EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[v0.18.0][Triton][Qwen3.5] delete expr for kernels args (#7646)

Browse Source 
### What this PR does / why we need it?
Some parameters of Triton operators are unnecessarily modified with the
"constexpr" modifier. When these parameters change, recompilation is
triggered, which significantly affects the model performance. Therefore,
these parameters need to be rectified.
backport: https://github.com/vllm-project/vllm-ascend/pull/7482


Signed-off-by: w30012745 <wangxiaoshuai2@h-partners.com>
Co-authored-by: w30012745 <wangxiaoshuai2@h-partners.com>
This commit is contained in:
Mr.WXS
2026-03-25 23:31:27 +08:00
committed by GitHub
parent dd55736ee4
commit dba34d4915
4 changed files with 13 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm_ascend/ops/triton/fla/l2norm.py
View File

@@ -14,8 +14,8 @@ from vllm.triton_utils import tl, triton
from vllm_ascend.ops.triton.triton_utils import get_vectorcore_num
@triton.jit
def l2norm_fwd_kernel2_loop(X, Y, eps, M, N: tl.constexpr, MBLOCK: tl.constexpr, NUM_CHUNKS: tl.constexpr):
@triton.jit(do_not_specialize=["eps", "M", "NUM_CHUNKS"])
def l2norm_fwd_kernel2_loop(X, Y, eps, M, N: tl.constexpr, MBLOCK: tl.constexpr, NUM_CHUNKS):
base_row = tl.program_id(0) * (NUM_CHUNKS * MBLOCK)
rindex = tl.arange(0, N)[None, :]