Skip to content

[fp8] refactor fp8 linear with compile #5993

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
flybird11111 merged 3 commits into from
Aug 13, 2024
Merged

[fp8] refactor fp8 linear with compile #5993

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
c3e1ae7
[fp8] refactor fp8 linear with compile
ver217 Aug 12, 2024
3529256
[fp8] fix linear test
ver217 Aug 12, 2024
bf4aa02
[fp8] fix linear test
ver217 Aug 13, 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
19 changes: 11 additions & 8 deletions colossalai/quantization/fp8.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,6 +7,8 @@
from packaging.version import Version
from torch.distributed import ReduceOp

SUPPORT_TORCH_COMPILE = Version(torch.__version__) >= Version("2.3.0")


def cast_to_fp8(inp: torch.Tensor, fp8_format="e4m3", per_channel_scale=False) -> Tuple[torch.Tensor, torch.Tensor]:
r"""
Expand Down Expand Up @@ -661,13 +663,14 @@ def backward(ctx: Any, out_grad) -> Any:
return x_grad.reshape(ctx.x_shape), w_grad, bias_grad


if Version(torch.__version__) >= Version("2.3.0"): # TODO failed on torch < 2.3.0

@torch.compile(mode="reduce-overhead", fullgraph=True)
def linear_fp8(input: torch.Tensor, weight: torch.Tensor, bias: Optional[torch.Tensor] = None) -> torch.Tensor:
return _LinearFp8.apply(input, weight, bias)
@torch.compile(mode="reduce-overhead", disable=not SUPPORT_TORCH_COMPILE)
def _linear_fp8(input: torch.Tensor, weight: torch.Tensor, bias: Optional[torch.Tensor] = None) -> torch.Tensor:
return _LinearFp8.apply(input, weight, bias)

else:

def linear_fp8(input: torch.Tensor, weight: torch.Tensor, bias: Optional[torch.Tensor] = None) -> torch.Tensor:
return _LinearFp8.apply(input, weight, bias)
def linear_fp8(input: torch.Tensor, weight: torch.Tensor, bias: Optional[torch.Tensor] = None) -> torch.Tensor:
out = _linear_fp8(input, weight, bias)
if SUPPORT_TORCH_COMPILE:
# avoid modifying the tensor created from cuda graph
out = out.clone()
return out