megatron-fsdp 0.2.0.dev100825__tar.gz → 0.2.0.dev101062__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (17) hide show
  1. {megatron_fsdp-0.2.0.dev100825 → megatron_fsdp-0.2.0.dev101062}/PKG-INFO +1 -1
  2. {megatron_fsdp-0.2.0.dev100825 → megatron_fsdp-0.2.0.dev101062}/megatron_fsdp/package_info.py +1 -1
  3. {megatron_fsdp-0.2.0.dev100825 → megatron_fsdp-0.2.0.dev101062}/megatron_fsdp/param_and_grad_buffer.py +0 -3
  4. {megatron_fsdp-0.2.0.dev100825 → megatron_fsdp-0.2.0.dev101062}/megatron_fsdp.egg-info/PKG-INFO +1 -1
  5. {megatron_fsdp-0.2.0.dev100825 → megatron_fsdp-0.2.0.dev101062}/README.md +0 -0
  6. {megatron_fsdp-0.2.0.dev100825 → megatron_fsdp-0.2.0.dev101062}/megatron_fsdp/__init__.py +0 -0
  7. {megatron_fsdp-0.2.0.dev100825 → megatron_fsdp-0.2.0.dev101062}/megatron_fsdp/distributed_data_parallel_config.py +0 -0
  8. {megatron_fsdp-0.2.0.dev100825 → megatron_fsdp-0.2.0.dev101062}/megatron_fsdp/fully_shard.py +0 -0
  9. {megatron_fsdp-0.2.0.dev100825 → megatron_fsdp-0.2.0.dev101062}/megatron_fsdp/megatron_fsdp.py +0 -0
  10. {megatron_fsdp-0.2.0.dev100825 → megatron_fsdp-0.2.0.dev101062}/megatron_fsdp/uneven_dtensor.py +0 -0
  11. {megatron_fsdp-0.2.0.dev100825 → megatron_fsdp-0.2.0.dev101062}/megatron_fsdp/utils.py +0 -0
  12. {megatron_fsdp-0.2.0.dev100825 → megatron_fsdp-0.2.0.dev101062}/megatron_fsdp.egg-info/SOURCES.txt +0 -0
  13. {megatron_fsdp-0.2.0.dev100825 → megatron_fsdp-0.2.0.dev101062}/megatron_fsdp.egg-info/dependency_links.txt +0 -0
  14. {megatron_fsdp-0.2.0.dev100825 → megatron_fsdp-0.2.0.dev101062}/megatron_fsdp.egg-info/requires.txt +0 -0
  15. {megatron_fsdp-0.2.0.dev100825 → megatron_fsdp-0.2.0.dev101062}/megatron_fsdp.egg-info/top_level.txt +0 -0
  16. {megatron_fsdp-0.2.0.dev100825 → megatron_fsdp-0.2.0.dev101062}/pyproject.toml +0 -0
  17. {megatron_fsdp-0.2.0.dev100825 → megatron_fsdp-0.2.0.dev101062}/setup.cfg +0 -0
{megatron_fsdp-0.2.0.dev100825 → megatron_fsdp-0.2.0.dev101062}/PKG-INFO RENAMED
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: megatron-fsdp
3
- Version: 0.2.0.dev100825
3
+ Version: 0.2.0.dev101062
4
4
  Summary: **Megatron-FSDP** is an NVIDIA-developed PyTorch extension that provides a high-performance implementation of Fully Sharded Data Parallelism (FSDP)
5
5
  Author-email: NVIDIA <nemo-toolkit@nvidia.com>
6
6
  Maintainer-email: NVIDIA <nemo-toolkit@nvidia.com>
{megatron_fsdp-0.2.0.dev100825 → megatron_fsdp-0.2.0.dev101062}/megatron_fsdp/package_info.py RENAMED
@@ -4,7 +4,7 @@
4
4
  MAJOR = 0
5
5
  MINOR = 2
6
6
  PATCH = 0
7
- PRE_RELEASE = '0.dev100825'
7
+ PRE_RELEASE = '0.dev101062'
8
8
 
9
9
  # Use the following formatting: (major, minor, patch, pre-release)
10
10
  VERSION = (MAJOR, MINOR, PATCH, PRE_RELEASE)
{megatron_fsdp-0.2.0.dev100825 → megatron_fsdp-0.2.0.dev101062}/megatron_fsdp/param_and_grad_buffer.py RENAMED
@@ -2782,9 +2782,6 @@ class GradReducePipeline:
2782
2782
  outer_fsdp_group_grad_reduce (bool, optional): Whether to reduce gradients
2783
2783
  across outer-DP groups. Defaults to False.
2784
2784
  """
2785
- # Sort parameters by their bucket IDs to ensure a deterministic processing order.
2786
- # Performing reduce-scatter operations out of order can lead to hangs.
2787
- params = sorted(list(params), key=lambda x: self.buffer.param_to_param_group[x])
2788
2785
  for param in params:
2789
2786
  bucket_id = self.buffer.param_to_param_group[param]
2790
2787
  param_group = self.buffer.parameter_groups[bucket_id]
{megatron_fsdp-0.2.0.dev100825 → megatron_fsdp-0.2.0.dev101062}/megatron_fsdp.egg-info/PKG-INFO RENAMED
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: megatron-fsdp
3
- Version: 0.2.0.dev100825
3
+ Version: 0.2.0.dev101062
4
4
  Summary: **Megatron-FSDP** is an NVIDIA-developed PyTorch extension that provides a high-performance implementation of Fully Sharded Data Parallelism (FSDP)
5
5
  Author-email: NVIDIA <nemo-toolkit@nvidia.com>
6
6
  Maintainer-email: NVIDIA <nemo-toolkit@nvidia.com>