Skip to content

[BugFix][Quant] Accumulate E5M6 BF16 amax in FP32 #16

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
JayceSu98 wants to merge 1 commit into
base: main
Choose a base branch
from
+1 −1
Open

[BugFix][Quant] Accumulate E5M6 BF16 amax in FP32 #16

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 tile_kernels/quant/per_token_cast_to_e5m6_kernel.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -119,7 +119,7 @@ def per_token_cast_to_e5m6_kernel(
# Copy input into registers
T.copy(x[pid_token * block_m, pid_hidden * block_k], x_fragment)

amax_fragment = T.alloc_fragment((block_m, num_groups), in_config.dtype)
amax_fragment = T.alloc_fragment((block_m, num_groups), T.float32)
x_fragment_reshaped = T.reshape(x_fragment, [block_m, num_groups, num_per_channels])
# Reduce SF
T.reduce_absmax(x_fragment_reshaped, amax_fragment, dim=2)
Expand Down