Skip to content

Triton current scaling: avoid casting amax input #458

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
matthiasdiener wants to merge 5 commits into
base: dev
Choose a base branch
from
Open

Triton current scaling: avoid casting amax input #458

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
261cb4b
Triton current scaling: avoid casting input to amax
matthiasdiener Feb 20, 2026
86a4103
Merge branch 'dev' into mdiener/triton-cast
matthiasdiener Feb 23, 2026
9492fc5
update copyright year
matthiasdiener Feb 23, 2026
c19cbd2
Merge remote-tracking branch 'origin/dev' into mdiener/triton-cast
matthiasdiener Feb 23, 2026
f8e059e
Merge remote-tracking branch 'origin/dev' into mdiener/triton-cast
matthiasdiener Feb 24, 2026
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
10 changes: 5 additions & 5 deletions transformer_engine/pytorch/triton_kernels/cast_transpose.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,4 +1,4 @@
# Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
# Copyright (c) 2024-2026, Advanced Micro Devices, Inc. All rights reserved.
# License for AMD contributions = MIT. See LICENSE for more information

import torch
Expand Down Expand Up @@ -53,10 +53,10 @@ def _amax_reduce_triton(
A_ptrs = A + rm[:, None] * stride_am + rn[None, :] * stride_an
mask = (rm < M)[:, None] & (rn < N)[None, :]

a = tl.load(A_ptrs, mask=mask, other=0).to(tl.float32)
a = tl.load(A_ptrs, mask=mask, other=0)
tile_amax = tl.max(tl.abs(a))
# accumulate tile-wise max into global amax
tl.atomic_max(amax_ptr, tile_amax, sem='relaxed')
tl.atomic_max(amax_ptr, tile_amax.to(tl.float32), sem='relaxed')
Comment on lines +56 to +59
Copy link

Copilot AI Feb 23, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

PR description still lists placeholder items (“Change A”, “Change B”) and doesn’t describe the actual change (moving float32 cast from load to the reduced amax result). Please update the description to match what this PR does so reviewers/users can understand intent and impact.

Copilot uses AI. Check for mistakes.


@triton.jit
Expand Down Expand Up @@ -229,11 +229,11 @@ def _amax_reduce_triton_stage1(
A_ptrs = A + rm[:, None] * stride_am + rn[None, :] * stride_an
mask = (rm < M)[:, None] & (rn < N)[None, :]

a = tl.load(A_ptrs, mask=mask, other=0).to(tl.float32)
a = tl.load(A_ptrs, mask=mask, other=0)
tile_amax = tl.max(tl.abs(a))

# Store per-program amax in workspace
tl.store(block_amax + pid, tile_amax)
tl.store(block_amax + pid, tile_amax.to(tl.float32))

if pid == 0:
tl.store(num_blocks, tl.num_programs(0))
Expand Down
Loading