A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 9, 10:00 UTC
Add Triton fused MoE config for B200 (Nemotron Nano) (#32804)
danisereb committed
Jan 29, 2026 at 21:21 UTC
8e2a469b3b2f67bc900ed72724fe3f05e3564994
1 file changed
+139
vllm/model_executor/layers/fused_moe/configs/E=128,N=1856,device_name=NVIDIA_B200.json
new
+139
@@ -0,0 +1,139 @@
1
+{
2
+ "triton_version": "3.5.1",
3
+ "1": {
4
+ "BLOCK_SIZE_M": 16,
5
+ "BLOCK_SIZE_N": 64,
6
+ "BLOCK_SIZE_K": 64,
7
+ "GROUP_SIZE_M": 64,
8
+ "num_warps": 4,
9
+ "num_stages": 5
10
+ },
11
+ "2": {
12
+ "BLOCK_SIZE_M": 16,
13
+ "BLOCK_SIZE_N": 64,
14
+ "BLOCK_SIZE_K": 128,
15
+ "GROUP_SIZE_M": 1,
16
+ "num_warps": 4,
17
+ "num_stages": 3
18
+ },
19
+ "4": {
20
+ "BLOCK_SIZE_M": 16,
21
+ "BLOCK_SIZE_N": 128,
22
+ "BLOCK_SIZE_K": 64,
23
+ "GROUP_SIZE_M": 1,
24
+ "num_warps": 4,
25
+ "num_stages": 3
26
+ },
27
+ "8": {
28
+ "BLOCK_SIZE_M": 16,
29
+ "BLOCK_SIZE_N": 128,
30
+ "BLOCK_SIZE_K": 128,
31
+ "GROUP_SIZE_M": 32,
32
+ "num_warps": 4,
33
+ "num_stages": 3
34
+ },
35
+ "16": {
36
+ "BLOCK_SIZE_M": 16,
37
+ "BLOCK_SIZE_N": 64,
38
+ "BLOCK_SIZE_K": 128,
39
+ "GROUP_SIZE_M": 1,
40
+ "num_warps": 4,
41
+ "num_stages": 3
42
+ },
43
+ "24": {
44
+ "BLOCK_SIZE_M": 16,
45
+ "BLOCK_SIZE_N": 64,
46
+ "BLOCK_SIZE_K": 128,
47
+ "GROUP_SIZE_M": 32,
48
+ "num_warps": 4,
49
+ "num_stages": 3
50
+ },
51
+ "32": {
52
+ "BLOCK_SIZE_M": 16,
53
+ "BLOCK_SIZE_N": 64,
54
+ "BLOCK_SIZE_K": 128,
55
+ "GROUP_SIZE_M": 1,
56
+ "num_warps": 4,
57
+ "num_stages": 3
58
+ },
59
+ "48": {
60
+ "BLOCK_SIZE_M": 16,
61
+ "BLOCK_SIZE_N": 64,
62
+ "BLOCK_SIZE_K": 128,
63
+ "GROUP_SIZE_M": 1,
64
+ "num_warps": 4,
65
+ "num_stages": 3
66
+ },
67
+ "64": {
68
+ "BLOCK_SIZE_M": 16,
69
+ "BLOCK_SIZE_N": 128,
70
+ "BLOCK_SIZE_K": 128,
71
+ "GROUP_SIZE_M": 1,
72
+ "num_warps": 4,
73
+ "num_stages": 3
74
+ },
75
+ "96": {
76
+ "BLOCK_SIZE_M": 16,
77
+ "BLOCK_SIZE_N": 64,
78
+ "BLOCK_SIZE_K": 128,
79
+ "GROUP_SIZE_M": 1,
80
+ "num_warps": 4,
81
+ "num_stages": 4
82
+ },
83
+ "128": {
84
+ "BLOCK_SIZE_M": 16,
85
+ "BLOCK_SIZE_N": 128,
86
+ "BLOCK_SIZE_K": 128,
87
+ "GROUP_SIZE_M": 1,
88
+ "num_warps": 4,
89
+ "num_stages": 4
90
+ },
91
+ "256": {
92
+ "BLOCK_SIZE_M": 16,
93
+ "BLOCK_SIZE_N": 128,
94
+ "BLOCK_SIZE_K": 128,
95
+ "GROUP_SIZE_M": 1,
96
+ "num_warps": 4,
97
+ "num_stages": 3
98
+ },
99
+ "320": {
100
+ "BLOCK_SIZE_M": 16,
101
+ "BLOCK_SIZE_N": 128,
102
+ "BLOCK_SIZE_K": 128,
103
+ "GROUP_SIZE_M": 1,
104
+ "num_warps": 4,
105
+ "num_stages": 3
106
+ },
107
+ "512": {
108
+ "BLOCK_SIZE_M": 32,
109
+ "BLOCK_SIZE_N": 128,
110
+ "BLOCK_SIZE_K": 128,
111
+ "GROUP_SIZE_M": 1,
112
+ "num_warps": 4,
113
+ "num_stages": 2
114
+ },
115
+ "768": {
116
+ "BLOCK_SIZE_M": 64,
117
+ "BLOCK_SIZE_N": 256,
118
+ "BLOCK_SIZE_K": 64,
119
+ "GROUP_SIZE_M": 1,
120
+ "num_warps": 8,
121
+ "num_stages": 5
122
+ },
123
+ "1024": {
124
+ "BLOCK_SIZE_M": 64,
125
+ "BLOCK_SIZE_N": 128,
126
+ "BLOCK_SIZE_K": 64,
127
+ "GROUP_SIZE_M": 1,
128
+ "num_warps": 8,
129
+ "num_stages": 3
130
+ },
131
+ "1536": {
132
+ "BLOCK_SIZE_M": 128,
133
+ "BLOCK_SIZE_N": 256,
134
+ "BLOCK_SIZE_K": 64,
135
+ "GROUP_SIZE_M": 1,
136
+ "num_warps": 8,
137
+ "num_stages": 4
138
+ }
139
+}