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 11, 06:15 UTC
[PERF] Qwen3-next. Add fp8 cutlass MoE tuned configs. `chmod -x *MI308X.json` (#29553)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Vadim Gimpelson committed
Dec 18, 2025 at 09:16 UTC
717ac33d9cfa82357ea57dc0f6ee2aa325eba9f2
7 files changed
+882
vllm/model_executor/layers/fused_moe/configs/E=128,N=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json
new
+147
@@ -0,0 +1,147 @@
1
+{
2
+ "triton_version": "3.5.0",
3
+ "1": {
4
+ "BLOCK_SIZE_M": 16,
5
+ "BLOCK_SIZE_N": 128,
6
+ "BLOCK_SIZE_K": 128,
7
+ "GROUP_SIZE_M": 1,
8
+ "num_warps": 4,
9
+ "num_stages": 3
10
+ },
11
+ "2": {
12
+ "BLOCK_SIZE_M": 16,
13
+ "BLOCK_SIZE_N": 128,
14
+ "BLOCK_SIZE_K": 128,
15
+ "GROUP_SIZE_M": 1,
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": 1,
24
+ "num_warps": 4,
25
+ "num_stages": 3
26
+ },
27
+ "8": {
28
+ "BLOCK_SIZE_M": 16,
29
+ "BLOCK_SIZE_N": 128,
30
+ "BLOCK_SIZE_K": 128,
31
+ "GROUP_SIZE_M": 1,
32
+ "num_warps": 4,
33
+ "num_stages": 3
34
+ },
35
+ "16": {
36
+ "BLOCK_SIZE_M": 16,
37
+ "BLOCK_SIZE_N": 128,
38
+ "BLOCK_SIZE_K": 256,
39
+ "GROUP_SIZE_M": 1,
40
+ "num_warps": 4,
41
+ "num_stages": 3
42
+ },
43
+ "24": {
44
+ "BLOCK_SIZE_M": 16,
45
+ "BLOCK_SIZE_N": 128,
46
+ "BLOCK_SIZE_K": 256,
47
+ "GROUP_SIZE_M": 1,
48
+ "num_warps": 4,
49
+ "num_stages": 3
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": 3
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": 3
66
+ },
67
+ "64": {
68
+ "BLOCK_SIZE_M": 16,
69
+ "BLOCK_SIZE_N": 128,
70
+ "BLOCK_SIZE_K": 256,
71
+ "GROUP_SIZE_M": 64,
72
+ "num_warps": 4,
73
+ "num_stages": 3
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": 3
82
+ },
83
+ "128": {
84
+ "BLOCK_SIZE_M": 16,
85
+ "BLOCK_SIZE_N": 128,
86
+ "BLOCK_SIZE_K": 256,
87
+ "GROUP_SIZE_M": 1,
88
+ "num_warps": 4,
89
+ "num_stages": 3
90
+ },
91
+ "256": {
92
+ "BLOCK_SIZE_M": 32,
93
+ "BLOCK_SIZE_N": 128,
94
+ "BLOCK_SIZE_K": 256,
95
+ "GROUP_SIZE_M": 1,
96
+ "num_warps": 4,
97
+ "num_stages": 3
98
+ },
99
+ "512": {
100
+ "BLOCK_SIZE_M": 64,
101
+ "BLOCK_SIZE_N": 128,
102
+ "BLOCK_SIZE_K": 128,
103
+ "GROUP_SIZE_M": 64,
104
+ "num_warps": 4,
105
+ "num_stages": 4
106
+ },
107
+ "1024": {
108
+ "BLOCK_SIZE_M": 64,
109
+ "BLOCK_SIZE_N": 128,
110
+ "BLOCK_SIZE_K": 256,
111
+ "GROUP_SIZE_M": 1,
112
+ "num_warps": 4,
113
+ "num_stages": 4
114
+ },
115
+ "1536": {
116
+ "BLOCK_SIZE_M": 64,
117
+ "BLOCK_SIZE_N": 128,
118
+ "BLOCK_SIZE_K": 256,
119
+ "GROUP_SIZE_M": 1,
120
+ "num_warps": 4,
121
+ "num_stages": 4
122
+ },
123
+ "2048": {
124
+ "BLOCK_SIZE_M": 64,
125
+ "BLOCK_SIZE_N": 128,
126
+ "BLOCK_SIZE_K": 128,
127
+ "GROUP_SIZE_M": 1,
128
+ "num_warps": 4,
129
+ "num_stages": 4
130
+ },
131
+ "3072": {
132
+ "BLOCK_SIZE_M": 64,
133
+ "BLOCK_SIZE_N": 128,
134
+ "BLOCK_SIZE_K": 128,
135
+ "GROUP_SIZE_M": 1,
136
+ "num_warps": 4,
137
+ "num_stages": 4
138
+ },
139
+ "4096": {
140
+ "BLOCK_SIZE_M": 64,
141
+ "BLOCK_SIZE_N": 128,
142
+ "BLOCK_SIZE_K": 256,
143
+ "GROUP_SIZE_M": 1,
144
+ "num_warps": 4,
145
+ "num_stages": 4
146
+ }
147
+}
vllm/model_executor/layers/fused_moe/configs/E=128,N=768,device_name=AMD_Instinct_MI308X.json
vllm/model_executor/layers/fused_moe/configs/E=256,N=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json
new
+147
@@ -0,0 +1,147 @@
1
+{
2
+ "triton_version": "3.5.0",
3
+ "1": {
4
+ "BLOCK_SIZE_M": 16,
5
+ "BLOCK_SIZE_N": 128,
6
+ "BLOCK_SIZE_K": 128,
7
+ "GROUP_SIZE_M": 1,
8
+ "num_warps": 4,
9
+ "num_stages": 4
10
+ },
11
+ "2": {
12
+ "BLOCK_SIZE_M": 16,
13
+ "BLOCK_SIZE_N": 128,
14
+ "BLOCK_SIZE_K": 128,
15
+ "GROUP_SIZE_M": 1,
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": 1,
24
+ "num_warps": 4,
25
+ "num_stages": 3
26
+ },
27
+ "8": {
28
+ "BLOCK_SIZE_M": 16,
29
+ "BLOCK_SIZE_N": 128,
30
+ "BLOCK_SIZE_K": 128,
31
+ "GROUP_SIZE_M": 1,
32
+ "num_warps": 4,
33
+ "num_stages": 3
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": 3
42
+ },
43
+ "24": {
44
+ "BLOCK_SIZE_M": 16,
45
+ "BLOCK_SIZE_N": 128,
46
+ "BLOCK_SIZE_K": 256,
47
+ "GROUP_SIZE_M": 32,
48
+ "num_warps": 4,
49
+ "num_stages": 3
50
+ },
51
+ "32": {
52
+ "BLOCK_SIZE_M": 16,
53
+ "BLOCK_SIZE_N": 128,
54
+ "BLOCK_SIZE_K": 128,
55
+ "GROUP_SIZE_M": 1,
56
+ "num_warps": 4,
57
+ "num_stages": 3
58
+ },
59
+ "48": {
60
+ "BLOCK_SIZE_M": 16,
61
+ "BLOCK_SIZE_N": 128,
62
+ "BLOCK_SIZE_K": 256,
63
+ "GROUP_SIZE_M": 32,
64
+ "num_warps": 4,
65
+ "num_stages": 3
66
+ },
67
+ "64": {
68
+ "BLOCK_SIZE_M": 16,
69
+ "BLOCK_SIZE_N": 128,
70
+ "BLOCK_SIZE_K": 256,
71
+ "GROUP_SIZE_M": 64,
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": 256,
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": 256,
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": 256,
95
+ "GROUP_SIZE_M": 1,
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": 3
106
+ },
107
+ "1024": {
108
+ "BLOCK_SIZE_M": 64,
109
+ "BLOCK_SIZE_N": 128,
110
+ "BLOCK_SIZE_K": 128,
111
+ "GROUP_SIZE_M": 16,
112
+ "num_warps": 4,
113
+ "num_stages": 4
114
+ },
115
+ "1536": {
116
+ "BLOCK_SIZE_M": 64,
117
+ "BLOCK_SIZE_N": 128,
118
+ "BLOCK_SIZE_K": 128,
119
+ "GROUP_SIZE_M": 16,
120
+ "num_warps": 4,
121
+ "num_stages": 4
122
+ },
123
+ "2048": {
124
+ "BLOCK_SIZE_M": 64,
125
+ "BLOCK_SIZE_N": 128,
126
+ "BLOCK_SIZE_K": 128,
127
+ "GROUP_SIZE_M": 1,
128
+ "num_warps": 4,
129
+ "num_stages": 4
130
+ },
131
+ "3072": {
132
+ "BLOCK_SIZE_M": 64,
133
+ "BLOCK_SIZE_N": 128,
134
+ "BLOCK_SIZE_K": 128,
135
+ "GROUP_SIZE_M": 16,
136
+ "num_warps": 4,
137
+ "num_stages": 4
138
+ },
139
+ "4096": {
140
+ "BLOCK_SIZE_M": 64,
141
+ "BLOCK_SIZE_N": 128,
142
+ "BLOCK_SIZE_K": 256,
143
+ "GROUP_SIZE_M": 1,
144
+ "num_warps": 4,
145
+ "num_stages": 4
146
+ }
147
+}
vllm/model_executor/layers/fused_moe/configs/E=512,N=128,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json
new
+147
@@ -0,0 +1,147 @@
1
+{
2
+ "triton_version": "3.5.0",
3
+ "1": {
4
+ "BLOCK_SIZE_M": 16,
5
+ "BLOCK_SIZE_N": 128,
6
+ "BLOCK_SIZE_K": 128,
7
+ "GROUP_SIZE_M": 1,
8
+ "num_warps": 4,
9
+ "num_stages": 3
10
+ },
11
+ "2": {
12
+ "BLOCK_SIZE_M": 16,
13
+ "BLOCK_SIZE_N": 128,
14
+ "BLOCK_SIZE_K": 128,
15
+ "GROUP_SIZE_M": 1,
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": 1,
24
+ "num_warps": 4,
25
+ "num_stages": 3
26
+ },
27
+ "8": {
28
+ "BLOCK_SIZE_M": 16,
29
+ "BLOCK_SIZE_N": 128,
30
+ "BLOCK_SIZE_K": 128,
31
+ "GROUP_SIZE_M": 1,
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": 256,
39
+ "GROUP_SIZE_M": 16,
40
+ "num_warps": 4,
41
+ "num_stages": 3
42
+ },
43
+ "24": {
44
+ "BLOCK_SIZE_M": 16,
45
+ "BLOCK_SIZE_N": 128,
46
+ "BLOCK_SIZE_K": 128,
47
+ "GROUP_SIZE_M": 16,
48
+ "num_warps": 4,
49
+ "num_stages": 3
50
+ },
51
+ "32": {
52
+ "BLOCK_SIZE_M": 16,
53
+ "BLOCK_SIZE_N": 128,
54
+ "BLOCK_SIZE_K": 128,
55
+ "GROUP_SIZE_M": 16,
56
+ "num_warps": 4,
57
+ "num_stages": 3
58
+ },
59
+ "48": {
60
+ "BLOCK_SIZE_M": 16,
61
+ "BLOCK_SIZE_N": 256,
62
+ "BLOCK_SIZE_K": 128,
63
+ "GROUP_SIZE_M": 1,
64
+ "num_warps": 4,
65
+ "num_stages": 2
66
+ },
67
+ "64": {
68
+ "BLOCK_SIZE_M": 16,
69
+ "BLOCK_SIZE_N": 256,
70
+ "BLOCK_SIZE_K": 128,
71
+ "GROUP_SIZE_M": 1,
72
+ "num_warps": 4,
73
+ "num_stages": 2
74
+ },
75
+ "96": {
76
+ "BLOCK_SIZE_M": 16,
77
+ "BLOCK_SIZE_N": 256,
78
+ "BLOCK_SIZE_K": 128,
79
+ "GROUP_SIZE_M": 16,
80
+ "num_warps": 4,
81
+ "num_stages": 2
82
+ },
83
+ "128": {
84
+ "BLOCK_SIZE_M": 16,
85
+ "BLOCK_SIZE_N": 128,
86
+ "BLOCK_SIZE_K": 256,
87
+ "GROUP_SIZE_M": 16,
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": 256,
95
+ "GROUP_SIZE_M": 64,
96
+ "num_warps": 4,
97
+ "num_stages": 3
98
+ },
99
+ "512": {
100
+ "BLOCK_SIZE_M": 16,
101
+ "BLOCK_SIZE_N": 128,
102
+ "BLOCK_SIZE_K": 256,
103
+ "GROUP_SIZE_M": 64,
104
+ "num_warps": 4,
105
+ "num_stages": 3
106
+ },
107
+ "1024": {
108
+ "BLOCK_SIZE_M": 32,
109
+ "BLOCK_SIZE_N": 128,
110
+ "BLOCK_SIZE_K": 128,
111
+ "GROUP_SIZE_M": 1,
112
+ "num_warps": 4,
113
+ "num_stages": 3
114
+ },
115
+ "1536": {
116
+ "BLOCK_SIZE_M": 32,
117
+ "BLOCK_SIZE_N": 128,
118
+ "BLOCK_SIZE_K": 256,
119
+ "GROUP_SIZE_M": 64,
120
+ "num_warps": 4,
121
+ "num_stages": 3
122
+ },
123
+ "2048": {
124
+ "BLOCK_SIZE_M": 64,
125
+ "BLOCK_SIZE_N": 128,
126
+ "BLOCK_SIZE_K": 128,
127
+ "GROUP_SIZE_M": 64,
128
+ "num_warps": 4,
129
+ "num_stages": 4
130
+ },
131
+ "3072": {
132
+ "BLOCK_SIZE_M": 64,
133
+ "BLOCK_SIZE_N": 128,
134
+ "BLOCK_SIZE_K": 128,
135
+ "GROUP_SIZE_M": 32,
136
+ "num_warps": 4,
137
+ "num_stages": 4
138
+ },
139
+ "4096": {
140
+ "BLOCK_SIZE_M": 64,
141
+ "BLOCK_SIZE_N": 128,
142
+ "BLOCK_SIZE_K": 128,
143
+ "GROUP_SIZE_M": 64,
144
+ "num_warps": 4,
145
+ "num_stages": 3
146
+ }
147
+}
vllm/model_executor/layers/fused_moe/configs/E=512,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json
new
+147
@@ -0,0 +1,147 @@
1
+{
2
+ "triton_version": "3.5.0",
3
+ "1": {
4
+ "BLOCK_SIZE_M": 16,
5
+ "BLOCK_SIZE_N": 128,
6
+ "BLOCK_SIZE_K": 128,
7
+ "GROUP_SIZE_M": 1,
8
+ "num_warps": 4,
9
+ "num_stages": 3
10
+ },
11
+ "2": {
12
+ "BLOCK_SIZE_M": 16,
13
+ "BLOCK_SIZE_N": 128,
14
+ "BLOCK_SIZE_K": 128,
15
+ "GROUP_SIZE_M": 1,
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": 1,
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": 1,
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": 32,
40
+ "num_warps": 4,
41
+ "num_stages": 3
42
+ },
43
+ "24": {
44
+ "BLOCK_SIZE_M": 16,
45
+ "BLOCK_SIZE_N": 256,
46
+ "BLOCK_SIZE_K": 256,
47
+ "GROUP_SIZE_M": 16,
48
+ "num_warps": 4,
49
+ "num_stages": 2
50
+ },
51
+ "32": {
52
+ "BLOCK_SIZE_M": 16,
53
+ "BLOCK_SIZE_N": 128,
54
+ "BLOCK_SIZE_K": 256,
55
+ "GROUP_SIZE_M": 32,
56
+ "num_warps": 4,
57
+ "num_stages": 3
58
+ },
59
+ "48": {
60
+ "BLOCK_SIZE_M": 16,
61
+ "BLOCK_SIZE_N": 128,
62
+ "BLOCK_SIZE_K": 256,
63
+ "GROUP_SIZE_M": 32,
64
+ "num_warps": 4,
65
+ "num_stages": 3
66
+ },
67
+ "64": {
68
+ "BLOCK_SIZE_M": 16,
69
+ "BLOCK_SIZE_N": 128,
70
+ "BLOCK_SIZE_K": 128,
71
+ "GROUP_SIZE_M": 64,
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": 32,
80
+ "num_warps": 4,
81
+ "num_stages": 3
82
+ },
83
+ "128": {
84
+ "BLOCK_SIZE_M": 16,
85
+ "BLOCK_SIZE_N": 128,
86
+ "BLOCK_SIZE_K": 128,
87
+ "GROUP_SIZE_M": 16,
88
+ "num_warps": 4,
89
+ "num_stages": 4
90
+ },
91
+ "256": {
92
+ "BLOCK_SIZE_M": 16,
93
+ "BLOCK_SIZE_N": 128,
94
+ "BLOCK_SIZE_K": 256,
95
+ "GROUP_SIZE_M": 64,
96
+ "num_warps": 4,
97
+ "num_stages": 3
98
+ },
99
+ "512": {
100
+ "BLOCK_SIZE_M": 16,
101
+ "BLOCK_SIZE_N": 128,
102
+ "BLOCK_SIZE_K": 128,
103
+ "GROUP_SIZE_M": 16,
104
+ "num_warps": 4,
105
+ "num_stages": 4
106
+ },
107
+ "1024": {
108
+ "BLOCK_SIZE_M": 32,
109
+ "BLOCK_SIZE_N": 128,
110
+ "BLOCK_SIZE_K": 256,
111
+ "GROUP_SIZE_M": 16,
112
+ "num_warps": 4,
113
+ "num_stages": 3
114
+ },
115
+ "1536": {
116
+ "BLOCK_SIZE_M": 32,
117
+ "BLOCK_SIZE_N": 128,
118
+ "BLOCK_SIZE_K": 128,
119
+ "GROUP_SIZE_M": 64,
120
+ "num_warps": 4,
121
+ "num_stages": 3
122
+ },
123
+ "2048": {
124
+ "BLOCK_SIZE_M": 64,
125
+ "BLOCK_SIZE_N": 128,
126
+ "BLOCK_SIZE_K": 256,
127
+ "GROUP_SIZE_M": 64,
128
+ "num_warps": 4,
129
+ "num_stages": 4
130
+ },
131
+ "3072": {
132
+ "BLOCK_SIZE_M": 64,
133
+ "BLOCK_SIZE_N": 128,
134
+ "BLOCK_SIZE_K": 256,
135
+ "GROUP_SIZE_M": 32,
136
+ "num_warps": 4,
137
+ "num_stages": 4
138
+ },
139
+ "4096": {
140
+ "BLOCK_SIZE_M": 64,
141
+ "BLOCK_SIZE_N": 128,
142
+ "BLOCK_SIZE_K": 256,
143
+ "GROUP_SIZE_M": 64,
144
+ "num_warps": 4,
145
+ "num_stages": 3
146
+ }
147
+}
vllm/model_executor/layers/fused_moe/configs/E=512,N=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json
new
+147
@@ -0,0 +1,147 @@
1
+{
2
+ "triton_version": "3.5.0",
3
+ "1": {
4
+ "BLOCK_SIZE_M": 16,
5
+ "BLOCK_SIZE_N": 128,
6
+ "BLOCK_SIZE_K": 128,
7
+ "GROUP_SIZE_M": 1,
8
+ "num_warps": 4,
9
+ "num_stages": 4
10
+ },
11
+ "2": {
12
+ "BLOCK_SIZE_M": 16,
13
+ "BLOCK_SIZE_N": 128,
14
+ "BLOCK_SIZE_K": 128,
15
+ "GROUP_SIZE_M": 1,
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": 1,
24
+ "num_warps": 4,
25
+ "num_stages": 3
26
+ },
27
+ "8": {
28
+ "BLOCK_SIZE_M": 16,
29
+ "BLOCK_SIZE_N": 128,
30
+ "BLOCK_SIZE_K": 128,
31
+ "GROUP_SIZE_M": 1,
32
+ "num_warps": 4,
33
+ "num_stages": 3
34
+ },
35
+ "16": {
36
+ "BLOCK_SIZE_M": 16,
37
+ "BLOCK_SIZE_N": 128,
38
+ "BLOCK_SIZE_K": 256,
39
+ "GROUP_SIZE_M": 1,
40
+ "num_warps": 4,
41
+ "num_stages": 3
42
+ },
43
+ "24": {
44
+ "BLOCK_SIZE_M": 16,
45
+ "BLOCK_SIZE_N": 128,
46
+ "BLOCK_SIZE_K": 256,
47
+ "GROUP_SIZE_M": 1,
48
+ "num_warps": 4,
49
+ "num_stages": 3
50
+ },
51
+ "32": {
52
+ "BLOCK_SIZE_M": 16,
53
+ "BLOCK_SIZE_N": 128,
54
+ "BLOCK_SIZE_K": 256,
55
+ "GROUP_SIZE_M": 1,
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": 64,
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": 3
82
+ },
83
+ "128": {
84
+ "BLOCK_SIZE_M": 16,
85
+ "BLOCK_SIZE_N": 128,
86
+ "BLOCK_SIZE_K": 128,
87
+ "GROUP_SIZE_M": 16,
88
+ "num_warps": 4,
89
+ "num_stages": 4
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": 16,
101
+ "BLOCK_SIZE_N": 128,
102
+ "BLOCK_SIZE_K": 128,
103
+ "GROUP_SIZE_M": 64,
104
+ "num_warps": 4,
105
+ "num_stages": 3
106
+ },
107
+ "1024": {
108
+ "BLOCK_SIZE_M": 32,
109
+ "BLOCK_SIZE_N": 128,
110
+ "BLOCK_SIZE_K": 128,
111
+ "GROUP_SIZE_M": 16,
112
+ "num_warps": 4,
113
+ "num_stages": 3
114
+ },
115
+ "1536": {
116
+ "BLOCK_SIZE_M": 64,
117
+ "BLOCK_SIZE_N": 128,
118
+ "BLOCK_SIZE_K": 128,
119
+ "GROUP_SIZE_M": 64,
120
+ "num_warps": 4,
121
+ "num_stages": 4
122
+ },
123
+ "2048": {
124
+ "BLOCK_SIZE_M": 64,
125
+ "BLOCK_SIZE_N": 128,
126
+ "BLOCK_SIZE_K": 128,
127
+ "GROUP_SIZE_M": 64,
128
+ "num_warps": 4,
129
+ "num_stages": 4
130
+ },
131
+ "3072": {
132
+ "BLOCK_SIZE_M": 64,
133
+ "BLOCK_SIZE_N": 128,
134
+ "BLOCK_SIZE_K": 128,
135
+ "GROUP_SIZE_M": 64,
136
+ "num_warps": 4,
137
+ "num_stages": 4
138
+ },
139
+ "4096": {
140
+ "BLOCK_SIZE_M": 64,
141
+ "BLOCK_SIZE_N": 128,
142
+ "BLOCK_SIZE_K": 256,
143
+ "GROUP_SIZE_M": 64,
144
+ "num_warps": 4,
145
+ "num_stages": 4
146
+ }
147
+}
vllm/model_executor/layers/fused_moe/configs/E=64,N=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json
new
+147
@@ -0,0 +1,147 @@
1
+{
2
+ "triton_version": "3.5.0",
3
+ "1": {
4
+ "BLOCK_SIZE_M": 16,
5
+ "BLOCK_SIZE_N": 128,
6
+ "BLOCK_SIZE_K": 128,
7
+ "GROUP_SIZE_M": 1,
8
+ "num_warps": 4,
9
+ "num_stages": 3
10
+ },
11
+ "2": {
12
+ "BLOCK_SIZE_M": 16,
13
+ "BLOCK_SIZE_N": 128,
14
+ "BLOCK_SIZE_K": 128,
15
+ "GROUP_SIZE_M": 1,
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": 1,
24
+ "num_warps": 4,
25
+ "num_stages": 3
26
+ },
27
+ "8": {
28
+ "BLOCK_SIZE_M": 16,
29
+ "BLOCK_SIZE_N": 128,
30
+ "BLOCK_SIZE_K": 128,
31
+ "GROUP_SIZE_M": 1,
32
+ "num_warps": 4,
33
+ "num_stages": 3
34
+ },
35
+ "16": {
36
+ "BLOCK_SIZE_M": 16,
37
+ "BLOCK_SIZE_N": 128,
38
+ "BLOCK_SIZE_K": 256,
39
+ "GROUP_SIZE_M": 1,
40
+ "num_warps": 4,
41
+ "num_stages": 3
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": 3
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": 3
58
+ },
59
+ "48": {
60
+ "BLOCK_SIZE_M": 16,
61
+ "BLOCK_SIZE_N": 128,
62
+ "BLOCK_SIZE_K": 256,
63
+ "GROUP_SIZE_M": 16,
64
+ "num_warps": 4,
65
+ "num_stages": 3
66
+ },
67
+ "64": {
68
+ "BLOCK_SIZE_M": 16,
69
+ "BLOCK_SIZE_N": 128,
70
+ "BLOCK_SIZE_K": 128,
71
+ "GROUP_SIZE_M": 16,
72
+ "num_warps": 4,
73
+ "num_stages": 3
74
+ },
75
+ "96": {
76
+ "BLOCK_SIZE_M": 16,
77
+ "BLOCK_SIZE_N": 128,
78
+ "BLOCK_SIZE_K": 256,
79
+ "GROUP_SIZE_M": 32,
80
+ "num_warps": 4,
81
+ "num_stages": 3
82
+ },
83
+ "128": {
84
+ "BLOCK_SIZE_M": 32,
85
+ "BLOCK_SIZE_N": 128,
86
+ "BLOCK_SIZE_K": 128,
87
+ "GROUP_SIZE_M": 16,
88
+ "num_warps": 4,
89
+ "num_stages": 3
90
+ },
91
+ "256": {
92
+ "BLOCK_SIZE_M": 64,
93
+ "BLOCK_SIZE_N": 128,
94
+ "BLOCK_SIZE_K": 128,
95
+ "GROUP_SIZE_M": 1,
96
+ "num_warps": 4,
97
+ "num_stages": 3
98
+ },
99
+ "512": {
100
+ "BLOCK_SIZE_M": 64,
101
+ "BLOCK_SIZE_N": 128,
102
+ "BLOCK_SIZE_K": 128,
103
+ "GROUP_SIZE_M": 1,
104
+ "num_warps": 4,
105
+ "num_stages": 4
106
+ },
107
+ "1024": {
108
+ "BLOCK_SIZE_M": 64,
109
+ "BLOCK_SIZE_N": 128,
110
+ "BLOCK_SIZE_K": 128,
111
+ "GROUP_SIZE_M": 1,
112
+ "num_warps": 4,
113
+ "num_stages": 4
114
+ },
115
+ "1536": {
116
+ "BLOCK_SIZE_M": 64,
117
+ "BLOCK_SIZE_N": 128,
118
+ "BLOCK_SIZE_K": 128,
119
+ "GROUP_SIZE_M": 1,
120
+ "num_warps": 4,
121
+ "num_stages": 4
122
+ },
123
+ "2048": {
124
+ "BLOCK_SIZE_M": 64,
125
+ "BLOCK_SIZE_N": 128,
126
+ "BLOCK_SIZE_K": 256,
127
+ "GROUP_SIZE_M": 1,
128
+ "num_warps": 4,
129
+ "num_stages": 4
130
+ },
131
+ "3072": {
132
+ "BLOCK_SIZE_M": 64,
133
+ "BLOCK_SIZE_N": 128,
134
+ "BLOCK_SIZE_K": 256,
135
+ "GROUP_SIZE_M": 1,
136
+ "num_warps": 4,
137
+ "num_stages": 3
138
+ },
139
+ "4096": {
140
+ "BLOCK_SIZE_M": 64,
141
+ "BLOCK_SIZE_N": 128,
142
+ "BLOCK_SIZE_K": 256,
143
+ "GROUP_SIZE_M": 16,
144
+ "num_warps": 4,
145
+ "num_stages": 4
146
+ }
147
+}