Skip to content

[BugFix] Fix SliceSampler for torch.compile compatibility #3298

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
vmoens wants to merge 4 commits into
base: gh/vmoens/179/base
Choose a base branch
from
Open

[BugFix] Fix SliceSampler for torch.compile compatibility #3298

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
674e513
Update
vmoens Jan 8, 2026
b3dea0d
Update
vmoens Jan 8, 2026
7304c9b
Update
vmoens Jan 8, 2026
55d08ab
Update
vmoens Jan 8, 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
20 changes: 18 additions & 2 deletions torchrl/data/replay_buffers/samplers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1556,14 +1556,30 @@ def _get_index(
out_of_traj = relative_starts < 0
if out_of_traj.any():
# a negative start means sampling fewer elements
# Convert seq_length to tensor to avoid torch.compile inductor C++ codegen
# bug with mixed scalar/tensor int64 in blendv operations (see PyTorch #xyz)
seq_length_t = torch.as_tensor(
seq_length,
dtype=relative_starts.dtype,
device=relative_starts.device,
)
seq_length = torch.where(
~out_of_traj, seq_length, seq_length + relative_starts
~out_of_traj, seq_length_t, seq_length_t + relative_starts
)
relative_starts = torch.where(
~out_of_traj, relative_starts, torch.zeros_like(relative_starts)
)
relative_starts = torch.where(~out_of_traj, relative_starts, 0)
if self.span[1]:
out_of_traj = relative_starts + seq_length > lengths[traj_idx]
if out_of_traj.any():
# a negative start means sampling fewer elements
# Convert seq_length to tensor if it's still a scalar
if not isinstance(seq_length, torch.Tensor):
seq_length = torch.as_tensor(
seq_length,
dtype=relative_starts.dtype,
device=relative_starts.device,
)
seq_length = torch.minimum(
seq_length, lengths[traj_idx] - relative_starts
)
Expand Down
Loading