[DO NOT MERGE] Integration: MaxText side of the 128-chip best run (#5718) by parambole · Pull Request #5723 · AI-Hypercomputer/maxtext

[MaxTextEngine] Prepare a micro-batch's inputs on the mesh ahead of its fwd_bwd

parambole

committed

ebc0191

[MaxTextEngine] Report how long a caller waited for the device

parambole

committed

ef37a0b

[MaxTextEngine] Test prepared micro-batches under dropless replay and the fused update

parambole

committed

b72884c

[MaxTextEngine] Test the order Tunix uses to stage before a replayed update

parambole

committed

0b107b4

[MaxTextEngine] Drive the staged-before-update order through Tunix's TrainerWorker

parambole

committed

c6b304a

[MaxTextEngine] Document that staging leaves a pending overflow check alone

parambole

committed

c79bcfd

Integration: merge #5718 parambole/engine-stage-before-update (c79bcfd16)

parambole

committed

5f1e9ad

Integration: merge #5701 sivaibhav/rl-trainer-bench (70d64441e)

parambole

committed

0968476

Integration only: STAGE_BEFORE_UPDATE switch in the RL trainer bench driver (calls the #5718 engine hooks)

parambole

committed

2320c99

Revert "Integration only: STAGE_BEFORE_UPDATE switch in the RL trainer bench driver (calls the #5718 engine hooks)"

parambole

committed

da29b18

Revert "Integration: merge #5701 sivaibhav/rl-trainer-bench (70d64441e)"

parambole

committed

d9848b5

Navigation Menu

Repository navigation

[DO NOT MERGE] Integration: MaxText side of the 128-chip best run (#5718) - #5723

[DO NOT MERGE] Integration: MaxText side of the 128-chip best run (#5718)#5723
parambole wants to merge 14 commits into
atwigg/mlperffrom
parambole/integration-128-best

Commits on Oct 13, 2026

Add an RL trainer bench under experimental/rl_trainer_bench

rl_trainer_bench: label example numbers, fail fast in aot_e2e.sh, note DRY_RUN context switch

rl_trainer_bench: rename e2e identifiers to bench/launch names

Commits on Oct 14, 2026

[MaxTextEngine] Prepare a micro-batch's inputs on the mesh ahead of its fwd_bwd

[MaxTextEngine] Report how long a caller waited for the device

[MaxTextEngine] Test prepared micro-batches under dropless replay and the fused update

[MaxTextEngine] Test the order Tunix uses to stage before a replayed update

[MaxTextEngine] Drive the staged-before-update order through Tunix's TrainerWorker

[MaxTextEngine] Document that staging leaves a pending overflow check alone

Integration: merge #5718 parambole/engine-stage-before-update (c79bcfd16)

Integration: merge #5701 sivaibhav/rl-trainer-bench (70d64441e)

Integration only: STAGE_BEFORE_UPDATE switch in the RL trainer bench driver (calls the #5718 engine hooks)

Revert "Integration only: STAGE_BEFORE_UPDATE switch in the RL trainer bench driver (calls the #5718 engine hooks)"

Revert "Integration: merge #5701 sivaibhav/rl-trainer-bench (70d64441e)"