AWQ calibration applied to the Qwen2-VL language layers. The visual encoder and Jev decision-token lm_head remain in their original dtype. The original Jev request schema and serving code are preserved. Validate with vLLM before use.
Chat template
Files info
Base model
