From 431d070fcb87a7b652e24ab40cd5ab4506b49223 Mon Sep 17 00:00:00 2001 From: Aidan Do Date: Thu, 22 Jan 2026 20:53:51 -0800 Subject: [PATCH] [docs] Add additional tip for generating less kernels in blockwise (#2940) - Running without this generates a lot of kernels - Clarified CMake configuration for selecting GEMM kernels and added details on kernel generation granularity. --- examples/81_blackwell_gemm_blockwise/README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/examples/81_blackwell_gemm_blockwise/README.md b/examples/81_blackwell_gemm_blockwise/README.md index 4ee65d6d..f42faec0 100644 --- a/examples/81_blackwell_gemm_blockwise/README.md +++ b/examples/81_blackwell_gemm_blockwise/README.md @@ -50,7 +50,7 @@ To determine the most performance Blockwise/Groupwise GEMM or Grouped GEMM kerne All Blockwise/Groupwise GEMMs and Group GEMMs with `f32` scaling of `e4m3` or runtime `f8` types can be selected by selecting a subset of kernels when configuring with CMake by passing: -`-DCUTLASS_LIBRARY_KERNELS="cutlass3x*f32xe4m3_*f32xe4m3*,cutlass3x*f32xf8_*f32xf8*"`. +`-DCUTLASS_LIBRARY_KERNELS="cutlass3x*f32xe4m3_*f32xe4m3*,cutlass3x*f32xf8_*f32xf8*"` (you can further reduce the amount of kernels generated by specifying the SFA and SFB scale granularities e.g., `cutlass3x*1x128f32xe4m3_*128x128f32xe4m3*`). The simplest way to use the profiler is to pass `m`, `n`, and `k` as well as your `scale_vec_size_m`, `scale_vec_size_n`, and `scale_vec_size_k`. Passing `enable-best-kernel-for-fixed-shape` will do some autotuning