Alex Morehead

Add learning rate scheduling support for `DeepSpeedStrategy`

11

## What does this PR do? - Adds learning rate scheduling support for `DeepSpeedStrategy` - Credit to [lvhoaa](https://github.com/lvhoaa) for suggesting this change to make Fabric's support for internal DeepSpeed features...

Registered buffers not moved to correct device when using DeepSpeed Stage 3

2

### Bug description Using the DeepSpeed `Strategy` configuration ```yaml _target_: lightning.pytorch.strategies.DeepSpeedStrategy zero_optimization: true stage: 3 allgather_bucket_size: 2e8 reduce_bucket_size: 2e8 offload_optimizer: false offload_parameters: false partition_activations: false cpu_checkpointing: false contiguous_gradients: false overlap_comm:...

* Fixes an error triggered by calling `\nointerlineskip` in horizontal mode. * This error only appeared for me when locally upgrading my Emacs version to 29.4.4 (which I use to...

* Fixes the expected returncode in `is_reduce_installed()`

Support `grad_clip_norm_()` for FSDP

7

## What does this PR do? Adds gradient norm clipping support for FSDP. Tests fine locally. For fun, here's a research deep dive ChatGPT came up with when comparing norm...

`DetermineBonds` causes a hard freeze for certain `Chem.Mol` inputs

12

**Describe the bug** When I pass certain `Chem.Mol` inputs to `DetermineBonds()` as `rdDetermineBonds.DetermineBonds(mol, allowChargedFragments=False)`, my Python script experiences a hard freeze, meaning one my CPU processes throttles to 100% while...

Add optional `start_method` argument to `_cli.py`

7

* Adds an optional `start_method` argument to `_cli.py`, to allow one to use `fork` over `spawn` for instance

Step counter reset upon resuming training job

3

### Python version 3.10.0 ### fairchem-core version 2.4.0 ### pytorch version 2.6.0 ### cuda version 12.4 ### Operating system version Ubuntu 22.04 LTS ### Minimal example ```Python fairchem -c training.yaml...

Summary: Makes `test_utils.py` (and `torchtnt` in general) safe to use `start_method=fork` for multi-GPU training with `torchelastic`. An example of a project that would benefit from this change is `fairchem`, which...