From f43ef1531d10d556e616abb8e069ecb24578bde7 Mon Sep 17 00:00:00 2001 From: Jiangyun Zhu Date: Mon, 7 Sep 2026 11:54:22 +0800 Subject: [PATCH] [CI] fix pre-commit (#55630) Signed-off-by: zjy0516 --- .../w8a8/cutlass/c3x/scaled_mm_blockwise_sm120_fp8.cu | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/csrc/libtorch_stable/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm120_fp8.cu b/csrc/libtorch_stable/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm120_fp8.cu index ecf9538d34..021c16d668 100644 --- a/csrc/libtorch_stable/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm120_fp8.cu +++ b/csrc/libtorch_stable/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm120_fp8.cu @@ -11,8 +11,8 @@ namespace { // On a GB10 (24 MiB L2) the blockwise kernel loses most of its throughput once // the weight is re-streamed from DRAM per row of M tiles: a 16384x2560 FP8 // weight runs at 165 TFLOPS at M=4096 but 90 at M=8192 and 52 at M>=16384; -// 8192x8192 is at 54 from M=6144. With the tile scheduler's max_swizzle_size = 8 -// the same launches run at 150-174 TFLOPS at every M, bit-identical to the +// 8192x8192 is at 54 from M=6144. With the tile scheduler's max_swizzle_size = +// 8 the same launches run at 150-174 TFLOPS at every M, bit-identical to the // default order (ten N/K shapes, M 2048-16384, all cells identical). The one // place the default order is better is a narrow band around M=4096 on the // 2560-wide weights (167 vs 153 at 16384x2560, 160 vs 154 at 12288x2560);