Skip to content

Consistent fp32 grads flow #8056

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
tohtana merged 8 commits into from
Jun 10, 2026
Merged

Consistent fp32 grads flow #8056

Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
9acf239
Consistent fp32 grads flow
sfc-gh-truwase Jun 9, 2026
b1e61c5
Merge branch 'master' into sfc-gh-truwase/fp32_grads
sfc-gh-truwase Jun 9, 2026
07a8441
Coverage
sfc-gh-truwase Jun 9, 2026
52b59b6
Merge branch 'sfc-gh-truwase/fp32_grads' of https://github.com/deepsp…
sfc-gh-truwase Jun 9, 2026
ae20d9d
Format
sfc-gh-truwase Jun 9, 2026
54cf4f2
Merge branch 'master' into sfc-gh-truwase/fp32_grads
sfc-gh-truwase Jun 9, 2026
3e7f58f
Merge branch 'master' into sfc-gh-truwase/fp32_grads
tohtana Jun 10, 2026
2be795f
Merge branch 'master' into sfc-gh-truwase/fp32_grads
tohtana Jun 10, 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: 10 additions & 6 deletions deepspeed/runtime/zero/stage3.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1540,7 +1540,9 @@ def _apply_distributed_muon_update(self, communication_data_type: torch.dtype, b
start_offset = rank * chunk_sz
end_offset = start_offset + chunk_sz
if end_offset > param.grad.numel():
buffer_to_update = torch.zeros(chunk_sz, device=param.grad.device, dtype=param.grad.dtype)
buffer_to_update = torch.zeros(chunk_sz,
device=param.grad.device,
dtype=self.gradient_accumulation_dtype)
buffer_to_update[:param.grad.numel() -
start_offset] = gathered_momentum.view(-1).data[start_offset:param.grad.numel()]
else:
Expand Down Expand Up @@ -1584,8 +1586,8 @@ def __avg_scatter_contiguous_grads(self, buffer_to_reduce: Tensor,
if self.postscale_gradients and self.gradient_predivide_factor != world_sz:
buffer_to_reduce = buffer_to_reduce.mul(self.gradient_predivide_factor)

if communication_data_type != self.dtype:
buffer_to_reduce = buffer_to_reduce.to(self.dtype)
if communication_data_type != self.gradient_accumulation_dtype:
buffer_to_reduce = buffer_to_reduce.to(self.gradient_accumulation_dtype)

@chatgpt-codex-connector chatgpt-codex-connector Bot Jun 9, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Badge Preserve fp32 dtype when padding contiguous partitions

When ZeRO-3 runs with contiguous_gradients=True, reduce_scatter=False, fp32 gradient accumulation, and a low-precision communication dtype, this conversion makes buffer_to_reduce fp32, but the padding branch below still allocates padded_partition with dtype=grad.dtype. For any parameter whose size is not evenly divisible by the data-parallel world size, the real tail partition is assigned into that low-precision padded tensor before partition_grads() copies it to the fp32 accumulation buffer, so those gradients are silently rounded back to fp16/bf16. The padded tensor should use the converted partition/buffer dtype, not the original parameter grad dtype.

Useful? React with 👍 / 👎.


grad_partitions = []
grad_offset_in_buffer = 0
Expand All @@ -1599,7 +1601,9 @@ def __avg_scatter_contiguous_grads(self, buffer_to_reduce: Tensor,

partition = buffer_to_reduce[start_offset:end_offset]
if param.partition_numel() != partition.numel():
padded_partition = torch.zeros(param.partition_numel(), device=grad.device, dtype=grad.dtype)
padded_partition = torch.zeros(param.partition_numel(),
device=grad.device,
dtype=self.gradient_accumulation_dtype)
if partition.numel() > 0:
padded_partition[:partition.numel()] = partition
grad_partitions.append(padded_partition)
Expand Down Expand Up @@ -1636,8 +1640,8 @@ def __avg_scatter_grads(self, params_to_reduce: List[Parameter],
self.dp_process_group):
grad_partitions_for_rank = [g.mul(self.gradient_predivide_factor) for g in grad_partitions_for_rank]

if communication_data_type != self.dtype:
grad_partitions_for_rank = [g.to(self.dtype) for g in grad_partitions_for_rank]
if communication_data_type != self.gradient_accumulation_dtype:
grad_partitions_for_rank = [g.to(self.gradient_accumulation_dtype) for g in grad_partitions_for_rank]

return grad_partitions_for_rank

Expand Down
Loading