Skip to content

dont let dynamo inline inside of NF4 constructors or __torch_dispatch__ #544

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
msaroufim merged 2 commits into from
Jul 28, 2024
Merged

dont let dynamo inline inside of NF4 constructors or __torch_dispatch__ #544

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
8c3b444
dont let dynamo inline inside of NF4 constructors or __torch_dispatch__
bdhirsh Jul 26, 2024
7b10019
allow_in_graph the ctr instead of disabling it
bdhirsh Jul 26, 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
29 changes: 28 additions & 1 deletion torchao/dtypes/nf4tensor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -410,6 +410,7 @@ def get_block_absmax(inpt_tensor: torch.Tensor, block_size: int) -> torch.Tensor
class NF4Tensor(torch.Tensor):
"""NF4Tensor class for converting a weight to the QLoRA NF4 format"""

@torch._dynamo.disable
def __new__(
cls,
# Args related for base tensor construction
Expand Down Expand Up @@ -450,6 +451,7 @@ def __new__(
)
return nf4tensor

@torch._dynamo.disable
def __init__(
self,
tensor_meta: SubclassTensorArgs,
Expand Down Expand Up @@ -758,6 +760,7 @@ def __str__(self):
return self.to(torch.float32).__str__()

@classmethod
@torch._dynamo.disable
def __torch_dispatch__(cls, func, types, args, kwargs=None):
"""TODO we are not supporting torch dispatch at the moment
instead we have created a Autograd.Function to handle the linear
Expand Down Expand Up @@ -849,7 +852,7 @@ def fsdp_post_all_gather(
), f"Expects out's data to be the all-gather output"
return

return NF4Tensor(
return nf4_constructor(
tensor_meta,
block_size,
n_blocks,
Expand Down Expand Up @@ -934,3 +937,27 @@ def function_cpu(*args, **kwargs):
updated_attrs = call_from_inner_tensors(nf4tensor, "cpu", args[1:], kwargs)
updated_attrs["device"] = "cpu"
return NF4Tensor(*construct_nf4_args(nf4tensor, updated_attrs))

@torch._dynamo.allow_in_graph
def nf4_constructor(
tensor_meta: SubclassTensorArgs,
block_size: int,
n_blocks: int,
scaler_block_size: int,
quantized_scalers: torch.Tensor,
quantization_factor: torch.Tensor,
scaler_mean: torch.Tensor,
quantized_data: torch.Tensor,
nf4: torch.Tensor,
):
return NF4Tensor(
tensor_meta,
block_size,
n_blocks,
scaler_block_size,
quantized_scalers,
quantization_factor,
scaler_mean,
quantized_data,
nf4,
)