Skip to content

vllm.model_executor.warmup.qwen_vl_triton_warmup

Warm Qwen3-VL ViT kernels and M-RoPE for SupportsMRoPE models.

Functions:

qwen_vl_triton_warmup(runner)

Warm Qwen3-VL ViT kernels and M-RoPE for SupportsMRoPE models.

Source code in vllm/model_executor/warmup/qwen_vl_triton_warmup.py 
106
107
108
109
110
111
112
@torch.inference_mode()
def qwen_vl_triton_warmup(runner: "GPUModelRunner") -> None:
    """Warm Qwen3-VL ViT kernels and M-RoPE for ``SupportsMRoPE`` models."""
    model = runner.get_model()
    _warm_vision(model, runner.device)
    _warm_mrope(runner, model)
    _synchronize_device(runner.device)