[CI] fix pre-commit (#55630)
Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
Jiangyun Zhu committed
Sep 7, 2026 at 11:54 UTC
f43ef1531d10d556e616abb8e069ecb24578bde7
1 file changed
+2
-2
csrc/libtorch_stable/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm120_fp8.cu
+2
-2
@@ -11,8 +11,8 @@ namespace {
11
// On a GB10 (24 MiB L2) the blockwise kernel loses most of its throughput once
12
// the weight is re-streamed from DRAM per row of M tiles: a 16384x2560 FP8
13
// weight runs at 165 TFLOPS at M=4096 but 90 at M=8192 and 52 at M>=16384;
14
-// 8192x8192 is at 54 from M=6144. With the tile scheduler's max_swizzle_size = 8
15
-// the same launches run at 150-174 TFLOPS at every M, bit-identical to the
14
+// 8192x8192 is at 54 from M=6144. With the tile scheduler's max_swizzle_size =
15
+// 8 the same launches run at 150-174 TFLOPS at every M, bit-identical to the
16
// default order (ten N/K shapes, M 2048-16384, all cells identical). The one
17
// place the default order is better is a narrow band around M=4096 on the
18
// 2560-wide weights (167 vs 153 at 16384x2560, 160 vs 154 at 12288x2560);