Support the lm_head of QWEN2.5_OMNI？

[LiMa-cas]

as the mentioned above，thanks a lot？

[Qubitium]

Is lm head of qwen 2.5 omni that big? We see little advantage in lm head compression unless its really so large as to impact vram or performance.

[LiMa-cas]

15% speed-up for thinking-LLM part

[Qubitium]

@LiMa-cas Did you check quality of output before and after lm-head quantization? For your 15% speed up, what quant bits are you using?

[LiMa-cas]

loss 0.5% for qwen2.5-llm， 4bit