[Minimax-M3] Add Encoder CUDA graph support (#58673)

Signed-off-by: Thien Tran <gau.nernst@yahoo.com.sg>
Co-authored-by: Codex <codex@openai.com>
This commit is contained in:
Thien Tran
2026-09-28 04:17:36 +00:00
committed by GitHub
co-authored by Codex
parent 6f65fde419
commit 55de40a2fc
5 changed files with 395 additions and 15 deletions
@@ -175,6 +175,24 @@ MODEL_CONFIGS: dict[str, VitCudagraphTestConfig] = {
},
marks=[pytest.mark.core_model],
),
"minimax_m3": VitCudagraphTestConfig(
model="MiniMaxAI/MiniMax-M3",
modalities=["image"],
image_prompt=("]~b]user\n]<]image[>[What is in this image?[e~[\n]~b]ai\n"),
compilation_config_overrides={
"encoder_cudagraph_token_budgets": [512],
},
vllm_runner_kwargs={
"trust_remote_code": True,
"load_format": "dummy",
"hf_overrides": partial(
dummy_hf_overrides,
model_arch="MiniMaxM3SparseForConditionalGeneration",
),
"mm_processor_kwargs": {"max_long_side_pixel": 336},
},
marks=[pytest.mark.core_model],
),
"qwen3_vl": VitCudagraphTestConfig(
model="Qwen/Qwen3-VL-2B-Instruct",
image_prompt=qwen_vl_chat_template(