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 16, 10:45 UTC
[Kernel] adding fused moe kernel config for L40S TP4 (#9245)
Burkhard Ringlein committed
Oct 11, 2024 at 08:54 UTC
f710090d8e40451879690b6a27b7d3b1a41b53ec
1 file changed
+173
vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=NVIDIA_L40S.json
new
+173
@@ -0,0 +1,173 @@
1
+{
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 32,
5
+ "BLOCK_SIZE_K": 64,
6
+ "GROUP_SIZE_M": 1,
7
+ "num_warps": 4,
8
+ "num_ctas": 1,
9
+ "num_stages": 2
10
+ },
11
+ "2": {
12
+ "BLOCK_SIZE_M": 32,
13
+ "BLOCK_SIZE_N": 64,
14
+ "BLOCK_SIZE_K": 32,
15
+ "GROUP_SIZE_M": 2,
16
+ "num_warps": 4,
17
+ "num_ctas": 1,
18
+ "num_stages": 7
19
+ },
20
+ "4": {
21
+ "BLOCK_SIZE_M": 32,
22
+ "BLOCK_SIZE_N": 128,
23
+ "BLOCK_SIZE_K": 32,
24
+ "GROUP_SIZE_M": 128,
25
+ "num_warps": 2,
26
+ "num_ctas": 1,
27
+ "num_stages": 4
28
+ },
29
+ "8": {
30
+ "BLOCK_SIZE_M": 16,
31
+ "BLOCK_SIZE_N": 32,
32
+ "BLOCK_SIZE_K": 256,
33
+ "GROUP_SIZE_M": 1,
34
+ "num_warps": 2,
35
+ "num_ctas": 1,
36
+ "num_stages": 1
37
+ },
38
+ "16": {
39
+ "BLOCK_SIZE_M": 16,
40
+ "BLOCK_SIZE_N": 32,
41
+ "BLOCK_SIZE_K": 256,
42
+ "GROUP_SIZE_M": 1,
43
+ "num_warps": 2,
44
+ "num_ctas": 1,
45
+ "num_stages": 1
46
+ },
47
+ "32": {
48
+ "BLOCK_SIZE_M": 16,
49
+ "BLOCK_SIZE_N": 256,
50
+ "BLOCK_SIZE_K": 128,
51
+ "GROUP_SIZE_M": 2,
52
+ "num_warps": 4,
53
+ "num_ctas": 1,
54
+ "num_stages": 2
55
+ },
56
+ "48": {
57
+ "BLOCK_SIZE_M": 16,
58
+ "BLOCK_SIZE_N": 256,
59
+ "BLOCK_SIZE_K": 128,
60
+ "GROUP_SIZE_M": 2,
61
+ "num_warps": 4,
62
+ "num_ctas": 1,
63
+ "num_stages": 2
64
+ },
65
+ "64": {
66
+ "BLOCK_SIZE_M": 16,
67
+ "BLOCK_SIZE_N": 256,
68
+ "BLOCK_SIZE_K": 128,
69
+ "GROUP_SIZE_M": 2,
70
+ "num_warps": 4,
71
+ "num_ctas": 1,
72
+ "num_stages": 2
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 32,
76
+ "BLOCK_SIZE_N": 256,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 8,
79
+ "num_warps": 8,
80
+ "num_ctas": 1,
81
+ "num_stages": 2
82
+ },
83
+ "128": {
84
+ "BLOCK_SIZE_M": 32,
85
+ "BLOCK_SIZE_N": 256,
86
+ "BLOCK_SIZE_K": 128,
87
+ "GROUP_SIZE_M": 8,
88
+ "num_warps": 8,
89
+ "num_ctas": 1,
90
+ "num_stages": 2
91
+ },
92
+ "192": {
93
+ "BLOCK_SIZE_M": 32,
94
+ "BLOCK_SIZE_N": 256,
95
+ "BLOCK_SIZE_K": 128,
96
+ "GROUP_SIZE_M": 8,
97
+ "num_warps": 8,
98
+ "num_ctas": 1,
99
+ "num_stages": 2
100
+ },
101
+ "256": {
102
+ "BLOCK_SIZE_M": 64,
103
+ "BLOCK_SIZE_N": 256,
104
+ "BLOCK_SIZE_K": 128,
105
+ "GROUP_SIZE_M": 1,
106
+ "num_warps": 16,
107
+ "num_ctas": 1,
108
+ "num_stages": 2
109
+ },
110
+ "512": {
111
+ "BLOCK_SIZE_M": 32,
112
+ "BLOCK_SIZE_N": 32,
113
+ "BLOCK_SIZE_K": 64,
114
+ "GROUP_SIZE_M": 128,
115
+ "num_warps": 2,
116
+ "num_ctas": 1,
117
+ "num_stages": 8
118
+ },
119
+ "1024": {
120
+ "BLOCK_SIZE_M": 64,
121
+ "BLOCK_SIZE_N": 128,
122
+ "BLOCK_SIZE_K": 128,
123
+ "GROUP_SIZE_M": 4,
124
+ "num_warps": 8,
125
+ "num_ctas": 1,
126
+ "num_stages": 3
127
+ },
128
+ "2048": {
129
+ "BLOCK_SIZE_M": 64,
130
+ "BLOCK_SIZE_N": 256,
131
+ "BLOCK_SIZE_K": 128,
132
+ "GROUP_SIZE_M": 1,
133
+ "num_warps": 16,
134
+ "num_ctas": 1,
135
+ "num_stages": 2
136
+ },
137
+ "3072": {
138
+ "BLOCK_SIZE_M": 64,
139
+ "BLOCK_SIZE_N": 256,
140
+ "BLOCK_SIZE_K": 128,
141
+ "GROUP_SIZE_M": 1,
142
+ "num_warps": 16,
143
+ "num_ctas": 1,
144
+ "num_stages": 2
145
+ },
146
+ "4096": {
147
+ "BLOCK_SIZE_M": 128,
148
+ "BLOCK_SIZE_N": 256,
149
+ "BLOCK_SIZE_K": 32,
150
+ "GROUP_SIZE_M": 64,
151
+ "num_warps": 8,
152
+ "num_ctas": 1,
153
+ "num_stages": 2
154
+ },
155
+ "6144": {
156
+ "BLOCK_SIZE_M": 128,
157
+ "BLOCK_SIZE_N": 256,
158
+ "BLOCK_SIZE_K": 32,
159
+ "GROUP_SIZE_M": 64,
160
+ "num_warps": 8,
161
+ "num_ctas": 1,
162
+ "num_stages": 2
163
+ },
164
+ "8192": {
165
+ "BLOCK_SIZE_M": 64,
166
+ "BLOCK_SIZE_N": 256,
167
+ "BLOCK_SIZE_K": 128,
168
+ "GROUP_SIZE_M": 1,
169
+ "num_warps": 16,
170
+ "num_ctas": 1,
171
+ "num_stages": 2
172
+ }
173
+}
\ No newline at end of file