[Annieren]

Overview

Add Q8_0 data type direct fp32 activation support.

Additional information

1. Performance improved for Q8_0 models, or the Q8_0 as output models.

model	test	original-baseline	this branch	Δ (B/A)
Cosmos-Reason2-8B.Q8_0	tg128	26.76	26.93	+0.6%
Qwen2.5-0.5B-Instruct-Q4_K_M	tg128	351.17	353.60	+0.7%
SmolVLM-500M-Instruct-Q8_0	tg128	310.65	348.00	+12%
gemma-4-26B-A4B-it-UD-Q4_K_M	tg128	48.45	49.21	+1.6%

3. No change on prefill (within noise level).
4. PPL agreed within +/- 0.1% comparing to previous version

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure:

Yes, test and documentation

[jimw567]

looks good