Skip to content

docs: fix stale dataloader-state callback references in custom_datase… #40

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
foreverlms merged 2 commits into from
Jun 15, 2026
+8 −8
Merged

docs: fix stale dataloader-state callback references in custom_datase… #40

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
74f0c2c
docs: fix stale dataloader-state callback references in custom_datase…
foreverlms Jun 12, 2026
8271b5f
Merge branch 'main' into maoshengl/doc-fix-dataloader-state-callback
lfengad Jun 15, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
16 changes: 8 additions & 8 deletions docs/custom_dataset.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -249,11 +249,11 @@ Override from the CLI like any Hydra node, e.g.

## 5. Checkpoint / resume

Resume is handled by the existing `DataLoaderStateCallback`:
Resume is handled by `CosmosDataLoaderStateCallback`:

```python
from cosmos_framework.callbacks.dataloader_state import DataLoaderStateCallback
cb = DataLoaderStateCallback(distributor_type="cosmos_dataloader")
from cosmos_framework.callbacks.cosmos_dataloader_state import CosmosDataLoaderStateCallback
cb = CosmosDataLoaderStateCallback()
```

- Use a **`MapDistributor`** source. On save, the callback records each worker's
Expand All @@ -266,8 +266,8 @@ cb = DataLoaderStateCallback(distributor_type="cosmos_dataloader")
- For multiple loaders sharing a process (e.g. inside `JointCosmosDataLoader`),
give each a distinct `name=` so resume env vars are namespaced
(`COSMOS_DL_STATE_{name}_WORKER_{id}_{EPOCH,INDEX}`), and use a single
`JointDataLoaderStateCallback(outer_loader=joint_loader, distributor_type="cosmos_dataloader")`
instead of one `DataLoaderStateCallback` per inner loader.
`JointCosmosDataLoaderStateCallback(outer_loader=joint_loader)`
instead of one `CosmosDataLoaderStateCallback` per inner loader.
- Use `ckpt_type=dcp` (the default) — not `ckpt_type=dummy`, which disables all
checkpointing. The on-disk checkpoint format is unchanged.

Expand Down Expand Up @@ -423,7 +423,7 @@ collator: VFMListCollator # media kept as p
- [ ] Pick a **collator**: `DefaultBatchCollator`, `VFMListCollator`, or your own
(must match the structure the model consumes).
- [ ] For real resume: use a `MapDistributor`, add
`DataLoaderStateCallback(distributor_type="cosmos_dataloader")`, and
`CosmosDataLoaderStateCallback()`, and
`ckpt_type=dcp` (not `dummy`).
- [ ] For FSDP+TP/PP, pass `parallel_dims=` so the correct DP rank is used.
- [ ] Register the experiment in the Hydra ConfigStore
Expand All @@ -437,8 +437,8 @@ collator: VFMListCollator # media kept as p
`name=` matching its key in `dataloaders` (namespaces resume env vars).
- [ ] Set each dataset's `ratio` (controls how often it is visited, per batch).
- [ ] Use a single
`JointDataLoaderStateCallback(outer_loader=joint_loader, distributor_type="cosmos_dataloader")`
— do **not** also register a standalone `DataLoaderStateCallback` per inner
`JointCosmosDataLoaderStateCallback(outer_loader=joint_loader)`
— do **not** also register a standalone `CosmosDataLoaderStateCallback` per inner
loader.
- [ ] Avoid `"global_id"` as a dataset name (reserved by the checkpoint state).
- [ ] Use `ckpt_type=dcp` for real checkpoint/resume.
Expand Down