Skip to content

fix: layernorm_fp8 kernel produces garbage scales for all but row 0 #13

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
pauliano22 wants to merge 1 commit into
base: main
Choose a base branch
from
Open

fix: layernorm_fp8 kernel produces garbage scales for all but row 0 #13

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
7 changes: 3 additions & 4 deletions kernels/layer_norm_fp8.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,12 +32,11 @@ def layernorm_quant_kernel(
max_val = tl.max(tl.abs(y_fp32), axis=0)
scale = max_val / 448.0 # 448 is the max value for FP8-E4M3

y_fp8 = (y_fp32 / scale).to(tl.float8e4m3fn)
y_fp8 = (y_fp32 / scale).to(tl.float8e4nv)

# 3. Store the FP8 data and the scale factor
# 3. Store the FP8 data and the per-row scale factor
tl.store(Y_ptr + cols, y_fp8, mask=mask)
if tl.program_id(0) == 0: # Store the scale once for the whole tensor or per row
tl.store(Scale_ptr + row_idx, scale)
tl.store(Scale_ptr + row_idx, scale)

def layernorm_fp8(x, w, b):
M, N = x.shape
Expand Down