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 12, 22:15 UTC
[Kernels] Improve H200 Fused MoE Config (#28992)
Signed-off-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
Robert Shaw committed
Nov 19, 2025 at 14:23 UTC
fe69f331f84d99541564dfe4852dd45220ed7875
1 file changed
+36
-36
vllm/model_executor/layers/fused_moe/configs/E=256,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json
+36
-36
@@ -1,11 +1,11 @@
1
{
2
"1": {
3
"BLOCK_SIZE_M": 16,
4
- "BLOCK_SIZE_N": 64,
4
+ "BLOCK_SIZE_N": 128,
5
"BLOCK_SIZE_K": 128,
6
- "GROUP_SIZE_M": 1,
6
+ "GROUP_SIZE_M": 16,
7
"num_warps": 4,
8
- "num_stages": 5
8
+ "num_stages": 4
9
},
10
"2": {
11
"BLOCK_SIZE_M": 16,
@@ -13,82 +13,82 @@
13
"BLOCK_SIZE_K": 128,
14
"GROUP_SIZE_M": 16,
15
"num_warps": 4,
16
- "num_stages": 3
16
+ "num_stages": 4
17
},
18
"4": {
19
- "BLOCK_SIZE_M": 64,
20
- "BLOCK_SIZE_N": 64,
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 128,
21
"BLOCK_SIZE_K": 128,
22
"GROUP_SIZE_M": 64,
23
"num_warps": 4,
24
"num_stages": 4
25
},
26
"8": {
27
- "BLOCK_SIZE_M": 64,
27
+ "BLOCK_SIZE_M": 16,
28
"BLOCK_SIZE_N": 128,
29
"BLOCK_SIZE_K": 128,
30
- "GROUP_SIZE_M": 64,
30
+ "GROUP_SIZE_M": 16,
31
"num_warps": 4,
32
- "num_stages": 3
32
+ "num_stages": 4
33
},
34
"16": {
35
"BLOCK_SIZE_M": 16,
36
- "BLOCK_SIZE_N": 256,
37
- "BLOCK_SIZE_K": 64,
38
- "GROUP_SIZE_M": 32,
36
+ "BLOCK_SIZE_N": 128,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 16,
39
"num_warps": 4,
40
"num_stages": 3
41
},
42
"24": {
43
- "BLOCK_SIZE_M": 64,
43
+ "BLOCK_SIZE_M": 16,
44
"BLOCK_SIZE_N": 128,
45
"BLOCK_SIZE_K": 128,
46
"GROUP_SIZE_M": 16,
47
"num_warps": 4,
48
- "num_stages": 3
48
+ "num_stages": 4
49
},
50
"32": {
51
- "BLOCK_SIZE_M": 64,
51
+ "BLOCK_SIZE_M": 16,
52
"BLOCK_SIZE_N": 128,
53
"BLOCK_SIZE_K": 128,
54
- "GROUP_SIZE_M": 1,
54
+ "GROUP_SIZE_M": 16,
55
"num_warps": 4,
56
- "num_stages": 3
56
+ "num_stages": 5
57
},
58
"48": {
59
- "BLOCK_SIZE_M": 64,
59
+ "BLOCK_SIZE_M": 16,
60
"BLOCK_SIZE_N": 128,
61
"BLOCK_SIZE_K": 128,
62
- "GROUP_SIZE_M": 32,
62
+ "GROUP_SIZE_M": 64,
63
"num_warps": 4,
64
"num_stages": 3
65
},
66
"64": {
67
- "BLOCK_SIZE_M": 64,
67
+ "BLOCK_SIZE_M": 16,
68
"BLOCK_SIZE_N": 128,
69
"BLOCK_SIZE_K": 128,
70
- "GROUP_SIZE_M": 32,
70
+ "GROUP_SIZE_M": 64,
71
"num_warps": 4,
72
"num_stages": 3
73
},
74
"96": {
75
- "BLOCK_SIZE_M": 64,
75
+ "BLOCK_SIZE_M": 16,
76
"BLOCK_SIZE_N": 128,
77
"BLOCK_SIZE_K": 128,
78
- "GROUP_SIZE_M": 32,
78
+ "GROUP_SIZE_M": 16,
79
"num_warps": 4,
80
"num_stages": 3
81
},
82
"128": {
83
- "BLOCK_SIZE_M": 64,
83
+ "BLOCK_SIZE_M": 16,
84
"BLOCK_SIZE_N": 128,
85
"BLOCK_SIZE_K": 128,
86
- "GROUP_SIZE_M": 16,
86
+ "GROUP_SIZE_M": 32,
87
"num_warps": 4,
88
"num_stages": 3
89
},
90
"256": {
91
- "BLOCK_SIZE_M": 64,
91
+ "BLOCK_SIZE_M": 16,
92
"BLOCK_SIZE_N": 128,
93
"BLOCK_SIZE_K": 128,
94
"GROUP_SIZE_M": 32,
@@ -96,10 +96,10 @@
96
"num_stages": 3
97
},
98
"512": {
99
- "BLOCK_SIZE_M": 64,
99
+ "BLOCK_SIZE_M": 16,
100
"BLOCK_SIZE_N": 128,
101
"BLOCK_SIZE_K": 128,
102
- "GROUP_SIZE_M": 32,
102
+ "GROUP_SIZE_M": 16,
103
"num_warps": 4,
104
"num_stages": 3
105
},
@@ -109,7 +109,7 @@
109
"BLOCK_SIZE_K": 128,
110
"GROUP_SIZE_M": 32,
111
"num_warps": 4,
112
- "num_stages": 3
112
+ "num_stages": 4
113
},
114
"1536": {
115
"BLOCK_SIZE_M": 64,
@@ -117,21 +117,21 @@
117
"BLOCK_SIZE_K": 128,
118
"GROUP_SIZE_M": 32,
119
"num_warps": 4,
120
- "num_stages": 3
120
+ "num_stages": 4
121
},
122
"2048": {
123
"BLOCK_SIZE_M": 64,
124
"BLOCK_SIZE_N": 128,
125
"BLOCK_SIZE_K": 128,
126
- "GROUP_SIZE_M": 64,
126
+ "GROUP_SIZE_M": 32,
127
"num_warps": 4,
128
- "num_stages": 3
128
+ "num_stages": 4
129
},
130
"3072": {
131
- "BLOCK_SIZE_M": 128,
132
- "BLOCK_SIZE_N": 64,
131
+ "BLOCK_SIZE_M": 64,
132
+ "BLOCK_SIZE_N": 128,
133
"BLOCK_SIZE_K": 128,
134
- "GROUP_SIZE_M": 16,
134
+ "GROUP_SIZE_M": 32,
135
"num_warps": 4,
136
"num_stages": 4
137
},
@@ -139,7 +139,7 @@
139
"BLOCK_SIZE_M": 64,
140
"BLOCK_SIZE_N": 128,
141
"BLOCK_SIZE_K": 128,
142
- "GROUP_SIZE_M": 64,
142
+ "GROUP_SIZE_M": 16,
143
"num_warps": 4,
144
"num_stages": 3
145
}