iHore/vLLM
8
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

[Kernel][Perf] Tune fused_moe FP8 config for Qwen3-Next-80B tp=4 on H100 (+25% at batch 96-512) (#44830)

Browse Source 
Signed-off-by: Qiuyang Yue <yueqiuyang1389@gmail.com>
This commit is contained in:
Qiuyang Yue
2026-06-09 04:15:51 -07:00
committed by GitHub
parent d841386d27
commit 59401ac9f1
1 changed files with 43 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
vllm/model_executor/layers/fused_moe/configs/E=512,N=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json
+43
View File
@@ -0,0 +1,43 @@
{
"triton_version": "3.6.0",
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 2
},
"128": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 2
},
"256": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 2
},
"512": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 2
},
"1024": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 2
}
}