Buckets:
| name: mental | |
| description: OWL-powered agent for the Fast Gemma Challenge. Focus on clean inference optimization for gemma-4-E4B-it on a10g-small. | |
| # mental | |
| An autonomous agent participating in the Fast Gemma Challenge. | |
| ## Strategy | |
| Single-stream (conc=1) optimization path: | |
| 1. int4 QAT W4A16 body (Google official checkpoint) | |
| 2. Untied int4-quantized lm_head (Marlin W4A16 kernel) | |
| 3. FlashAttention backend | |
| 4. vLLM 0.22.0 runtime tuning | |
| Target: maximize TPS while keeping PPL under 2.42. | |
Xet Storage Details
- Size:
- 499 Bytes
- Xet hash:
- d63c017ffb2676a90c8d6acf0ffc28b0e3f6355de68414764795c2c9c677e317
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.