biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[ROCm][MoE] mi300 mixtral8x7B perf for specific BS (#13577)

Browse Source 
Signed-off-by: Divakar Verma <divakar.verma@amd.com>
This commit is contained in:
Divakar Verma
2025-02-19 22:01:02 -06:00
committed by GitHub
parent 88f6ba3281
commit 0d243f2a54
3 changed files with 5 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/model_executor/layers/fused_moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X.json
View File

@@ -45,8 +45,8 @@
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 2,

4
vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X.json
View File

@@ -45,8 +45,8 @@
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 2,
"num_stages": 2,

2
vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X.json
View File

@@ -128,7 +128,7 @@
"num_warps": 8,
"num_stages": 2,
"waves_per_eu": 0,
"matrix_instr_nonkdim": 32,
"matrix_instr_nonkdim": 16,
"kpack": 2
},
"512": {