Phase 3: real FSDP, jit+NamedSharding, measurement, mid-epoch resume - #15

Closed

AshishKumar4 wants to merge 5 commits into

revamp/phase-4-flow-matchingfrom

revamp/phase-3-parallelism

Commits on Jul 27, 2026

perf(trainer): jit + NamedSharding replaces shard_map, with real FSDP

AshishKumar4
and
claude
committed
e249bb8
feat(models): rematerialize transformer blocks behind --remat

AshishKumar4
and
claude
committed
85dd63d
test(trainer): cover throughput accounting, divergence and the profiler

AshishKumar4
and
claude
committed
a644bf5
fix(trainer): restore the data iterator position, and drop code the migration orphaned

AshishKumar4
and
claude
committed
aef3728
test(trainer): cover gradient accumulation across the sharding heuristic

AshishKumar4
and
claude
committed
1587b09