From a69583f70f8f94c16e30f42ee535a897f8ccb8c4 Mon Sep 17 00:00:00 2001 From: Alan Kao Date: Thu, 24 Sep 2026 19:06:40 +0800 Subject: [PATCH] [AMD] Add tuned dsv4 shape (#40996) --- ...dtype=fp8_w8a8,block_shape=[128, 128].json | 119 ++++++++++++++++++ ...dtype=fp8_w8a8,block_shape=[128, 128].json | 47 +++++++ ...dtype=fp8_w8a8,block_shape=[128, 128].json | 119 ++++++++++++++++++ ...dtype=fp8_w8a8,block_shape=[128, 128].json | 119 ++++++++++++++++++ ...dtype=fp8_w8a8,block_shape=[128, 128].json | 119 ++++++++++++++++++ ...dtype=fp8_w8a8,block_shape=[128, 128].json | 119 ++++++++++++++++++ ...dtype=fp8_w8a8,block_shape=[128, 128].json | 119 ++++++++++++++++++ ...dtype=fp8_w8a8,block_shape=[128, 128].json | 119 ++++++++++++++++++ 8 files changed, 880 insertions(+) create mode 100644 python/sglang/kernels/ops/quantization/configs/N=1024,K=4096,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json create mode 100644 python/sglang/kernels/ops/quantization/configs/N=2048,K=4096,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json create mode 100644 python/sglang/kernels/ops/quantization/configs/N=32768,K=1024,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json create mode 100644 python/sglang/kernels/ops/quantization/configs/N=4096,K=2048,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json create mode 100644 python/sglang/kernels/ops/quantization/configs/N=4096,K=4096,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json create mode 100644 python/sglang/kernels/ops/quantization/configs/N=4096,K=8192,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json create mode 100644 python/sglang/kernels/ops/quantization/configs/N=512,K=4096,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json create mode 100644 python/sglang/kernels/ops/quantization/configs/N=8192,K=4096,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json diff --git a/python/sglang/kernels/ops/quantization/configs/N=1024,K=4096,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json b/python/sglang/kernels/ops/quantization/configs/N=1024,K=4096,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000..1de0eb9909 --- /dev/null +++ b/python/sglang/kernels/ops/quantization/configs/N=1024,K=4096,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,119 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 2 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 1, + "num_stages": 1, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 1, + "num_stages": 1, + "waves_per_eu": 0 + } +} diff --git a/python/sglang/kernels/ops/quantization/configs/N=2048,K=4096,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json b/python/sglang/kernels/ops/quantization/configs/N=2048,K=4096,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000..adf2a60b42 --- /dev/null +++ b/python/sglang/kernels/ops/quantization/configs/N=2048,K=4096,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,47 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + } +} diff --git a/python/sglang/kernels/ops/quantization/configs/N=32768,K=1024,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json b/python/sglang/kernels/ops/quantization/configs/N=32768,K=1024,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000..1b944b5ced --- /dev/null +++ b/python/sglang/kernels/ops/quantization/configs/N=32768,K=1024,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,119 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 1, + "waves_per_eu": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 2 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 1, + "num_stages": 1, + "waves_per_eu": 2 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 1, + "num_stages": 1, + "waves_per_eu": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 1, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 32, + "num_warps": 1, + "num_stages": 1, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 2 + } +} diff --git a/python/sglang/kernels/ops/quantization/configs/N=4096,K=2048,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json b/python/sglang/kernels/ops/quantization/configs/N=4096,K=2048,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000..1b4346a290 --- /dev/null +++ b/python/sglang/kernels/ops/quantization/configs/N=4096,K=2048,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,119 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 2 + } +} diff --git a/python/sglang/kernels/ops/quantization/configs/N=4096,K=4096,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json b/python/sglang/kernels/ops/quantization/configs/N=4096,K=4096,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000..2b7ac9cad8 --- /dev/null +++ b/python/sglang/kernels/ops/quantization/configs/N=4096,K=4096,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,119 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 1, + "num_stages": 1, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 0 + } +} diff --git a/python/sglang/kernels/ops/quantization/configs/N=4096,K=8192,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json b/python/sglang/kernels/ops/quantization/configs/N=4096,K=8192,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000..d352310e25 --- /dev/null +++ b/python/sglang/kernels/ops/quantization/configs/N=4096,K=8192,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,119 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 1, + "num_stages": 1, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 1, + "num_stages": 1, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 1, + "num_stages": 1, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 2 + } +} diff --git a/python/sglang/kernels/ops/quantization/configs/N=512,K=4096,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json b/python/sglang/kernels/ops/quantization/configs/N=512,K=4096,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000..b99fa915e9 --- /dev/null +++ b/python/sglang/kernels/ops/quantization/configs/N=512,K=4096,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,119 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 2 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 2 + }, + "1536": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 2 + } +} diff --git a/python/sglang/kernels/ops/quantization/configs/N=8192,K=4096,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json b/python/sglang/kernels/ops/quantization/configs/N=8192,K=4096,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000..3a52f57fb6 --- /dev/null +++ b/python/sglang/kernels/ops/quantization/configs/N=8192,K=4096,device_name=AMD_Instinct_MI455X,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,119 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 1, + "waves_per_eu": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 8, + "num_warps": 1, + "num_stages": 1, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 1, + "waves_per_eu": 0 + } +}