Buckets:

gemma-challenge/gemma-mental / mental_agent.md
ArifKemall's picture
|
download
raw
499 Bytes
metadata
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.