Jev2 vision — experimental AWQ W4A16
Text and visual encoder Linear layers use W4A16. AWQ smoothing maps the visual norm1/qkv and norm2/fc1 pairs; other visual Linear layers use W4A16 without AWQ smoothing. Conv3d, embeddings and Jev decision lm_head remain in original dtype. Validate text and image requests in vLLM before deployment.
- Downloads last month
- 35
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support