w
gcw_uQ09W7jl/videoflextok_d18_d18_k600-NPU
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

EPFL-VILAB/videoflextok_d18_d18_k600 on Ascend NPU

模型与任务

  • 模型:EPFL-VILAB/videoflextok_d18_d18_k600
  • Revision:b1e468168ba7a1869206d3fc74e369c095895722
  • 适配路线:video-tokenizer
  • 设备:npu:0

环境依赖

{
  "torch_version": "2.9.0+cpu",
  "torch_npu_available": true,
  "device_name": "Ascend910_9362",
  "device_count": 2
}

安装:pip install -r requirements.txt

NPU 推理

默认执行:python inference.py

真实输入

{
  "video_shape": [
    1,
    3,
    17,
    128,
    128
  ],
  "resolution": "128x128",
  "frames": 17,
  "normalized": "-1~1",
  "seed": 42
}

真实推理结果

{
  "tokens_shape": [
    1,
    5,
    128
  ],
  "reconst_shape": [
    1,
    3,
    17,
    128,
    128
  ],
  "tokens_device": "npu:0",
  "reconst_device": "npu:0"
}

CPU-NPU 一致性

{
  "token_ids_match": true,
  "max_abs_diff": 0.0,
  "mean_abs_diff": 0.0,
  "reconst_psnr": 2.8424744033592826
}

性能

{
  "npu_inference_ms": 175.72569847106934,
  "frames_per_s": 96.74168404457247,
  "tokens_per_s": 3642.039869913316,
  "dtype": "float32",
  "resolution": "128x128",
  "frames": 17
}

自验证截图

Agent workflow

NPU device call

Model result

已知限制

  • 仅对本次记录的模型 revision、输入契约、dtype 与 Ascend 环境完成验证。

标签

#NPU #Ascend #Ascend910