Files
miles/docker/amd_patch/latest/tile_kernels.patch
T

37 lines
1.1 KiB
Diff

diff -ru a/tile_kernels/__init__.py b/tile_kernels/__init__.py
--- a/tile_kernels/__init__.py
+++ b/tile_kernels/__init__.py
@@ -1,15 +1,3 @@
-import tilelang
-from . import (
- config,
- engram,
- mhc,
- modeling,
- moe,
- quant,
- transpose,
- torch,
- testing,
-)
from .config import get_num_sms, get_device_num_sms, set_num_sms
diff -ru a/tile_kernels/mhc/post_kernel.py b/tile_kernels/mhc/post_kernel.py
--- a/tile_kernels/mhc/post_kernel.py
+++ b/tile_kernels/mhc/post_kernel.py
@@ -39,7 +39,6 @@
c_local = T.alloc_fragment(mhc, T.float32)
T.copy(a[pid_n, 0, 0], a_local)
T.copy(c[pid_n, 0], c_local)
- T.pdl_sync()
for i0_h in T.Pipelined(T.ceildiv(h, h_blk), num_stages=2):
T.copy(b[pid_n, 0, i0_h * h_blk], b_shared, disable_tma=True)
diff -ru a/tile_kernels/modeling/__init__.py b/tile_kernels/modeling/__init__.py
--- a/tile_kernels/modeling/__init__.py
+++ b/tile_kernels/modeling/__init__.py
@@ -1,2 +0,0 @@
-from . import engram
-from . import mhc