Skip to content

Add portable quantized stacked-halves RoPE operator (#22423) #22423

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
ethansfng wants to merge 1 commit into
base: main
Choose a base branch
from
+53 −0
Open

Add portable quantized stacked-halves RoPE operator (#22423) #22423

Show file tree
Hide file tree
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
21 changes: 21 additions & 0 deletions backends/cadence/aot/ops_registrations.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -489,6 +489,13 @@ def register_fake(
"rope_rotate_stacked_halves.out(Tensor input, Tensor sin_tensor, Tensor cos_tensor, Tensor? pos, *, Tensor(a!) out) -> Tensor(a!)"
)

lib.define(
"quantized_rope_rotate_stacked_halves(Tensor input, Tensor sin_tensor, Tensor cos_tensor, Tensor? pos, float in_scale, int in_zero_point, float out_scale, int out_zero_point) -> (Tensor out)"
)
lib.define(
"quantized_rope_rotate_stacked_halves.out(Tensor input, Tensor sin_tensor, Tensor cos_tensor, Tensor? pos, float in_scale, int in_zero_point, float out_scale, int out_zero_point, *, Tensor(a!) out) -> Tensor(a!)"
)

lib.define(
"quantized_softmax(Tensor input, Tensor mask, int dim, int mask_type, Tensor pos, Tensor in_scale, Tensor in_zero_point, Tensor out_scale, Tensor out_zero_point) -> (Tensor out)"
)
Expand Down Expand Up @@ -3136,6 +3143,20 @@ def rope_rotate_stacked_halves_meta(
return input.new_empty(input.shape, dtype=input.dtype)


@register_fake("cadence::quantized_rope_rotate_stacked_halves")
def quantized_rope_rotate_stacked_halves_meta(
input: torch.Tensor,
sin_tensor: torch.Tensor,
cos_tensor: torch.Tensor,
pos: Optional[torch.Tensor],
in_scale: float,
in_zero_point: int,
out_scale: float,
out_zero_point: int,
) -> torch.Tensor:
return rope_rotate_stacked_halves_meta(input, sin_tensor, cos_tensor, pos)


@register_fake("cadence::idma_copy")
def copy_idma_copy_impl(
src: torch.Tensor,
Expand Down
32 changes: 32 additions & 0 deletions backends/cadence/aot/ref_implementations.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2265,6 +2265,38 @@ def rope_rotate_stacked_halves(
return rotated.view(original_shape)


@impl_tracked(m, "quantized_rope_rotate_stacked_halves")
def quantized_rope_rotate_stacked_halves(
input_tensor: torch.Tensor,
sin_tensor: torch.Tensor,
cos_tensor: torch.Tensor,
pos: torch.Tensor | None,
in_scale: float,
in_zero_point: int,
out_scale: float,
out_zero_point: int,
) -> torch.Tensor:
dtype = input_tensor.dtype
dtype_limits = torch.iinfo(dtype)
dequantized = dequantize_per_tensor_common(
input_tensor,
in_scale,
in_zero_point,
dtype_limits.min,
dtype_limits.max,
dtype,
)
rotated = rope_rotate_stacked_halves(dequantized, sin_tensor, cos_tensor, pos)
return quantize_per_tensor_common(
rotated,
out_scale,
out_zero_point,
dtype_limits.min,
dtype_limits.max,
dtype,
)


@impl_tracked(m, "im2row")
def im2row(
input_tensor: torch.Tensor,
Expand Down
Loading