Skip to content

Important bugfixes in local CG implementation that were leading to loss curve gaps for latent MoE models #4433

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
ericharper merged 3 commits into from
Apr 24, 2026
Merged

Important bugfixes in local CG implementation that were leading to loss curve gaps for latent MoE models #4433

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
edf80e8
fix cg convergence
jiemingz Apr 24, 2026
5a1a920
Merge branch 'main' into jiemingz/cg_diverge
jiemingz Apr 24, 2026
db9e078
Address comments
jiemingz Apr 24, 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
31 changes: 13 additions & 18 deletions megatron/core/transformer/cuda_graphs.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -796,7 +796,14 @@ def create_fwd_graph(self, args, kwargs, outputs=None, clone_inputs=True):
self.kwargs = kwargs
self.outputs = outputs

# save grads and other variables that may be affected by graph warmup
# Save buffers, grads, and other variables that may be affected by graph warmup.
# For example, megatron/core/transformer/moe/router.py's expert_bias is a persistent
# buffer updated each forward pass by '_apply_expert_bias()'. So we need to ensure
# graph capture's forward passes do not corrupt its value.
buffer_backup = []
for buf in self.base_module.buffers():
buffer_backup.append(buf.clone())
Comment thread
jiemingz marked this conversation as resolved.

if self.training and torch.is_grad_enabled():
grad_backup = []
for param in self.base_module.parameters():
Expand Down Expand Up @@ -842,7 +849,6 @@ def create_fwd_graph(self, args, kwargs, outputs=None, clone_inputs=True):
def _resolve_input_buffer(ten):
if not isinstance(ten, ArgMetadata):
return ten

# the input tensor is resued from another cudagraph's input or output
if (
hasattr(ten, "cg_buffer_metadata")
Expand Down Expand Up @@ -913,7 +919,7 @@ def _resolve_input_buffer(ten):
def clone_ten(ten):
if not torch.is_tensor(ten):
return ten
return torch.zeros_like(ten).requires_grad_(ten.requires_grad)
return torch.clone(ten).detach().requires_grad_(ten.requires_grad)
Comment thread
jiemingz marked this conversation as resolved.

warmup_args = tree_map(clone_ten, self.fwd_graph_input_args)
warmup_kwargs = tree_map(clone_ten, self.fwd_graph_input_kwargs)
Expand Down Expand Up @@ -986,17 +992,6 @@ def clone_ten(ten):
o.cg_buffer_metadata.fwd_cudagraph_buffer = fwd_graph_out
fwd_buffer_reuse_ref_count += 1

# if an input buffer requires a copy, and does not have metadata attached to it at this

@deepakn94 deepakn94 Apr 24, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

Why is this being deleted?

@jiemingz jiemingz Apr 24, 2026

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

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

This was causing a convergence issue, due to buffers being reused too early and affecting the correctness of the backward pass.
This optimization was not actually valid since the tensors this optimization was targeting were not actually able to be freed, so I'm just deleting it here. Because it only targets tensors that require an explicit copy into the graph buffer, its also not used for inference.

# point, it will not be reused after this forward pass, so return it to the pool
for buf in self.fwd_graph_input_surface:
if (
hasattr(buf, "can_skip_replay_copy")
and not buf.can_skip_replay_copy
and not hasattr(buf, "cg_buffer_metadata")
):
assert _CudagraphGlobalRecord.tensor_reuse_pool.owns(buf)
_CudagraphGlobalRecord.tensor_reuse_pool.insert(buf)

if self.training and torch.is_grad_enabled():
assert (
len(self.fwd_graph_output_surface) > 0
Expand All @@ -1016,6 +1011,10 @@ def clone_ten(ten):
if main_grad_copy is not None:
param.main_grad.copy_(main_grad_copy)

# restore cached buffers
for buf_copy, buf in zip(buffer_backup, self.base_module.buffers()):
buf.copy_(buf_copy)

if is_moe:
for name in tracker:
tracker[name]["values"].copy_(cached_aux_losses[name])
Expand Down Expand Up @@ -1643,10 +1642,6 @@ def __call__(self, megatron_module, args, kwargs):
runner = self.get_cudagraph_runner(
megatron_module, args, kwargs, self.reuse_cudagraphs
)
# check if a layer is frozen during training.
Comment thread
jiemingz marked this conversation as resolved.
if not torch.is_grad_enabled():
# If the layer is frozen, we need to set the runner to eval mode.
runner.eval()
out = runner.record_graph_capture(args, kwargs)
else:
# No cudagraphs were found in training mode with grad disabled, so fallback to
Expand Down
Loading