feat(pi07,fsdp): enable FSDP-FULL_SHARD for full unfreeze + profile_step audit + ZeRO-2 vs FSDP matrix#273

Merged

shuheng-liu merged 21 commits into

mainfrom

claude/naughty-franklin-bb8864

May 6, 2026

Commits on May 5, 2026

feat(pi07): add Gemma 3 example config + default expandable_segments
shuheng-liu
committed
feat(train): allow gradient_checkpointing under FSDP, ship FSDP config
shuheng-liu
committed
fix(pi07): add n_obs_history/history_interval to dataset_mixture in example
shuheng-liu
committed
feat(train): wrap make_policy in deepspeed.zero.Init(enabled=False)
shuheng-liu
committed
fix(train): unset HfDeepSpeedConfig (not deepspeed.zero.Init) for ZeRO-3 init
shuheng-liu
committed
refactor(pi07): bundle interleaved decoder layer pair into one nn.Module
shuheng-liu
committed
fix(profile_step): also skip MasterWeightOptimizer wrap under FSDP
shuheng-liu
committed
fix(pi07): set deterministic drop_probs in pi07 example for FSDP compat
shuheng-liu
committed
fix(pi07): lazy attention dispatch lookup + tests for refactor
shuheng-liu
committed
fix(datasets): make missing 'count' tolerable in empty-camera padding (#264)
shuheng-liu
committed
feat(normalize): materialize DTensor stats for FSDP2 compatibility
shuheng-liu
committed
chore(fsdp2): add experimental FSDP2 yaml (partial — see comment)
shuheng-liu
committed
fix(pi07): globally OR-reduce embed_prefix branch decisions for FSDP
shuheng-liu
committed
Merge main into FSDP branch
shuheng-liu
committed
style: ruff-format collapse two-line f-strings in _build_interleaved_layers
shuheng-liu
committed
docs: honestly describe FSDP precision regime as pure bf16
shuheng-liu
committed
fix(pi07,fsdp): keep fp32 master + fp32 Adam under FSDP via gated bf16 cast
shuheng-liu
committed
fix(profile_step): rank-aggregated throughput, peak HBM, prune share table
shuheng-liu
committed
chore(configs): pi07_low_level_libero n_obs_history 8->6 for benchmark matrix
shuheng-liu
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat(pi07,fsdp): enable FSDP-FULL_SHARD for full unfreeze + profile_step audit + ZeRO-2 vs FSDP matrix#273

feat(pi07,fsdp): enable FSDP-FULL_SHARD for full unfreeze + profile_step audit + ZeRO-2 vs FSDP matrix#273
shuheng-liu merged 21 commits into
mainfrom
claude/naughty-franklin-bb8864

Commits on May 5, 2026

feat(pi07): add Gemma 3 example config + default expandable_segments

feat(train): allow gradient_checkpointing under FSDP, ship FSDP config

fix(pi07): add n_obs_history/history_interval to dataset_mixture in example

feat(train): wrap make_policy in deepspeed.zero.Init(enabled=False)

fix(train): unset HfDeepSpeedConfig (not deepspeed.zero.Init) for ZeRO-3 init

refactor(pi07): bundle interleaved decoder layer pair into one nn.Module

fix(profile_step): also skip MasterWeightOptimizer wrap under FSDP

fix(pi07): set deterministic drop_probs in pi07 example for FSDP compat

fix(pi07): lazy attention dispatch lookup + tests for refactor

fix(datasets): make missing 'count' tolerable in empty-camera padding (#264)

feat(normalize): materialize DTensor stats for FSDP2 compatibility

chore(fsdp2): add experimental FSDP2 yaml (partial — see comment)

fix(pi07): globally OR-reduce embed_prefix branch decisions for FSDP

Merge main into FSDP branch

style: ruff-format collapse two-line f-strings in _build_interleaved_layers

docs: honestly describe FSDP precision regime as pure bf16

fix(pi07,fsdp): keep fp32 master + fp32 Adam under FSDP via gated bf16 cast

fix(profile_step): rank-aggregated throughput, peak HBM, prune share table

chore(configs): pi07_low_level_libero n_obs_history 8->6 for benchmark matrix

Commits on May 6, 2026

[claude-fix] address review feedback on #273

fix(profile_step): correct stale FSDP MixedPrecision dtype in comment