Skip to content

Add empty_like for NF4Tensor to support offloading #881

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
janeyx99 merged 5 commits into from
Sep 16, 2024
Merged

Add empty_like for NF4Tensor to support offloading #881

Changes from 1 commit
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
80b0325
Add overloads for NF4Tensor to support offloading
janeyx99 Sep 12, 2024
8fe70c6
Remove unneeded API, add test
janeyx99 Sep 16, 2024
3c3e090
lint
janeyx99 Sep 16, 2024
659fa8f
Fix up test and realize it was not doing what you thought
janeyx99 Sep 16, 2024
04a5b92
ruff
janeyx99 Sep 16, 2024
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
8 changes: 8 additions & 0 deletions torchao/dtypes/nf4tensor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -156,6 +156,7 @@ def clone(func, *args, **kwargs):
@implements(
[
aten.detach.default,
aten.empty_like.default,
]
)
def nf4_detach(aten_op, args, kwargs=None):
Expand Down Expand Up @@ -956,6 +957,13 @@ def decorator(func):
@implements_torch_function(torch.Tensor.to)
def function_to_dtype(*args, **kwargs):
tensor = args[0]
if len(args) <= 1:
if "device" in kwargs:
# Tensor.to(device, non_blocking)
device = kwargs["device"]
updated_attrs = call_from_inner_tensors(tensor, "to", args[1:], kwargs)
updated_attrs["device"] = device
return NF4Tensor(*construct_nf4_args(tensor, updated_attrs))
if isinstance(args[1], torch.dtype):
# Tensor.to(dtype, non_blocking, copy, memory_format)
return tensor.get_original_weight().to(*args[1:], **kwargs)
Expand Down