Hugging Face's logo

AbstractFramework
/
wan2.2-t2v-a14b-diffusers-8bit

Text-to-Video
Safetensors
MLX
mflux
Wan2.2
mlx-gen
apple-silicon
8-bit precision
mixed-q8-bf16
wan
video-generation
wan-a14b
Model card Files
xet
 Community
1

I downloaded this model and ran it with mlxgen 0.35.0, but I got an error.

#1
by Howek - opened
Discussion
Howek

I downloaded this model and tried running it with mlxgen 0.35.0, but it threw an error saying the tokenizer was missing. However, the tokenizer files are actually inside the model folder. Why is this happening?
lpalbou
 AbstractFramework org

Thanks for the report. The tokenizer files in this repo load correctly with mlx-gen 0.35.0 and both transformers 5.0 and 5.20, so this looks like the resolver picking a different folder than the one you are looking at. Could you share:

  1. The exact command you ran (the model card's flow is mlxgen download --model AbstractFramework/wan2.2-t2v-a14b-diffusers-8bit followed by mlxgen generate --model AbstractFramework/wan2.2-t2v-a14b-diffusers-8bit --task text-to-video ...).
  2. The full error output. The message names the folder it checked.
  3. ls -la <that folder>/tokenizer.
  4. Whether a models/wan2.2-t2v-a14b-diffusers-8bit folder exists in your working directory or home directory. mlx-gen prefers such a folder over the Hugging Face cache.
  5. pip show mlx-gen transformers.

Re-running the mlxgen download command is safe and completes a partial cache if that is what happened.

Sign up or log in to comment