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 17, 00:30 UTC
[Kernel] Re-tune Mixtral MoE configurations for FP8 on H100 (#5238)
Philipp Moritz committed
Jun 5, 2024 at 10:59 UTC
51a08e7d8f0f11411d380c007ab606fc2d5e3cf9
3 files changed
+85
-84
benchmarks/kernels/benchmark_moe.py
+2
-1
@@ -255,7 +255,8 @@ def main(args: argparse.Namespace):
255
256
if args.batch_size is None:
257
batch_sizes = [
258
- 1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024, 1536, 2048, 3072, 4096
258
+ 1, 2, 4, 8, 16, 24, 32, 48, 64, 96, 128, 256, 512, 1024, 1536,
259
+ 2048, 3072, 4096
260
]
261
else:
262
batch_sizes = [args.batch_size]
vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=float8.json
+52
-52
@@ -1,113 +1,113 @@
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": 8,
8
- "num_stages": 4
3
+ "BLOCK_SIZE_M": 64,
4
+ "BLOCK_SIZE_N": 128,
5
+ "BLOCK_SIZE_K": 128,
6
+ "GROUP_SIZE_M": 32,
7
+ "num_warps": 4,
8
+ "num_stages": 5
9
},
10
"2": {
11
- "BLOCK_SIZE_M": 16,
12
- "BLOCK_SIZE_N": 256,
13
- "BLOCK_SIZE_K": 64,
11
+ "BLOCK_SIZE_M": 64,
12
+ "BLOCK_SIZE_N": 128,
13
+ "BLOCK_SIZE_K": 128,
14
"GROUP_SIZE_M": 1,
15
- "num_warps": 8,
16
- "num_stages": 3
15
+ "num_warps": 4,
16
+ "num_stages": 4
17
},
18
"4": {
19
- "BLOCK_SIZE_M": 256,
20
- "BLOCK_SIZE_N": 64,
21
- "BLOCK_SIZE_K": 128,
22
- "GROUP_SIZE_M": 1,
19
+ "BLOCK_SIZE_M": 64,
20
+ "BLOCK_SIZE_N": 128,
21
+ "BLOCK_SIZE_K": 256,
22
+ "GROUP_SIZE_M": 16,
23
"num_warps": 4,
24
- "num_stages": 2
24
+ "num_stages": 4
25
},
26
"8": {
27
"BLOCK_SIZE_M": 64,
28
- "BLOCK_SIZE_N": 256,
29
- "BLOCK_SIZE_K": 64,
28
+ "BLOCK_SIZE_N": 128,
29
+ "BLOCK_SIZE_K": 128,
30
"GROUP_SIZE_M": 32,
31
- "num_warps": 8,
32
- "num_stages": 2
31
+ "num_warps": 4,
32
+ "num_stages": 4
33
},
34
"16": {
35
- "BLOCK_SIZE_M": 256,
36
- "BLOCK_SIZE_N": 128,
37
- "BLOCK_SIZE_K": 64,
35
+ "BLOCK_SIZE_M": 64,
36
+ "BLOCK_SIZE_N": 64,
37
+ "BLOCK_SIZE_K": 128,
38
"GROUP_SIZE_M": 1,
39
- "num_warps": 8,
40
- "num_stages": 5
39
+ "num_warps": 4,
40
+ "num_stages": 3
41
},
42
"24": {
43
"BLOCK_SIZE_M": 64,
44
- "BLOCK_SIZE_N": 256,
44
+ "BLOCK_SIZE_N": 64,
45
"BLOCK_SIZE_K": 128,
46
"GROUP_SIZE_M": 64,
47
"num_warps": 4,
48
- "num_stages": 4
48
+ "num_stages": 3
49
},
50
"32": {
51
"BLOCK_SIZE_M": 64,
52
- "BLOCK_SIZE_N": 128,
53
- "BLOCK_SIZE_K": 256,
54
- "GROUP_SIZE_M": 1,
52
+ "BLOCK_SIZE_N": 64,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 64,
55
"num_warps": 4,
56
- "num_stages": 4
56
+ "num_stages": 3
57
},
58
"48": {
59
"BLOCK_SIZE_M": 64,
60
- "BLOCK_SIZE_N": 256,
61
- "BLOCK_SIZE_K": 64,
62
- "GROUP_SIZE_M": 64,
60
+ "BLOCK_SIZE_N": 64,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 16,
63
"num_warps": 4,
64
- "num_stages": 4
64
+ "num_stages": 3
65
},
66
"64": {
67
"BLOCK_SIZE_M": 64,
68
- "BLOCK_SIZE_N": 128,
69
- "BLOCK_SIZE_K": 256,
68
+ "BLOCK_SIZE_N": 64,
69
+ "BLOCK_SIZE_K": 128,
70
"GROUP_SIZE_M": 1,
71
"num_warps": 4,
72
- "num_stages": 2
72
+ "num_stages": 3
73
},
74
"96": {
75
- "BLOCK_SIZE_M": 128,
76
- "BLOCK_SIZE_N": 32,
75
+ "BLOCK_SIZE_M": 64,
76
+ "BLOCK_SIZE_N": 64,
77
"BLOCK_SIZE_K": 128,
78
"GROUP_SIZE_M": 1,
79
"num_warps": 4,
80
- "num_stages": 5
80
+ "num_stages": 3
81
},
82
"128": {
83
"BLOCK_SIZE_M": 64,
84
"BLOCK_SIZE_N": 64,
85
- "BLOCK_SIZE_K": 256,
86
- "GROUP_SIZE_M": 32,
87
- "num_warps": 8,
88
- "num_stages": 5
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": 128,
93
- "BLOCK_SIZE_K": 256,
94
- "GROUP_SIZE_M": 1,
95
- "num_warps": 8,
93
+ "BLOCK_SIZE_K": 128,
94
+ "GROUP_SIZE_M": 16,
95
+ "num_warps": 4,
96
"num_stages": 3
97
},
98
"512": {
99
"BLOCK_SIZE_M": 128,
100
- "BLOCK_SIZE_N": 64,
100
+ "BLOCK_SIZE_N": 256,
101
"BLOCK_SIZE_K": 128,
102
"GROUP_SIZE_M": 16,
103
"num_warps": 8,
104
- "num_stages": 3
104
+ "num_stages": 4
105
},
106
"1024": {
107
"BLOCK_SIZE_M": 128,
108
"BLOCK_SIZE_N": 256,
109
"BLOCK_SIZE_K": 128,
110
- "GROUP_SIZE_M": 64,
110
+ "GROUP_SIZE_M": 32,
111
"num_warps": 8,
112
"num_stages": 4
113
},
@@ -115,7 +115,7 @@
115
"BLOCK_SIZE_M": 128,
116
"BLOCK_SIZE_N": 256,
117
"BLOCK_SIZE_K": 128,
118
- "GROUP_SIZE_M": 64,
118
+ "GROUP_SIZE_M": 16,
119
"num_warps": 8,
120
"num_stages": 4
121
},
vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=float8.json
+31
-31
@@ -1,67 +1,67 @@
1
{
2
"1": {
3
- "BLOCK_SIZE_M": 16,
4
- "BLOCK_SIZE_N": 32,
5
- "BLOCK_SIZE_K": 64,
6
- "GROUP_SIZE_M": 1,
3
+ "BLOCK_SIZE_M": 64,
4
+ "BLOCK_SIZE_N": 128,
5
+ "BLOCK_SIZE_K": 256,
6
+ "GROUP_SIZE_M": 64,
7
"num_warps": 4,
8
- "num_stages": 5
8
+ "num_stages": 4
9
},
10
"2": {
11
"BLOCK_SIZE_M": 64,
12
"BLOCK_SIZE_N": 64,
13
- "BLOCK_SIZE_K": 128,
14
- "GROUP_SIZE_M": 32,
15
- "num_warps": 8,
16
- "num_stages": 4
13
+ "BLOCK_SIZE_K": 256,
14
+ "GROUP_SIZE_M": 1,
15
+ "num_warps": 4,
16
+ "num_stages": 5
17
},
18
"4": {
19
"BLOCK_SIZE_M": 64,
20
- "BLOCK_SIZE_N": 128,
20
+ "BLOCK_SIZE_N": 64,
21
"BLOCK_SIZE_K": 256,
22
"GROUP_SIZE_M": 1,
23
"num_warps": 4,
24
- "num_stages": 2
24
+ "num_stages": 5
25
},
26
"8": {
27
- "BLOCK_SIZE_M": 128,
28
- "BLOCK_SIZE_N": 128,
27
+ "BLOCK_SIZE_M": 64,
28
+ "BLOCK_SIZE_N": 256,
29
"BLOCK_SIZE_K": 128,
30
- "GROUP_SIZE_M": 1,
30
+ "GROUP_SIZE_M": 32,
31
"num_warps": 4,
32
- "num_stages": 3
32
+ "num_stages": 4
33
},
34
"16": {
35
"BLOCK_SIZE_M": 64,
36
"BLOCK_SIZE_N": 128,
37
- "BLOCK_SIZE_K": 256,
38
- "GROUP_SIZE_M": 1,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 32,
39
"num_warps": 4,
40
- "num_stages": 4
40
+ "num_stages": 5
41
},
42
"24": {
43
"BLOCK_SIZE_M": 64,
44
- "BLOCK_SIZE_N": 128,
45
- "BLOCK_SIZE_K": 128,
44
+ "BLOCK_SIZE_N": 64,
45
+ "BLOCK_SIZE_K": 256,
46
"GROUP_SIZE_M": 1,
47
"num_warps": 4,
48
- "num_stages": 5
48
+ "num_stages": 3
49
},
50
"32": {
51
"BLOCK_SIZE_M": 64,
52
"BLOCK_SIZE_N": 128,
53
"BLOCK_SIZE_K": 256,
54
- "GROUP_SIZE_M": 64,
54
+ "GROUP_SIZE_M": 1,
55
"num_warps": 4,
56
"num_stages": 4
57
},
58
"48": {
59
"BLOCK_SIZE_M": 64,
60
- "BLOCK_SIZE_N": 64,
60
+ "BLOCK_SIZE_N": 128,
61
"BLOCK_SIZE_K": 256,
62
"GROUP_SIZE_M": 1,
63
"num_warps": 4,
64
- "num_stages": 3
64
+ "num_stages": 4
65
},
66
"64": {
67
"BLOCK_SIZE_M": 64,
@@ -81,19 +81,19 @@
81
},
82
"128": {
83
"BLOCK_SIZE_M": 64,
84
- "BLOCK_SIZE_N": 64,
84
+ "BLOCK_SIZE_N": 128,
85
"BLOCK_SIZE_K": 256,
86
"GROUP_SIZE_M": 1,
87
"num_warps": 4,
88
- "num_stages": 3
88
+ "num_stages": 4
89
},
90
"256": {
91
- "BLOCK_SIZE_M": 128,
91
+ "BLOCK_SIZE_M": 64,
92
"BLOCK_SIZE_N": 128,
93
"BLOCK_SIZE_K": 128,
94
- "GROUP_SIZE_M": 1,
95
- "num_warps": 8,
96
- "num_stages": 5
94
+ "GROUP_SIZE_M": 64,
95
+ "num_warps": 4,
96
+ "num_stages": 3
97
},
98
"512": {
99
"BLOCK_SIZE_M": 128,
@@ -107,7 +107,7 @@
107
"BLOCK_SIZE_M": 128,
108
"BLOCK_SIZE_N": 256,
109
"BLOCK_SIZE_K": 128,
110
- "GROUP_SIZE_M": 64,
110
+ "GROUP_SIZE_M": 32,
111
"num_warps": 8,
112
"num_stages": 4
113
},