Skip to content

fixup missing callsites for _scale_e8m0 -> scale rename #3173

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
vkuzo merged 1 commit into from
Oct 14, 2025
Merged

fixup missing callsites for _scale_e8m0 -> scale rename #3173

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
2 changes: 1 addition & 1 deletion benchmarks/float8/profile_lowp_training.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -392,7 +392,7 @@ def cast_with_to_blocked(x_hp):
gemm_kernel_choice=config.gemm_kernel_choice,
)
m, k = x_hp.shape
scale_blocked = to_blocked(x_mx._scale_e8m0.reshape(m, k // config.block_size))
scale_blocked = to_blocked(x_mx.scale.reshape(m, k // config.block_size))
return x_mx._data, scale_blocked

# this function is used for cast_only_dim0_dim1
Expand Down
6 changes: 3 additions & 3 deletions torchao/prototype/moe_training/scaled_grouped_mm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -394,7 +394,7 @@ def backward(ctx, grad_out: torch.Tensor):
scale_calculation_mode=ScaleCalculationMode.FLOOR,
)
grad_out_t_data = grad_out_t_mx.qdata
grad_out_t_scales = grad_out_t_mx._scale_e8m0
grad_out_t_scales = grad_out_t_mx.scale

# Transpose A so we can scale along the M dimension, then un-transpose.
# A shape: (M, K)
Expand All @@ -410,7 +410,7 @@ def backward(ctx, grad_out: torch.Tensor):
scale_calculation_mode=ScaleCalculationMode.FLOOR,
)
A_t_data = A_t_mx.qdata
A_t_scales = A_t_mx._scale_e8m0
A_t_scales = A_t_mx.scale

# Convert scales to blocked format for 2d-2d grouped mm
scale_group_offsets = offs // block_size
Expand Down Expand Up @@ -463,7 +463,7 @@ def _to_mxfp8_dim1_3d(
)
B_data = B_t_mx.qdata.t() # (K, E*N) -> (E*N, K)
B_data = B_data.reshape(E, N, K) # (E*N, K) -> (E, N, K)
B_scales = B_t_mx._scale_e8m0.view(torch.uint8) # (K, E*N//block_size)
B_scales = B_t_mx.scale.view(torch.uint8) # (K, E*N//block_size)
B_scales = B_scales.reshape(
K, E, N // block_size
) # (K, E*N//block_size) -> (K, E, N//block_size)
Expand Down
Loading