[MM][Perf][CG] Support ViT full cudagraphs for mllama4 (#40660)

Signed-off-by: allgather <all2allops@gmail.com>
Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
allgather
2026-06-11 22:17:55 -07:00
committed by GitHub
co-authored by Isotr0py
parent eb28452b10
commit 39dee1114a
5 changed files with 193 additions and 14 deletions
@@ -55,6 +55,26 @@ def step3_vl_chat_template(content: str) -> str:
MODEL_CONFIGS: dict[str, VitCudagraphTestConfig] = {
"llama4": VitCudagraphTestConfig(
model="meta-llama/Llama-4-Scout-17B-16E-Instruct",
modalities=["image"],
image_prompt=(
"<|begin_of_text|><|header_start|>user<|header_end|>\n\n"
"<|image|>What is in this image?<|eot|>"
"<|header_start|>assistant<|header_end|>\n\n"
),
max_model_len=4096,
max_tokens=32,
max_num_seqs=2,
vllm_runner_kwargs={
"load_format": "dummy",
"hf_overrides": partial(
dummy_hf_overrides,
model_arch="Llama4ForConditionalGeneration",
),
},
marks=[pytest.mark.core_model],
),
"internvl": VitCudagraphTestConfig(
model="OpenGVLab/InternVL3-1B",
num_video_frames=8,
+3 -2
View File
@@ -507,12 +507,13 @@ def dummy_hf_overrides(
# Only set MoE related config when the model has MoE layers.
# Otherwise all models detected as MoE by _get_transformers_backend_cls.
if model_arch_config.num_experts > 0:
num_experts_per_tok = 1 if model_arch == "Llama4ForConditionalGeneration" else 2
update_dict.update(
{
"num_experts": num_experts,
"num_experts_per_tok": 2,
"num_experts_per_tok": num_experts_per_tok,
# Kimi uses `num_experts_per_token`.
"num_experts_per_token": 2,
"num_experts_per_token": num_experts_per_tok,
"num_local_experts": num_experts,
# Otherwise there will not be any expert layers
"first_k_dense_replace": 0,