saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-06-06 00:03:22 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Fix PoliCheck (#18180)

Browse source 
Fix PoliCheck by changing some words, which was from Triton flash
attention's original code.
This commit is contained in:
Vincent Wang 2023-10-31 13:53:11 +08:00 committed by GitHub
parent 08dce54266
commit 1c25fe5580
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
1 changed files with 1 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
orttraining/orttraining/python/training/ort_triton/kernel/_flash_attn.py
View file

@ -393,7 +393,7 @@ def _bwd_kernel_one_col_block(
dk = tl.zeros([BLOCK_N, BLOCK_HEADDIM], dtype=tl.float32)
# There seems to be some problem with Triton pipelining that makes results wrong for
# headdim=64, seqlen=(113, 255), bias_type='matrix'. In this case the for loop
# may have zero step, and pipelining with the bias matrix could screw it up.
# may have zero step, and pipelining with the bias matrix could cause the problem.
# So we just exit early.
if begin_m >= seqlen_q:
dv_ptrs = DV + (offs_n[:, None] * stride_dvn + offs_d[None, :])