[ROCm] add tuned moe_wna16_triton kernel configs for CDNA4 (#35093)

Signed-off-by: salykova <amsalykov@gmail.com>
Signed-off-by: amd-asalykov <asalykov@amd.com>
This commit is contained in:
Amanzhol Salykov
2026-03-11 20:00:08 +01:00
committed by GitHub
parent 65986db6ba
commit 8a24842765
4 changed files with 768 additions and 0 deletions

View File

@@ -0,0 +1,192 @@
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 1,
"matrix_instr_nonkdim": 16
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 1,
"matrix_instr_nonkdim": 16
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 1,
"matrix_instr_nonkdim": 16
},
"128": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"256": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"512": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 32
},
"1024": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 32
},
"1536": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 32
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 1,
"matrix_instr_nonkdim": 16
},
"3072": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 1,
"matrix_instr_nonkdim": 16
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 8,
"num_stages": 2,
"matrix_instr_nonkdim": 32
}
}

View File

@@ -0,0 +1,192 @@
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 1,
"matrix_instr_nonkdim": 16
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 1,
"matrix_instr_nonkdim": 16
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 1,
"matrix_instr_nonkdim": 16
},
"128": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"256": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"512": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 32
},
"1024": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 32
},
"1536": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 32
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 1,
"matrix_instr_nonkdim": 16
},
"3072": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 1,
"matrix_instr_nonkdim": 16
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 8,
"num_stages": 2,
"matrix_instr_nonkdim": 32
}
}

View File

@@ -0,0 +1,192 @@
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 1,
"matrix_instr_nonkdim": 16
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 1,
"matrix_instr_nonkdim": 16
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 1,
"matrix_instr_nonkdim": 16
},
"128": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"256": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"512": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 32
},
"1024": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 32
},
"1536": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 32
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 1,
"matrix_instr_nonkdim": 16
},
"3072": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 1,
"matrix_instr_nonkdim": 16
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 8,
"num_stages": 2,
"matrix_instr_nonkdim": 32
}
}

View File

@@ -0,0 +1,192 @@
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 1,
"matrix_instr_nonkdim": 16
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 1,
"matrix_instr_nonkdim": 16
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 1,
"matrix_instr_nonkdim": 16
},
"128": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"256": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"512": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 32
},
"1024": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 32
},
"1536": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 32
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 1,
"matrix_instr_nonkdim": 16
},
"3072": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 1,
"matrix_instr_nonkdim": 16
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 4,
"num_stages": 2,
"matrix_instr_nonkdim": 16
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_warps": 8,
"num_stages": 2,
"matrix_instr_nonkdim": 32
}
}