Skip to content

[bug]: fixed comm_dtype in extra_large_param_to_reduce #7660

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 2 commits into from
Oct 31, 2025
+3 −2
Merged

[bug]: fixed comm_dtype in extra_large_param_to_reduce #7660

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
33c9a9c
fixed comm_dtype in extra_large_param_to_reduce
therealnaveenkamal Oct 31, 2025
edd2db3
Merge branch 'master' into ipg_bug
tohtana Oct 31, 2025
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
5 changes: 3 additions & 2 deletions deepspeed/runtime/zero/stage_1_and_2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1002,7 +1002,8 @@ def flatten_dense_tensors_aligned(self, tensor_list, alignment, use_cpu_data=Fal
def reduce_independent_p_g_buckets_and_remove_grads(self, param, i):

grad_reduc = self.get_gradient_for_reduction(param)
bucket = self.ipg_buckets[self.get_param_comm_dtype(param)]
comm_dtype = self.get_param_comm_dtype(param)
bucket = self.ipg_buckets[comm_dtype]
if bucket.elements + param.numel() > self.reduce_bucket_size:
self.report_ipg_memory_usage("In ipg_remove_grads before reduce_ipg_grads", param.numel())
self.reduce_ipg_grads()
Expand All @@ -1022,7 +1023,7 @@ def reduce_independent_p_g_buckets_and_remove_grads(self, param, i):

if self.contiguous_gradients:
if param.numel() > self.reduce_bucket_size:
self.extra_large_param_to_reduce[param.dtype] = param
self.extra_large_param_to_reduce[comm_dtype] = param
else:
# keeping the gradients contiguous to prevent memory fragmentation, and avoid flattening
new_grad_tensor = bucket.buffer[bucket.index].narrow(0, bucket.elements, param.numel())
Expand Down