Qwen3.5-397B RL trainer on 128 v7x: pad-first sequence packing, fused sharded log-prob head VJP, and the tuned trainer recipe - #2913

Qwen3.5-397B RL trainer on 128 v7x: pad-first sequence packing, fused sharded log-prob head VJP, and the tuned trainer recipe#2913

parambole wants to merge 4 commits into

atwigg/mlperfgoogle/tunix:atwigg/mlperffrom

parambole/rl-trainer-4600google/tunix:parambole/rl-trainer-4600Copy head branch name to clipboard