Wan-AI

/

Wan2.1-T2V-1.3B

like 458

Follow

Wan-AI 10.7k

Manually add --no-deps to the install command: pip install -r requirements.txt --no-deps and make sure you have a ROCm-specific torch install. I have it running on my RX 7900 XTX but the 25 steps is running for ~24 mins at 832*480 with a non-trivial VAE decode too. I need to get sage-attn installed again, as I did when I played with Hunyuan.

When you talk about "non-trivial VAE decode" you mean ComfyUI?
--no-deps option didn't help unfortunately:
Wan2.1# pip3.12 install -r requirements.txt --no-deps --break-system-packages
Requirement already satisfied: torch>=2.4.0 in /usr/local/lib/python3.12/dist-packages (from -r requirements.txt (line 1)) (2.7.0.dev20250119+rocm6.3)
Requirement already satisfied: torchvision>=0.19.0 in /usr/local/lib/python3.12/dist-packages (from -r requirements.txt (line 2)) (0.22.0.dev20250119+rocm6.3)
Requirement already satisfied: opencv-python>=4.9.0.80 in /usr/local/lib/python3.12/dist-packages (from -r requirements.txt (line 3)) (4.11.0.86)
Requirement already satisfied: diffusers>=0.31.0 in /usr/local/lib/python3.12/dist-packages (from -r requirements.txt (line 4)) (0.32.2)
Collecting transformers>=4.49.0 (from -r requirements.txt (line 5))
Downloading transformers-4.49.0-py3-none-any.whl.metadata (44 kB)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 44.0/44.0 kB 321.4 kB/s eta 0:00:00
Requirement already satisfied: tokenizers>=0.20.3 in /usr/local/lib/python3.12/dist-packages (from -r requirements.txt (line 6)) (0.21.0)
Requirement already satisfied: accelerate>=1.1.1 in /usr/local/lib/python3.12/dist-packages (from -r requirements.txt (line 7)) (1.3.0)
Requirement already satisfied: tqdm in /usr/local/lib/python3.12/dist-packages (from -r requirements.txt (line 8)) (4.67.1)
Collecting imageio (from -r requirements.txt (line 9))
Downloading imageio-2.37.0-py3-none-any.whl.metadata (5.2 kB)
Collecting easydict (from -r requirements.txt (line 10))
Downloading easydict-1.13-py3-none-any.whl.metadata (4.2 kB)
Collecting ftfy (from -r requirements.txt (line 11))
Downloading ftfy-6.3.1-py3-none-any.whl.metadata (7.3 kB)
Collecting dashscope (from -r requirements.txt (line 12))
Downloading dashscope-1.22.2-py3-none-any.whl.metadata (6.8 kB)
Requirement already satisfied: imageio-ffmpeg in /usr/local/lib/python3.12/dist-packages (from -r requirements.txt (line 13)) (0.6.0)
Collecting flash_attn (from -r requirements.txt (line 14))
Downloading flash_attn-2.7.4.post1.tar.gz (6.0 MB)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 6.0/6.0 MB 1.1 MB/s eta 0:00:00
Preparing metadata (setup.py) ... error
error: subprocess-exited-with-error

× python setup.py egg_info did not run successfully.
│ exit code: 1
╰─> [21 lines of output]
/tmp/pip-install-ly5kyotc/flash-attn_a10f07a52fe1407e8b9f7bc672ef64d3/setup.py:106: UserWarning: flash_attn was requested, but nvcc was not found. Are you sure your environment has nvcc available? If you're installing within a container from https://hub.docker.com/r/pytorch/pytorch, only images whose names contain 'devel' will provide nvcc.
warnings.warn(
Traceback (most recent call last):
File "", line 2, in
File "", line 34, in
File "/tmp/pip-install-ly5kyotc/flash-attn_a10f07a52fe1407e8b9f7bc672ef64d3/setup.py", line 198, in
CUDAExtension(
File "/usr/local/lib/python3.12/dist-packages/torch/utils/cpp_extension.py", line 1139, in CUDAExtension
library_dirs += library_paths(device_type="cuda")
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/dist-packages/torch/utils/cpp_extension.py", line 1274, in library_paths
if (not os.path.exists(_join_cuda_home(lib_dir)) and
^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/dist-packages/torch/utils/cpp_extension.py", line 2535, in _join_cuda_home
raise OSError('CUDA_HOME environment variable is not set. '
OSError: CUDA_HOME environment variable is not set. Please set it to your CUDA install root.

  torch.__version__  = 2.7.0.dev20250119+rocm6.3
  
  
  [end of output]

note: This error originates from a subprocess, and is likely not a problem with pip.
error: metadata-generation-failed

× Encountered error while generating package metadata.
╰─> See above for output.

note: This is an issue with the package mentioned above, not pip.
hint: See above for details.

When you talk about "non-trivial VAE decode" you mean ComfyUI?

Yep.

As for your error - my apologies, I actually installed and ran the Kijai workflows which provide a different requirements.txt which does work on my AMD GPU: https://github.com/kijai/ComfyUI-WanVideoWrapper

Also, if it helps, I am using Python 3.10 and Ubuntu 24.04, with the latest version of ComfyUI.

Incase it is useful to yourself or others: Fixed my VAE issues by not pre-empting OOMs and changing the tile size to 256x256 - a 720*480 video will decode in 16s now.

With TeaCache, SLG, and torch.compile(), I'm able to generate a 81 frame video from an input image in just under 20 minutes, with 30 steps, with the 480P 14B FP8_e5m2 model. This is using PyTorch's native Flash attention (via SDP) on PyTorch 2.6+rocm6.2.4.

I attempted to use flash attention 2 also, but it increased the generation time during the sampling by ~50%.