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, 03:45 UTC
Add MoE config for Super B200 TP2 (#33510)
shaharmor98 committed
Feb 1, 2026 at 20:48 UTC
8869cd8ec1b2586df96e6556f2d2e3b3760bfb93
1 file changed
+131
vllm/model_executor/layers/fused_moe/configs/E=512,N=1344,device_name=NVIDIA_B200.json
new
+131
@@ -0,0 +1,131 @@
1
+{
2
+ "triton_version": "3.5.1",
3
+ "1": {
4
+ "BLOCK_SIZE_M": 16,
5
+ "BLOCK_SIZE_N": 128,
6
+ "BLOCK_SIZE_K": 128,
7
+ "GROUP_SIZE_M": 16,
8
+ "num_warps": 4,
9
+ "num_stages": 2
10
+ },
11
+ "2": {
12
+ "BLOCK_SIZE_M": 16,
13
+ "BLOCK_SIZE_N": 64,
14
+ "BLOCK_SIZE_K": 128,
15
+ "GROUP_SIZE_M": 64,
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": 128,
23
+ "GROUP_SIZE_M": 64,
24
+ "num_warps": 4,
25
+ "num_stages": 4
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": 4
34
+ },
35
+ "16": {
36
+ "BLOCK_SIZE_M": 16,
37
+ "BLOCK_SIZE_N": 128,
38
+ "BLOCK_SIZE_K": 128,
39
+ "GROUP_SIZE_M": 64,
40
+ "num_warps": 4,
41
+ "num_stages": 4
42
+ },
43
+ "24": {
44
+ "BLOCK_SIZE_M": 16,
45
+ "BLOCK_SIZE_N": 128,
46
+ "BLOCK_SIZE_K": 128,
47
+ "GROUP_SIZE_M": 64,
48
+ "num_warps": 4,
49
+ "num_stages": 4
50
+ },
51
+ "32": {
52
+ "BLOCK_SIZE_M": 16,
53
+ "BLOCK_SIZE_N": 128,
54
+ "BLOCK_SIZE_K": 128,
55
+ "GROUP_SIZE_M": 64,
56
+ "num_warps": 4,
57
+ "num_stages": 4
58
+ },
59
+ "48": {
60
+ "BLOCK_SIZE_M": 16,
61
+ "BLOCK_SIZE_N": 128,
62
+ "BLOCK_SIZE_K": 128,
63
+ "GROUP_SIZE_M": 64,
64
+ "num_warps": 4,
65
+ "num_stages": 4
66
+ },
67
+ "64": {
68
+ "BLOCK_SIZE_M": 16,
69
+ "BLOCK_SIZE_N": 128,
70
+ "BLOCK_SIZE_K": 128,
71
+ "GROUP_SIZE_M": 32,
72
+ "num_warps": 4,
73
+ "num_stages": 4
74
+ },
75
+ "96": {
76
+ "BLOCK_SIZE_M": 16,
77
+ "BLOCK_SIZE_N": 128,
78
+ "BLOCK_SIZE_K": 128,
79
+ "GROUP_SIZE_M": 64,
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": 64,
88
+ "num_warps": 4,
89
+ "num_stages": 3
90
+ },
91
+ "256": {
92
+ "BLOCK_SIZE_M": 16,
93
+ "BLOCK_SIZE_N": 128,
94
+ "BLOCK_SIZE_K": 128,
95
+ "GROUP_SIZE_M": 64,
96
+ "num_warps": 4,
97
+ "num_stages": 3
98
+ },
99
+ "512": {
100
+ "BLOCK_SIZE_M": 32,
101
+ "BLOCK_SIZE_N": 128,
102
+ "BLOCK_SIZE_K": 128,
103
+ "GROUP_SIZE_M": 1,
104
+ "num_warps": 4,
105
+ "num_stages": 2
106
+ },
107
+ "768": {
108
+ "BLOCK_SIZE_M": 64,
109
+ "BLOCK_SIZE_N": 256,
110
+ "BLOCK_SIZE_K": 64,
111
+ "GROUP_SIZE_M": 1,
112
+ "num_warps": 8,
113
+ "num_stages": 5
114
+ },
115
+ "1024": {
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
+ "1536": {
124
+ "BLOCK_SIZE_M": 128,
125
+ "BLOCK_SIZE_N": 256,
126
+ "BLOCK_SIZE_K": 64,
127
+ "GROUP_SIZE_M": 1,
128
+ "num_warps": 8,
129
+ "num_stages": 4
130
+ }
131
+}