From c0215ce663e71824ebed72e8c56f5f0c693ae3b3 Mon Sep 17 00:00:00 2001 From: okone1995 <161808120+okone1995@users.noreply.github.com> Date: Sat, 15 Aug 2026 23:20:37 +0800 Subject: [PATCH 1/2] csrc: allow gfx1100 (RDNA3) in GPU_ARCHS allow-list (ref #4604) --- csrc/cpp_itfs/utils.py | 1 + 1 file changed, 1 insertion(+) diff --git a/csrc/cpp_itfs/utils.py b/csrc/cpp_itfs/utils.py index 5687e194cb..11bdc6e670 100644 --- a/csrc/cpp_itfs/utils.py +++ b/csrc/cpp_itfs/utils.py @@ -179,6 +179,7 @@ def validate_and_update_archs(): "gfx941", "gfx942", "gfx950", + "gfx1100", "gfx1151", ] From a22484305f5e6dc1bbe0604a4d3903c3c739cd7e Mon Sep 17 00:00:00 2001 From: okone1995 <161808120+okone1995@users.noreply.github.com> Date: Sat, 15 Aug 2026 23:20:50 +0800 Subject: [PATCH 2/2] triton/gemm: add gfx1100 GEMM-A8W8 tuning config (ref #4604) --- .../configs/gemm/gfx1100-GEMM-A8W8.json | 38 +++++++++++++++++++ 1 file changed, 38 insertions(+) create mode 100644 aiter/ops/triton/configs/gemm/gfx1100-GEMM-A8W8.json diff --git a/aiter/ops/triton/configs/gemm/gfx1100-GEMM-A8W8.json b/aiter/ops/triton/configs/gemm/gfx1100-GEMM-A8W8.json new file mode 100644 index 0000000000..06cbffaa02 --- /dev/null +++ b/aiter/ops/triton/configs/gemm/gfx1100-GEMM-A8W8.json @@ -0,0 +1,38 @@ +{ + "M_LEQ_32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 3, + "waves_per_eu": 2, + "matrix_instr_nonkdim": 16, + "kpack": 1, + "NUM_KSPLIT": 1 + }, + "M_GEQ_64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 2, + "matrix_instr_nonkdim": 16, + "kpack": 1, + "NUM_KSPLIT": 1 + }, + "any": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 3, + "waves_per_eu": 2, + "matrix_instr_nonkdim": 16, + "kpack": 1, + "NUM_KSPLIT": 1 + } +}