A high-throughput and memory-efficient inference and serving engine for LLMs
[Kernel][Perf] Tune fused_moe FP8 config for Qwen3-Next-80B tp=4 on H100 (+25% at batch 96-512) (#44830)
Signed-off-by: Qiuyang Yue <yueqiuyang1389@gmail.com>
Qiuyang Yue committed
Jun 9, 2026 at 04:15 UTC
59401ac9f18c246dceffce28bc878881179405e4
1 file changed
+43
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
new
+43
@@ -0,0 +1,43 @@
1
+{
2
+ "triton_version": "3.6.0",
3
+ "96": {
4
+ "BLOCK_SIZE_M": 16,
5
+ "BLOCK_SIZE_N": 128,
6
+ "BLOCK_SIZE_K": 128,
7
+ "GROUP_SIZE_M": 64,
8
+ "num_warps": 4,
9
+ "num_stages": 2
10
+ },
11
+ "128": {
12
+ "BLOCK_SIZE_M": 16,
13
+ "BLOCK_SIZE_N": 128,
14
+ "BLOCK_SIZE_K": 128,
15
+ "GROUP_SIZE_M": 64,
16
+ "num_warps": 4,
17
+ "num_stages": 2
18
+ },
19
+ "256": {
20
+ "BLOCK_SIZE_M": 16,
21
+ "BLOCK_SIZE_N": 128,
22
+ "BLOCK_SIZE_K": 128,
23
+ "GROUP_SIZE_M": 64,
24
+ "num_warps": 4,
25
+ "num_stages": 2
26
+ },
27
+ "512": {
28
+ "BLOCK_SIZE_M": 16,
29
+ "BLOCK_SIZE_N": 128,
30
+ "BLOCK_SIZE_K": 256,
31
+ "GROUP_SIZE_M": 64,
32
+ "num_warps": 4,
33
+ "num_stages": 2
34
+ },
35
+ "1024": {
36
+ "BLOCK_SIZE_M": 32,
37
+ "BLOCK_SIZE_N": 128,
38
+ "BLOCK_SIZE_K": 128,
39
+ "GROUP_SIZE_M": 64,
40
+ "num_warps": 4,
41
+ "num_stages": 2
42
+ }
43
+}
\ No newline at end of file