraphTestConfig(
model="zai-org/GLM-4.1V-9B-Thinking",
image_prompt=(