Skip to content

Loosen the late-iteration gradient tolerance in the checkpointing test #8648

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
pengdurice merged 1 commit into from
Sep 24, 2026
+16 −6
Merged

Loosen the late-iteration gradient tolerance in the checkpointing test #8648

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
22 changes: 16 additions & 6 deletions tests/unit/v1/zero/test_zero_user_backward.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -165,12 +165,13 @@ def collect_ddp_gradients(model_ddp):
return grads


def compare_gradients(grads_ddp, grads_ds, step_info=""):
def compare_gradients(grads_ddp, grads_ds, step_info="", **tolerance_kwargs):
"""Compare gradients between DDP and DeepSpeed.

Uses PyTorch's default tolerances for the tensor dtype (e.g., for bfloat16:
rtol=1.6e-2, atol=1e-5). The 2-layer model keeps differences small enough
to pass with default tolerances even after multiple optimizer steps.
rtol=1.6e-2, atol=1e-5) unless tolerance_kwargs overrides them. The 2-layer
model keeps differences small enough to pass with default tolerances even
after multiple optimizer steps.
"""
step_suffix = f" at {step_info}" if step_info else ""
assert len(grads_ddp) == len(grads_ds), \
Expand All @@ -184,7 +185,10 @@ def compare_gradients(grads_ddp, grads_ds, step_info=""):
if grad_ds.dtype != grad_ddp.dtype:
grad_ds = grad_ds.to(grad_ddp.dtype)
# Use PyTorch's default tolerances for the dtype
allclose_on_all_ranks(grad_ddp, grad_ds, assert_message=f"Gradients differ for parameter {name}{step_suffix}")
allclose_on_all_ranks(grad_ddp,
grad_ds,
assert_message=f"Gradients differ for parameter {name}{step_suffix}",
**tolerance_kwargs)


def collect_ddp_parameters(model_ddp):
Expand Down Expand Up @@ -1563,8 +1567,14 @@ def test_checkpointed_multiple_backward(self, zero_stage, use_reentrant):
f"No gradients at iteration {iteration} with use_reentrant={use_reentrant}"

# Compare gradients with DDP - using same optimizer so should match closely
# Small differences at later iterations are expected due to bfloat16 precision
compare_gradients(ddp_grads, ds_grads, f"iteration {iteration} with use_reentrant={use_reentrant}")
# Small differences at later iterations are expected due to bfloat16 precision:
# the engine keeps fp32 accounting while the DDP reference steps in bf16, so
# weights drift apart by ~one bf16 ulp per step; by iteration 2 that crosses a
# reduction/relu rounding boundary and yields ~8e-3 absolute gradient diffs on
# O(1) gradients, which dtype-default tolerances reject on some bf16 kernels.
late_iteration_tol = {"rtol": 1.6e-2, "atol": 1e-2} if iteration >= 2 else {}
compare_gradients(ddp_grads, ds_grads, f"iteration {iteration} with use_reentrant={use_reentrant}",
**late_iteration_tol)

# Run optimizer steps on both models
optimizer_ddp.step()
Expand Down
Loading