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, 23:15 UTC
[Kernel] Tune Qwen2MoE kernel configurations with tp2,4 (#5497)
Tune Qwen2-57B-A14B configs based on #4921 Throughput Performance command: python benchmarks/benchmark_throughput.py --model=Qwen/Qwen2-57B-A14B-Instruct --input-len 1000 --output-len 50 -tp 2 A100 GPU benchmark no config w/ PR tp=2 10.53 requests/s, 11058.17 tokens/s 12.47 requests/s, 13088.57 tokens/s tp=4 17.77 requests/s, 18662.95 tokens/s 20.20 requests/s, 21212.32 tokens/s
wenyujin333 committed
Jun 14, 2024 at 00:01 UTC
bd43973522ea17be50e10fbb222a22f673c8067e
4 files changed
+584
vllm/model_executor/layers/fused_moe/configs/E=64,N=1280,device_name=NVIDIA_A100-SXM4-80GB.json
new
+146
@@ -0,0 +1,146 @@
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_stages": 5
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 128,
13
+ "BLOCK_SIZE_K": 64,
14
+ "GROUP_SIZE_M": 64,
15
+ "num_warps": 8,
16
+ "num_stages": 5
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 64,
21
+ "BLOCK_SIZE_K": 64,
22
+ "GROUP_SIZE_M": 1,
23
+ "num_warps": 4,
24
+ "num_stages": 5
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 128,
29
+ "BLOCK_SIZE_K": 128,
30
+ "GROUP_SIZE_M": 1,
31
+ "num_warps": 8,
32
+ "num_stages": 5
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 256,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 32,
39
+ "num_warps": 4,
40
+ "num_stages": 3
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 128,
45
+ "BLOCK_SIZE_K": 256,
46
+ "GROUP_SIZE_M": 1,
47
+ "num_warps": 4,
48
+ "num_stages": 2
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 16,
52
+ "BLOCK_SIZE_N": 64,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 16,
55
+ "num_warps": 4,
56
+ "num_stages": 3
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 16,
60
+ "BLOCK_SIZE_N": 128,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 1,
63
+ "num_warps": 4,
64
+ "num_stages": 2
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 16,
68
+ "BLOCK_SIZE_N": 128,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 3
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 32,
76
+ "BLOCK_SIZE_N": 128,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 1,
79
+ "num_warps": 4,
80
+ "num_stages": 3
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 32,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 1,
87
+ "num_warps": 4,
88
+ "num_stages": 3
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 64,
92
+ "BLOCK_SIZE_N": 256,
93
+ "BLOCK_SIZE_K": 128,
94
+ "GROUP_SIZE_M": 1,
95
+ "num_warps": 8,
96
+ "num_stages": 3
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 64,
100
+ "BLOCK_SIZE_N": 256,
101
+ "BLOCK_SIZE_K": 128,
102
+ "GROUP_SIZE_M": 1,
103
+ "num_warps": 8,
104
+ "num_stages": 3
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 64,
108
+ "BLOCK_SIZE_N": 256,
109
+ "BLOCK_SIZE_K": 128,
110
+ "GROUP_SIZE_M": 16,
111
+ "num_warps": 8,
112
+ "num_stages": 3
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 64,
116
+ "BLOCK_SIZE_N": 256,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 1,
119
+ "num_warps": 4,
120
+ "num_stages": 3
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 64,
124
+ "BLOCK_SIZE_N": 256,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 1,
127
+ "num_warps": 4,
128
+ "num_stages": 3
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 128,
132
+ "BLOCK_SIZE_N": 128,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 1,
135
+ "num_warps": 4,
136
+ "num_stages": 3
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 128,
140
+ "BLOCK_SIZE_N": 128,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 1,
143
+ "num_warps": 4,
144
+ "num_stages": 3
145
+ }
146
+}
vllm/model_executor/layers/fused_moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3.json
new
+146
@@ -0,0 +1,146 @@
1
+{
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 256,
5
+ "BLOCK_SIZE_K": 128,
6
+ "GROUP_SIZE_M": 16,
7
+ "num_warps": 4,
8
+ "num_stages": 3
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 64,
13
+ "BLOCK_SIZE_K": 128,
14
+ "GROUP_SIZE_M": 16,
15
+ "num_warps": 4,
16
+ "num_stages": 4
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 64,
21
+ "BLOCK_SIZE_K": 256,
22
+ "GROUP_SIZE_M": 16,
23
+ "num_warps": 4,
24
+ "num_stages": 3
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 32,
29
+ "BLOCK_SIZE_K": 256,
30
+ "GROUP_SIZE_M": 1,
31
+ "num_warps": 4,
32
+ "num_stages": 2
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 32,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 16,
39
+ "num_warps": 4,
40
+ "num_stages": 5
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 128,
45
+ "BLOCK_SIZE_K": 256,
46
+ "GROUP_SIZE_M": 32,
47
+ "num_warps": 4,
48
+ "num_stages": 2
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 16,
52
+ "BLOCK_SIZE_N": 256,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 1,
55
+ "num_warps": 4,
56
+ "num_stages": 3
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 16,
60
+ "BLOCK_SIZE_N": 256,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 1,
63
+ "num_warps": 4,
64
+ "num_stages": 3
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 16,
68
+ "BLOCK_SIZE_N": 256,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 3
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 32,
76
+ "BLOCK_SIZE_N": 256,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 1,
79
+ "num_warps": 4,
80
+ "num_stages": 3
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 32,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 1,
87
+ "num_warps": 4,
88
+ "num_stages": 3
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 64,
92
+ "BLOCK_SIZE_N": 64,
93
+ "BLOCK_SIZE_K": 64,
94
+ "GROUP_SIZE_M": 1,
95
+ "num_warps": 4,
96
+ "num_stages": 3
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 128,
100
+ "BLOCK_SIZE_N": 128,
101
+ "BLOCK_SIZE_K": 64,
102
+ "GROUP_SIZE_M": 1,
103
+ "num_warps": 8,
104
+ "num_stages": 3
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 128,
108
+ "BLOCK_SIZE_N": 256,
109
+ "BLOCK_SIZE_K": 64,
110
+ "GROUP_SIZE_M": 1,
111
+ "num_warps": 8,
112
+ "num_stages": 4
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 128,
116
+ "BLOCK_SIZE_N": 256,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 1,
119
+ "num_warps": 8,
120
+ "num_stages": 4
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 128,
124
+ "BLOCK_SIZE_N": 256,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 1,
127
+ "num_warps": 8,
128
+ "num_stages": 4
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 128,
132
+ "BLOCK_SIZE_N": 256,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 32,
135
+ "num_warps": 8,
136
+ "num_stages": 4
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 128,
140
+ "BLOCK_SIZE_N": 256,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 1,
143
+ "num_warps": 8,
144
+ "num_stages": 4
145
+ }
146
+}
vllm/model_executor/layers/fused_moe/configs/E=64,N=640,device_name=NVIDIA_A100-SXM4-80GB.json
new
+146
@@ -0,0 +1,146 @@
1
+{
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 128,
5
+ "BLOCK_SIZE_K": 128,
6
+ "GROUP_SIZE_M": 64,
7
+ "num_warps": 4,
8
+ "num_stages": 4
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 32,
12
+ "BLOCK_SIZE_N": 128,
13
+ "BLOCK_SIZE_K": 64,
14
+ "GROUP_SIZE_M": 1,
15
+ "num_warps": 4,
16
+ "num_stages": 5
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 32,
21
+ "BLOCK_SIZE_K": 64,
22
+ "GROUP_SIZE_M": 16,
23
+ "num_warps": 4,
24
+ "num_stages": 4
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 64,
29
+ "BLOCK_SIZE_K": 256,
30
+ "GROUP_SIZE_M": 16,
31
+ "num_warps": 4,
32
+ "num_stages": 5
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 64,
37
+ "BLOCK_SIZE_K": 256,
38
+ "GROUP_SIZE_M": 1,
39
+ "num_warps": 4,
40
+ "num_stages": 2
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 64,
45
+ "BLOCK_SIZE_K": 256,
46
+ "GROUP_SIZE_M": 16,
47
+ "num_warps": 4,
48
+ "num_stages": 2
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 16,
52
+ "BLOCK_SIZE_N": 128,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 16,
55
+ "num_warps": 4,
56
+ "num_stages": 3
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 16,
60
+ "BLOCK_SIZE_N": 128,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 1,
63
+ "num_warps": 4,
64
+ "num_stages": 3
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 16,
68
+ "BLOCK_SIZE_N": 128,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 3
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 32,
76
+ "BLOCK_SIZE_N": 128,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 1,
79
+ "num_warps": 4,
80
+ "num_stages": 3
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 32,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 1,
87
+ "num_warps": 4,
88
+ "num_stages": 3
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 64,
92
+ "BLOCK_SIZE_N": 256,
93
+ "BLOCK_SIZE_K": 128,
94
+ "GROUP_SIZE_M": 1,
95
+ "num_warps": 8,
96
+ "num_stages": 3
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 64,
100
+ "BLOCK_SIZE_N": 256,
101
+ "BLOCK_SIZE_K": 128,
102
+ "GROUP_SIZE_M": 1,
103
+ "num_warps": 8,
104
+ "num_stages": 3
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 64,
108
+ "BLOCK_SIZE_N": 128,
109
+ "BLOCK_SIZE_K": 64,
110
+ "GROUP_SIZE_M": 1,
111
+ "num_warps": 4,
112
+ "num_stages": 3
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 64,
116
+ "BLOCK_SIZE_N": 256,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 1,
119
+ "num_warps": 4,
120
+ "num_stages": 3
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 128,
124
+ "BLOCK_SIZE_N": 128,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 1,
127
+ "num_warps": 4,
128
+ "num_stages": 3
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 128,
132
+ "BLOCK_SIZE_N": 128,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 1,
135
+ "num_warps": 4,
136
+ "num_stages": 3
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 128,
140
+ "BLOCK_SIZE_N": 128,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 1,
143
+ "num_warps": 4,
144
+ "num_stages": 3
145
+ }
146
+}
vllm/model_executor/layers/fused_moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3.json
new
+146
@@ -0,0 +1,146 @@
1
+{
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 128,
5
+ "BLOCK_SIZE_K": 128,
6
+ "GROUP_SIZE_M": 1,
7
+ "num_warps": 8,
8
+ "num_stages": 4
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 64,
13
+ "BLOCK_SIZE_K": 128,
14
+ "GROUP_SIZE_M": 16,
15
+ "num_warps": 4,
16
+ "num_stages": 4
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 64,
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": 32,
28
+ "BLOCK_SIZE_N": 256,
29
+ "BLOCK_SIZE_K": 128,
30
+ "GROUP_SIZE_M": 32,
31
+ "num_warps": 4,
32
+ "num_stages": 3
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 128,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 32,
39
+ "num_warps": 8,
40
+ "num_stages": 4
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 32,
45
+ "BLOCK_SIZE_K": 128,
46
+ "GROUP_SIZE_M": 64,
47
+ "num_warps": 4,
48
+ "num_stages": 3
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 16,
52
+ "BLOCK_SIZE_N": 64,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 16,
55
+ "num_warps": 4,
56
+ "num_stages": 2
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 16,
60
+ "BLOCK_SIZE_N": 128,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 1,
63
+ "num_warps": 4,
64
+ "num_stages": 2
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 16,
68
+ "BLOCK_SIZE_N": 64,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 3
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 32,
76
+ "BLOCK_SIZE_N": 64,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 1,
79
+ "num_warps": 4,
80
+ "num_stages": 2
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 32,
84
+ "BLOCK_SIZE_N": 64,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 1,
87
+ "num_warps": 4,
88
+ "num_stages": 2
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 64,
92
+ "BLOCK_SIZE_N": 128,
93
+ "BLOCK_SIZE_K": 64,
94
+ "GROUP_SIZE_M": 1,
95
+ "num_warps": 4,
96
+ "num_stages": 3
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 64,
100
+ "BLOCK_SIZE_N": 128,
101
+ "BLOCK_SIZE_K": 64,
102
+ "GROUP_SIZE_M": 1,
103
+ "num_warps": 4,
104
+ "num_stages": 3
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 128,
108
+ "BLOCK_SIZE_N": 256,
109
+ "BLOCK_SIZE_K": 64,
110
+ "GROUP_SIZE_M": 1,
111
+ "num_warps": 8,
112
+ "num_stages": 4
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 128,
116
+ "BLOCK_SIZE_N": 256,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 16,
119
+ "num_warps": 8,
120
+ "num_stages": 4
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 128,
124
+ "BLOCK_SIZE_N": 256,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 1,
127
+ "num_warps": 8,
128
+ "num_stages": 4
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 128,
132
+ "BLOCK_SIZE_N": 256,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 1,
135
+ "num_warps": 8,
136
+ "num_stages": 4
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 128,
140
+ "BLOCK_SIZE_N": 256,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 16,
143
+ "num_warps": 8,
144
+ "num_stages": 4
145
+ }
146
+}