Skip to content

feat(tools): add backward graph generation and validation tools #711

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

Open
Dayuxiaoshui wants to merge 2 commits into
base: develop
Choose a base branch
from
+9 −2
Open

feat(tools): add backward graph generation and validation tools #711

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
67a81d6
feat(tools): add backward graph generation and validation tools
Dayuxiaoshui May 15, 2026
ea832a1
fix: address review feedback - remove is_pure_shape_graph and standal…
Dayuxiaoshui May 18, 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
5 changes: 4 additions & 1 deletion graph_net/torch/extractor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,7 +8,10 @@

torch._dynamo.config.capture_scalar_outputs = True
torch._dynamo.config.capture_dynamic_output_shape_ops = True
torch._dynamo.config.capture_sparse_compute = True
try:
torch._dynamo.config.capture_sparse_compute = True
except AttributeError:
pass
torch._dynamo.config.raise_on_ctx_manager_usage = False
torch._dynamo.config.allow_rnn = True

Expand Down
6 changes: 5 additions & 1 deletion graph_net/torch/sample_pass/backward_graph_extractor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,14 +27,18 @@ def __call__(self):
module, forward_inputs = get_torch_module_and_inputs(
self.model_path, use_dummy_inputs=False, device=self.device
)
module.train()
module.eval()
Copy link
Copy Markdown
Collaborator

@Xreki Xreki May 18, 2026

Choose a reason for hiding this comment

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

eval模式下不会生成反向图吧？

Copy link
Copy Markdown
Contributor Author

@Dayuxiaoshui Dayuxiaoshui May 18, 2026

Choose a reason for hiding this comment

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

model.eval() 不会禁用梯度计算，只有 torch.no_grad() / torch.inference_mode() 才会。eval 仅改变特定层的前向行为（dropout → identity，BatchNorm → 用 running stats 而非 batch stats），反向传播完全正常。而且使用 eval 模式反而更好

Copy link
Copy Markdown
Collaborator

@Xreki Xreki May 18, 2026

Choose a reason for hiding this comment

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

eval模式下不会有反向图吧？


eval_forward_dir = os.path.join(
self.output_dir, "eval_forward", self.rel_model_path
)
if not os.path.exists(eval_forward_dir):
shutil.copytree(self.model_path, eval_forward_dir)

forward_inputs = [
inp.detach().clone() if isinstance(inp, torch.Tensor) else inp
for inp in forward_inputs
]
forward_inputs = self.set_requires_grad_for_forward_inputs(
self.model_path, module, forward_inputs
)
Expand Down
Loading