Gemma 4 in Pure JAX: What Changes Between Turing and Ada, and What Doesn't

Chronological Source Flow
Back

AI Fusion Summary

This measurement report analyzes a hand-written Gemma 4 port in pure JAX across two NVIDIA GPUs: the Turing-based T4G and the Ada-based L4. Using a single build and the google/gemma-4-E2B-it checkpoint, the study identifies two abstraction leaks in JAX. One specific leak results in an 87% loss in decode performance, despite no errors appearing in the logs. The tests were conducted on g5g.2xlarge and g6.2xlarge spot instances using Graviton2 and x86_64 hosts respectively.
Community Comments
Loading updates...
0