Skip to content

perf(expt_data): skip redundant TokenStart/TileStart stores in stage1 #4361

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
amd-hhashemi wants to merge 2 commits into
base: main
Choose a base branch
from
Open

perf(expt_data): skip redundant TokenStart/TileStart stores in stage1 #4361

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
ff43f7c
perf(expt_data): skip redundant TokenStart/TileStart stores in stage1
amd-hhashemi Jul 23, 2026
2a4b11e
fix for combined_rounting path
amd-hhashemi Jul 23, 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
28 changes: 18 additions & 10 deletions aiter/ops/triton/_triton_kernels/moe/moe_routing/expt_data.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,17 +33,25 @@ def _expt_data_compute_stage1(
token_acc = tl.zeros([BLOCK], dtype=TokenStart.dtype.element_ty)
tile_acc = tl.zeros([BLOCK], dtype=TileStart.dtype.element_ty)
offs_n = tl.arange(0, BLOCK)
# pid==0 must run the full loop: it reads TileStart[n_expts_tot-1]
# below for the sentinel, so it must have written it first.
# All other blocks can stop once they've written their own chunk.
done = False
for i in range(0, n_expts_tot, BLOCK):
mask_n = offs_n < n_expts_tot
hist_token = tl.load(Hist + offs_n, mask=mask_n, other=0)
hist_tile = _cdiv_pow2(hist_token, tile_dim_log2)
token_starts = tl.cumsum(hist_token, 0) - hist_token + token_acc
tile_starts = tl.cumsum(hist_tile, 0) - hist_tile + tile_acc
token_acc += tl.sum(hist_token, 0)
tile_acc += tl.sum(hist_tile, 0)
tl.store(TokenStart + offs_n, token_starts)
tl.store(TileStart + offs_n, tile_starts)
offs_n += BLOCK
if not done:
mask_n = offs_n < n_expts_tot
hist_token = tl.load(Hist + offs_n, mask=mask_n, other=0)
hist_tile = _cdiv_pow2(hist_token, tile_dim_log2)
token_starts = tl.cumsum(hist_token, 0) - hist_token + token_acc
tile_starts = tl.cumsum(hist_tile, 0) - hist_tile + tile_acc
token_acc += tl.sum(hist_token, 0)
tile_acc += tl.sum(hist_tile, 0)
if pid == 0 or pid < i + BLOCK:
tl.store(TokenStart + offs_n, token_starts, mask=mask_n)
tl.store(TileStart + offs_n, tile_starts, mask=mask_n)
if pid != 0 and pid < i + BLOCK:
done = True
Comment thread
amd-hhashemi marked this conversation as resolved.
Outdated
offs_n += BLOCK

if pid == 0:
tl.store(TokenStart + n_expts_tot, n_gates)
Expand Down
Loading