Skip to content

[Bug] VBench SceneMetric loads Qwen2.5-Omni on CPU on any non-CUDA accelerator (device_map hardcodes cuda) #1815

Description

@li-lizhe

Problem

in only passes a when :

On any non-CUDA accelerator (Ascend NPU, Apple MPS, Intel XPU, AMD ROCm) the Qwen2.5-Omni model silently loads onto the CPU, so the scene metric runs on CPU (slow) even when a GPU/NPU is available. The model weights are loaded via .

Expected

Follow the actual device instead of assuming CUDA: set for every non-CPU device.

Environment

  • repo: hao-ai-lab/FastVideo
  • device: Ascend 910B (torch.accelerator.current_accelerator() == npu)

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    installationInstallation and setup issuesplatformPlatform-specific (Windows/macOS)scope: inferenceInference pipeline, serving, CLI

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions