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 6, 20:30 UTC
[ROCm] add tuned moe_wna16_triton kernel configs for CDNA4 (#35093)
Signed-off-by: salykova <amsalykov@gmail.com> Signed-off-by: amd-asalykov <asalykov@amd.com>
Amanzhol Salykov committed
Mar 11, 2026 at 20:00 UTC
8a24842765ba9b45b0116d65b16c2d5b1fcb7e05
4 files changed
+768
vllm/model_executor/layers/fused_moe/configs/E=384,N=256,device_name=AMD_Instinct_MI350X,dtype=int4_w4a16.json
new
+192
@@ -0,0 +1,192 @@
1
+{
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 128,
5
+ "BLOCK_SIZE_K": 128,
6
+ "GROUP_SIZE_M": 2,
7
+ "SPLIT_K": 1,
8
+ "num_warps": 4,
9
+ "num_stages": 2,
10
+ "matrix_instr_nonkdim": 16
11
+ },
12
+ "2": {
13
+ "BLOCK_SIZE_M": 16,
14
+ "BLOCK_SIZE_N": 128,
15
+ "BLOCK_SIZE_K": 128,
16
+ "GROUP_SIZE_M": 2,
17
+ "SPLIT_K": 1,
18
+ "num_warps": 4,
19
+ "num_stages": 2,
20
+ "matrix_instr_nonkdim": 16
21
+ },
22
+ "4": {
23
+ "BLOCK_SIZE_M": 16,
24
+ "BLOCK_SIZE_N": 128,
25
+ "BLOCK_SIZE_K": 128,
26
+ "GROUP_SIZE_M": 2,
27
+ "SPLIT_K": 1,
28
+ "num_warps": 4,
29
+ "num_stages": 2,
30
+ "matrix_instr_nonkdim": 16
31
+ },
32
+ "8": {
33
+ "BLOCK_SIZE_M": 16,
34
+ "BLOCK_SIZE_N": 128,
35
+ "BLOCK_SIZE_K": 128,
36
+ "GROUP_SIZE_M": 2,
37
+ "SPLIT_K": 1,
38
+ "num_warps": 4,
39
+ "num_stages": 2,
40
+ "matrix_instr_nonkdim": 16
41
+ },
42
+ "16": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 128,
45
+ "BLOCK_SIZE_K": 128,
46
+ "GROUP_SIZE_M": 1,
47
+ "SPLIT_K": 1,
48
+ "num_warps": 4,
49
+ "num_stages": 2,
50
+ "matrix_instr_nonkdim": 16
51
+ },
52
+ "24": {
53
+ "BLOCK_SIZE_M": 16,
54
+ "BLOCK_SIZE_N": 128,
55
+ "BLOCK_SIZE_K": 64,
56
+ "GROUP_SIZE_M": 1,
57
+ "SPLIT_K": 1,
58
+ "num_warps": 4,
59
+ "num_stages": 2,
60
+ "matrix_instr_nonkdim": 16
61
+ },
62
+ "32": {
63
+ "BLOCK_SIZE_M": 16,
64
+ "BLOCK_SIZE_N": 128,
65
+ "BLOCK_SIZE_K": 128,
66
+ "GROUP_SIZE_M": 2,
67
+ "SPLIT_K": 1,
68
+ "num_warps": 4,
69
+ "num_stages": 1,
70
+ "matrix_instr_nonkdim": 16
71
+ },
72
+ "48": {
73
+ "BLOCK_SIZE_M": 16,
74
+ "BLOCK_SIZE_N": 256,
75
+ "BLOCK_SIZE_K": 64,
76
+ "GROUP_SIZE_M": 1,
77
+ "SPLIT_K": 1,
78
+ "num_warps": 4,
79
+ "num_stages": 2,
80
+ "matrix_instr_nonkdim": 16
81
+ },
82
+ "64": {
83
+ "BLOCK_SIZE_M": 16,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 1,
87
+ "SPLIT_K": 1,
88
+ "num_warps": 4,
89
+ "num_stages": 1,
90
+ "matrix_instr_nonkdim": 16
91
+ },
92
+ "96": {
93
+ "BLOCK_SIZE_M": 16,
94
+ "BLOCK_SIZE_N": 128,
95
+ "BLOCK_SIZE_K": 128,
96
+ "GROUP_SIZE_M": 1,
97
+ "SPLIT_K": 1,
98
+ "num_warps": 4,
99
+ "num_stages": 1,
100
+ "matrix_instr_nonkdim": 16
101
+ },
102
+ "128": {
103
+ "BLOCK_SIZE_M": 16,
104
+ "BLOCK_SIZE_N": 256,
105
+ "BLOCK_SIZE_K": 128,
106
+ "GROUP_SIZE_M": 1,
107
+ "SPLIT_K": 1,
108
+ "num_warps": 4,
109
+ "num_stages": 2,
110
+ "matrix_instr_nonkdim": 16
111
+ },
112
+ "256": {
113
+ "BLOCK_SIZE_M": 16,
114
+ "BLOCK_SIZE_N": 256,
115
+ "BLOCK_SIZE_K": 128,
116
+ "GROUP_SIZE_M": 1,
117
+ "SPLIT_K": 1,
118
+ "num_warps": 4,
119
+ "num_stages": 2,
120
+ "matrix_instr_nonkdim": 16
121
+ },
122
+ "512": {
123
+ "BLOCK_SIZE_M": 32,
124
+ "BLOCK_SIZE_N": 128,
125
+ "BLOCK_SIZE_K": 128,
126
+ "GROUP_SIZE_M": 2,
127
+ "SPLIT_K": 1,
128
+ "num_warps": 4,
129
+ "num_stages": 2,
130
+ "matrix_instr_nonkdim": 32
131
+ },
132
+ "1024": {
133
+ "BLOCK_SIZE_M": 32,
134
+ "BLOCK_SIZE_N": 128,
135
+ "BLOCK_SIZE_K": 128,
136
+ "GROUP_SIZE_M": 2,
137
+ "SPLIT_K": 1,
138
+ "num_warps": 4,
139
+ "num_stages": 2,
140
+ "matrix_instr_nonkdim": 32
141
+ },
142
+ "1536": {
143
+ "BLOCK_SIZE_M": 32,
144
+ "BLOCK_SIZE_N": 128,
145
+ "BLOCK_SIZE_K": 128,
146
+ "GROUP_SIZE_M": 2,
147
+ "SPLIT_K": 1,
148
+ "num_warps": 4,
149
+ "num_stages": 2,
150
+ "matrix_instr_nonkdim": 32
151
+ },
152
+ "2048": {
153
+ "BLOCK_SIZE_M": 64,
154
+ "BLOCK_SIZE_N": 128,
155
+ "BLOCK_SIZE_K": 128,
156
+ "GROUP_SIZE_M": 1,
157
+ "SPLIT_K": 1,
158
+ "num_warps": 4,
159
+ "num_stages": 1,
160
+ "matrix_instr_nonkdim": 16
161
+ },
162
+ "3072": {
163
+ "BLOCK_SIZE_M": 64,
164
+ "BLOCK_SIZE_N": 128,
165
+ "BLOCK_SIZE_K": 128,
166
+ "GROUP_SIZE_M": 1,
167
+ "SPLIT_K": 1,
168
+ "num_warps": 4,
169
+ "num_stages": 1,
170
+ "matrix_instr_nonkdim": 16
171
+ },
172
+ "4096": {
173
+ "BLOCK_SIZE_M": 128,
174
+ "BLOCK_SIZE_N": 128,
175
+ "BLOCK_SIZE_K": 64,
176
+ "GROUP_SIZE_M": 2,
177
+ "SPLIT_K": 1,
178
+ "num_warps": 4,
179
+ "num_stages": 2,
180
+ "matrix_instr_nonkdim": 16
181
+ },
182
+ "8192": {
183
+ "BLOCK_SIZE_M": 256,
184
+ "BLOCK_SIZE_N": 256,
185
+ "BLOCK_SIZE_K": 32,
186
+ "GROUP_SIZE_M": 2,
187
+ "SPLIT_K": 1,
188
+ "num_warps": 8,
189
+ "num_stages": 2,
190
+ "matrix_instr_nonkdim": 32
191
+ }
192
+}
\ No newline at end of file
vllm/model_executor/layers/fused_moe/configs/E=384,N=256,device_name=AMD_Instinct_MI350_OAM,dtype=int4_w4a16.json
new
+192
@@ -0,0 +1,192 @@
1
+{
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 128,
5
+ "BLOCK_SIZE_K": 128,
6
+ "GROUP_SIZE_M": 2,
7
+ "SPLIT_K": 1,
8
+ "num_warps": 4,
9
+ "num_stages": 2,
10
+ "matrix_instr_nonkdim": 16
11
+ },
12
+ "2": {
13
+ "BLOCK_SIZE_M": 16,
14
+ "BLOCK_SIZE_N": 128,
15
+ "BLOCK_SIZE_K": 128,
16
+ "GROUP_SIZE_M": 2,
17
+ "SPLIT_K": 1,
18
+ "num_warps": 4,
19
+ "num_stages": 2,
20
+ "matrix_instr_nonkdim": 16
21
+ },
22
+ "4": {
23
+ "BLOCK_SIZE_M": 16,
24
+ "BLOCK_SIZE_N": 128,
25
+ "BLOCK_SIZE_K": 128,
26
+ "GROUP_SIZE_M": 2,
27
+ "SPLIT_K": 1,
28
+ "num_warps": 4,
29
+ "num_stages": 2,
30
+ "matrix_instr_nonkdim": 16
31
+ },
32
+ "8": {
33
+ "BLOCK_SIZE_M": 16,
34
+ "BLOCK_SIZE_N": 128,
35
+ "BLOCK_SIZE_K": 128,
36
+ "GROUP_SIZE_M": 2,
37
+ "SPLIT_K": 1,
38
+ "num_warps": 4,
39
+ "num_stages": 2,
40
+ "matrix_instr_nonkdim": 16
41
+ },
42
+ "16": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 128,
45
+ "BLOCK_SIZE_K": 128,
46
+ "GROUP_SIZE_M": 1,
47
+ "SPLIT_K": 1,
48
+ "num_warps": 4,
49
+ "num_stages": 2,
50
+ "matrix_instr_nonkdim": 16
51
+ },
52
+ "24": {
53
+ "BLOCK_SIZE_M": 16,
54
+ "BLOCK_SIZE_N": 128,
55
+ "BLOCK_SIZE_K": 64,
56
+ "GROUP_SIZE_M": 1,
57
+ "SPLIT_K": 1,
58
+ "num_warps": 4,
59
+ "num_stages": 2,
60
+ "matrix_instr_nonkdim": 16
61
+ },
62
+ "32": {
63
+ "BLOCK_SIZE_M": 16,
64
+ "BLOCK_SIZE_N": 128,
65
+ "BLOCK_SIZE_K": 128,
66
+ "GROUP_SIZE_M": 2,
67
+ "SPLIT_K": 1,
68
+ "num_warps": 4,
69
+ "num_stages": 1,
70
+ "matrix_instr_nonkdim": 16
71
+ },
72
+ "48": {
73
+ "BLOCK_SIZE_M": 16,
74
+ "BLOCK_SIZE_N": 256,
75
+ "BLOCK_SIZE_K": 64,
76
+ "GROUP_SIZE_M": 1,
77
+ "SPLIT_K": 1,
78
+ "num_warps": 4,
79
+ "num_stages": 2,
80
+ "matrix_instr_nonkdim": 16
81
+ },
82
+ "64": {
83
+ "BLOCK_SIZE_M": 16,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 1,
87
+ "SPLIT_K": 1,
88
+ "num_warps": 4,
89
+ "num_stages": 1,
90
+ "matrix_instr_nonkdim": 16
91
+ },
92
+ "96": {
93
+ "BLOCK_SIZE_M": 16,
94
+ "BLOCK_SIZE_N": 128,
95
+ "BLOCK_SIZE_K": 128,
96
+ "GROUP_SIZE_M": 1,
97
+ "SPLIT_K": 1,
98
+ "num_warps": 4,
99
+ "num_stages": 1,
100
+ "matrix_instr_nonkdim": 16
101
+ },
102
+ "128": {
103
+ "BLOCK_SIZE_M": 16,
104
+ "BLOCK_SIZE_N": 256,
105
+ "BLOCK_SIZE_K": 128,
106
+ "GROUP_SIZE_M": 1,
107
+ "SPLIT_K": 1,
108
+ "num_warps": 4,
109
+ "num_stages": 2,
110
+ "matrix_instr_nonkdim": 16
111
+ },
112
+ "256": {
113
+ "BLOCK_SIZE_M": 16,
114
+ "BLOCK_SIZE_N": 256,
115
+ "BLOCK_SIZE_K": 128,
116
+ "GROUP_SIZE_M": 1,
117
+ "SPLIT_K": 1,
118
+ "num_warps": 4,
119
+ "num_stages": 2,
120
+ "matrix_instr_nonkdim": 16
121
+ },
122
+ "512": {
123
+ "BLOCK_SIZE_M": 32,
124
+ "BLOCK_SIZE_N": 128,
125
+ "BLOCK_SIZE_K": 128,
126
+ "GROUP_SIZE_M": 2,
127
+ "SPLIT_K": 1,
128
+ "num_warps": 4,
129
+ "num_stages": 2,
130
+ "matrix_instr_nonkdim": 32
131
+ },
132
+ "1024": {
133
+ "BLOCK_SIZE_M": 32,
134
+ "BLOCK_SIZE_N": 128,
135
+ "BLOCK_SIZE_K": 128,
136
+ "GROUP_SIZE_M": 2,
137
+ "SPLIT_K": 1,
138
+ "num_warps": 4,
139
+ "num_stages": 2,
140
+ "matrix_instr_nonkdim": 32
141
+ },
142
+ "1536": {
143
+ "BLOCK_SIZE_M": 32,
144
+ "BLOCK_SIZE_N": 128,
145
+ "BLOCK_SIZE_K": 128,
146
+ "GROUP_SIZE_M": 2,
147
+ "SPLIT_K": 1,
148
+ "num_warps": 4,
149
+ "num_stages": 2,
150
+ "matrix_instr_nonkdim": 32
151
+ },
152
+ "2048": {
153
+ "BLOCK_SIZE_M": 64,
154
+ "BLOCK_SIZE_N": 128,
155
+ "BLOCK_SIZE_K": 128,
156
+ "GROUP_SIZE_M": 1,
157
+ "SPLIT_K": 1,
158
+ "num_warps": 4,
159
+ "num_stages": 1,
160
+ "matrix_instr_nonkdim": 16
161
+ },
162
+ "3072": {
163
+ "BLOCK_SIZE_M": 64,
164
+ "BLOCK_SIZE_N": 128,
165
+ "BLOCK_SIZE_K": 128,
166
+ "GROUP_SIZE_M": 1,
167
+ "SPLIT_K": 1,
168
+ "num_warps": 4,
169
+ "num_stages": 1,
170
+ "matrix_instr_nonkdim": 16
171
+ },
172
+ "4096": {
173
+ "BLOCK_SIZE_M": 128,
174
+ "BLOCK_SIZE_N": 128,
175
+ "BLOCK_SIZE_K": 64,
176
+ "GROUP_SIZE_M": 2,
177
+ "SPLIT_K": 1,
178
+ "num_warps": 4,
179
+ "num_stages": 2,
180
+ "matrix_instr_nonkdim": 16
181
+ },
182
+ "8192": {
183
+ "BLOCK_SIZE_M": 256,
184
+ "BLOCK_SIZE_N": 256,
185
+ "BLOCK_SIZE_K": 32,
186
+ "GROUP_SIZE_M": 2,
187
+ "SPLIT_K": 1,
188
+ "num_warps": 8,
189
+ "num_stages": 2,
190
+ "matrix_instr_nonkdim": 32
191
+ }
192
+}
\ No newline at end of file
vllm/model_executor/layers/fused_moe/configs/E=384,N=256,device_name=AMD_Instinct_MI355X,dtype=int4_w4a16.json
new
+192
@@ -0,0 +1,192 @@
1
+{
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 128,
5
+ "BLOCK_SIZE_K": 128,
6
+ "GROUP_SIZE_M": 2,
7
+ "SPLIT_K": 1,
8
+ "num_warps": 4,
9
+ "num_stages": 2,
10
+ "matrix_instr_nonkdim": 16
11
+ },
12
+ "2": {
13
+ "BLOCK_SIZE_M": 16,
14
+ "BLOCK_SIZE_N": 128,
15
+ "BLOCK_SIZE_K": 128,
16
+ "GROUP_SIZE_M": 2,
17
+ "SPLIT_K": 1,
18
+ "num_warps": 4,
19
+ "num_stages": 2,
20
+ "matrix_instr_nonkdim": 16
21
+ },
22
+ "4": {
23
+ "BLOCK_SIZE_M": 16,
24
+ "BLOCK_SIZE_N": 128,
25
+ "BLOCK_SIZE_K": 128,
26
+ "GROUP_SIZE_M": 2,
27
+ "SPLIT_K": 1,
28
+ "num_warps": 4,
29
+ "num_stages": 2,
30
+ "matrix_instr_nonkdim": 16
31
+ },
32
+ "8": {
33
+ "BLOCK_SIZE_M": 16,
34
+ "BLOCK_SIZE_N": 128,
35
+ "BLOCK_SIZE_K": 128,
36
+ "GROUP_SIZE_M": 2,
37
+ "SPLIT_K": 1,
38
+ "num_warps": 4,
39
+ "num_stages": 2,
40
+ "matrix_instr_nonkdim": 16
41
+ },
42
+ "16": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 128,
45
+ "BLOCK_SIZE_K": 128,
46
+ "GROUP_SIZE_M": 1,
47
+ "SPLIT_K": 1,
48
+ "num_warps": 4,
49
+ "num_stages": 2,
50
+ "matrix_instr_nonkdim": 16
51
+ },
52
+ "24": {
53
+ "BLOCK_SIZE_M": 16,
54
+ "BLOCK_SIZE_N": 128,
55
+ "BLOCK_SIZE_K": 64,
56
+ "GROUP_SIZE_M": 1,
57
+ "SPLIT_K": 1,
58
+ "num_warps": 4,
59
+ "num_stages": 2,
60
+ "matrix_instr_nonkdim": 16
61
+ },
62
+ "32": {
63
+ "BLOCK_SIZE_M": 16,
64
+ "BLOCK_SIZE_N": 128,
65
+ "BLOCK_SIZE_K": 128,
66
+ "GROUP_SIZE_M": 2,
67
+ "SPLIT_K": 1,
68
+ "num_warps": 4,
69
+ "num_stages": 1,
70
+ "matrix_instr_nonkdim": 16
71
+ },
72
+ "48": {
73
+ "BLOCK_SIZE_M": 16,
74
+ "BLOCK_SIZE_N": 256,
75
+ "BLOCK_SIZE_K": 64,
76
+ "GROUP_SIZE_M": 1,
77
+ "SPLIT_K": 1,
78
+ "num_warps": 4,
79
+ "num_stages": 2,
80
+ "matrix_instr_nonkdim": 16
81
+ },
82
+ "64": {
83
+ "BLOCK_SIZE_M": 16,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 1,
87
+ "SPLIT_K": 1,
88
+ "num_warps": 4,
89
+ "num_stages": 1,
90
+ "matrix_instr_nonkdim": 16
91
+ },
92
+ "96": {
93
+ "BLOCK_SIZE_M": 16,
94
+ "BLOCK_SIZE_N": 128,
95
+ "BLOCK_SIZE_K": 128,
96
+ "GROUP_SIZE_M": 1,
97
+ "SPLIT_K": 1,
98
+ "num_warps": 4,
99
+ "num_stages": 1,
100
+ "matrix_instr_nonkdim": 16
101
+ },
102
+ "128": {
103
+ "BLOCK_SIZE_M": 16,
104
+ "BLOCK_SIZE_N": 256,
105
+ "BLOCK_SIZE_K": 128,
106
+ "GROUP_SIZE_M": 1,
107
+ "SPLIT_K": 1,
108
+ "num_warps": 4,
109
+ "num_stages": 2,
110
+ "matrix_instr_nonkdim": 16
111
+ },
112
+ "256": {
113
+ "BLOCK_SIZE_M": 16,
114
+ "BLOCK_SIZE_N": 256,
115
+ "BLOCK_SIZE_K": 128,
116
+ "GROUP_SIZE_M": 1,
117
+ "SPLIT_K": 1,
118
+ "num_warps": 4,
119
+ "num_stages": 2,
120
+ "matrix_instr_nonkdim": 16
121
+ },
122
+ "512": {
123
+ "BLOCK_SIZE_M": 32,
124
+ "BLOCK_SIZE_N": 128,
125
+ "BLOCK_SIZE_K": 128,
126
+ "GROUP_SIZE_M": 2,
127
+ "SPLIT_K": 1,
128
+ "num_warps": 4,
129
+ "num_stages": 2,
130
+ "matrix_instr_nonkdim": 32
131
+ },
132
+ "1024": {
133
+ "BLOCK_SIZE_M": 32,
134
+ "BLOCK_SIZE_N": 128,
135
+ "BLOCK_SIZE_K": 128,
136
+ "GROUP_SIZE_M": 2,
137
+ "SPLIT_K": 1,
138
+ "num_warps": 4,
139
+ "num_stages": 2,
140
+ "matrix_instr_nonkdim": 32
141
+ },
142
+ "1536": {
143
+ "BLOCK_SIZE_M": 32,
144
+ "BLOCK_SIZE_N": 128,
145
+ "BLOCK_SIZE_K": 128,
146
+ "GROUP_SIZE_M": 2,
147
+ "SPLIT_K": 1,
148
+ "num_warps": 4,
149
+ "num_stages": 2,
150
+ "matrix_instr_nonkdim": 32
151
+ },
152
+ "2048": {
153
+ "BLOCK_SIZE_M": 64,
154
+ "BLOCK_SIZE_N": 128,
155
+ "BLOCK_SIZE_K": 128,
156
+ "GROUP_SIZE_M": 1,
157
+ "SPLIT_K": 1,
158
+ "num_warps": 4,
159
+ "num_stages": 1,
160
+ "matrix_instr_nonkdim": 16
161
+ },
162
+ "3072": {
163
+ "BLOCK_SIZE_M": 64,
164
+ "BLOCK_SIZE_N": 128,
165
+ "BLOCK_SIZE_K": 128,
166
+ "GROUP_SIZE_M": 1,
167
+ "SPLIT_K": 1,
168
+ "num_warps": 4,
169
+ "num_stages": 1,
170
+ "matrix_instr_nonkdim": 16
171
+ },
172
+ "4096": {
173
+ "BLOCK_SIZE_M": 128,
174
+ "BLOCK_SIZE_N": 128,
175
+ "BLOCK_SIZE_K": 64,
176
+ "GROUP_SIZE_M": 2,
177
+ "SPLIT_K": 1,
178
+ "num_warps": 4,
179
+ "num_stages": 2,
180
+ "matrix_instr_nonkdim": 16
181
+ },
182
+ "8192": {
183
+ "BLOCK_SIZE_M": 256,
184
+ "BLOCK_SIZE_N": 256,
185
+ "BLOCK_SIZE_K": 32,
186
+ "GROUP_SIZE_M": 2,
187
+ "SPLIT_K": 1,
188
+ "num_warps": 8,
189
+ "num_stages": 2,
190
+ "matrix_instr_nonkdim": 32
191
+ }
192
+}
\ No newline at end of file
vllm/model_executor/layers/fused_moe/configs/E=384,N=256,device_name=AMD_Instinct_MI355_OAM,dtype=int4_w4a16.json
new
+192
@@ -0,0 +1,192 @@
1
+{
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 128,
5
+ "BLOCK_SIZE_K": 128,
6
+ "GROUP_SIZE_M": 2,
7
+ "SPLIT_K": 1,
8
+ "num_warps": 4,
9
+ "num_stages": 2,
10
+ "matrix_instr_nonkdim": 16
11
+ },
12
+ "2": {
13
+ "BLOCK_SIZE_M": 16,
14
+ "BLOCK_SIZE_N": 128,
15
+ "BLOCK_SIZE_K": 128,
16
+ "GROUP_SIZE_M": 2,
17
+ "SPLIT_K": 1,
18
+ "num_warps": 4,
19
+ "num_stages": 2,
20
+ "matrix_instr_nonkdim": 16
21
+ },
22
+ "4": {
23
+ "BLOCK_SIZE_M": 16,
24
+ "BLOCK_SIZE_N": 128,
25
+ "BLOCK_SIZE_K": 128,
26
+ "GROUP_SIZE_M": 2,
27
+ "SPLIT_K": 1,
28
+ "num_warps": 4,
29
+ "num_stages": 2,
30
+ "matrix_instr_nonkdim": 16
31
+ },
32
+ "8": {
33
+ "BLOCK_SIZE_M": 16,
34
+ "BLOCK_SIZE_N": 128,
35
+ "BLOCK_SIZE_K": 128,
36
+ "GROUP_SIZE_M": 2,
37
+ "SPLIT_K": 1,
38
+ "num_warps": 4,
39
+ "num_stages": 2,
40
+ "matrix_instr_nonkdim": 16
41
+ },
42
+ "16": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 128,
45
+ "BLOCK_SIZE_K": 128,
46
+ "GROUP_SIZE_M": 1,
47
+ "SPLIT_K": 1,
48
+ "num_warps": 4,
49
+ "num_stages": 2,
50
+ "matrix_instr_nonkdim": 16
51
+ },
52
+ "24": {
53
+ "BLOCK_SIZE_M": 16,
54
+ "BLOCK_SIZE_N": 128,
55
+ "BLOCK_SIZE_K": 64,
56
+ "GROUP_SIZE_M": 1,
57
+ "SPLIT_K": 1,
58
+ "num_warps": 4,
59
+ "num_stages": 2,
60
+ "matrix_instr_nonkdim": 16
61
+ },
62
+ "32": {
63
+ "BLOCK_SIZE_M": 16,
64
+ "BLOCK_SIZE_N": 128,
65
+ "BLOCK_SIZE_K": 128,
66
+ "GROUP_SIZE_M": 2,
67
+ "SPLIT_K": 1,
68
+ "num_warps": 4,
69
+ "num_stages": 1,
70
+ "matrix_instr_nonkdim": 16
71
+ },
72
+ "48": {
73
+ "BLOCK_SIZE_M": 16,
74
+ "BLOCK_SIZE_N": 256,
75
+ "BLOCK_SIZE_K": 64,
76
+ "GROUP_SIZE_M": 1,
77
+ "SPLIT_K": 1,
78
+ "num_warps": 4,
79
+ "num_stages": 2,
80
+ "matrix_instr_nonkdim": 16
81
+ },
82
+ "64": {
83
+ "BLOCK_SIZE_M": 16,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 1,
87
+ "SPLIT_K": 1,
88
+ "num_warps": 4,
89
+ "num_stages": 1,
90
+ "matrix_instr_nonkdim": 16
91
+ },
92
+ "96": {
93
+ "BLOCK_SIZE_M": 16,
94
+ "BLOCK_SIZE_N": 128,
95
+ "BLOCK_SIZE_K": 128,
96
+ "GROUP_SIZE_M": 1,
97
+ "SPLIT_K": 1,
98
+ "num_warps": 4,
99
+ "num_stages": 1,
100
+ "matrix_instr_nonkdim": 16
101
+ },
102
+ "128": {
103
+ "BLOCK_SIZE_M": 16,
104
+ "BLOCK_SIZE_N": 256,
105
+ "BLOCK_SIZE_K": 128,
106
+ "GROUP_SIZE_M": 1,
107
+ "SPLIT_K": 1,
108
+ "num_warps": 4,
109
+ "num_stages": 2,
110
+ "matrix_instr_nonkdim": 16
111
+ },
112
+ "256": {
113
+ "BLOCK_SIZE_M": 16,
114
+ "BLOCK_SIZE_N": 256,
115
+ "BLOCK_SIZE_K": 128,
116
+ "GROUP_SIZE_M": 1,
117
+ "SPLIT_K": 1,
118
+ "num_warps": 4,
119
+ "num_stages": 2,
120
+ "matrix_instr_nonkdim": 16
121
+ },
122
+ "512": {
123
+ "BLOCK_SIZE_M": 32,
124
+ "BLOCK_SIZE_N": 128,
125
+ "BLOCK_SIZE_K": 128,
126
+ "GROUP_SIZE_M": 2,
127
+ "SPLIT_K": 1,
128
+ "num_warps": 4,
129
+ "num_stages": 2,
130
+ "matrix_instr_nonkdim": 32
131
+ },
132
+ "1024": {
133
+ "BLOCK_SIZE_M": 32,
134
+ "BLOCK_SIZE_N": 128,
135
+ "BLOCK_SIZE_K": 128,
136
+ "GROUP_SIZE_M": 2,
137
+ "SPLIT_K": 1,
138
+ "num_warps": 4,
139
+ "num_stages": 2,
140
+ "matrix_instr_nonkdim": 32
141
+ },
142
+ "1536": {
143
+ "BLOCK_SIZE_M": 32,
144
+ "BLOCK_SIZE_N": 128,
145
+ "BLOCK_SIZE_K": 128,
146
+ "GROUP_SIZE_M": 2,
147
+ "SPLIT_K": 1,
148
+ "num_warps": 4,
149
+ "num_stages": 2,
150
+ "matrix_instr_nonkdim": 32
151
+ },
152
+ "2048": {
153
+ "BLOCK_SIZE_M": 64,
154
+ "BLOCK_SIZE_N": 128,
155
+ "BLOCK_SIZE_K": 128,
156
+ "GROUP_SIZE_M": 1,
157
+ "SPLIT_K": 1,
158
+ "num_warps": 4,
159
+ "num_stages": 1,
160
+ "matrix_instr_nonkdim": 16
161
+ },
162
+ "3072": {
163
+ "BLOCK_SIZE_M": 64,
164
+ "BLOCK_SIZE_N": 128,
165
+ "BLOCK_SIZE_K": 128,
166
+ "GROUP_SIZE_M": 1,
167
+ "SPLIT_K": 1,
168
+ "num_warps": 4,
169
+ "num_stages": 1,
170
+ "matrix_instr_nonkdim": 16
171
+ },
172
+ "4096": {
173
+ "BLOCK_SIZE_M": 128,
174
+ "BLOCK_SIZE_N": 128,
175
+ "BLOCK_SIZE_K": 64,
176
+ "GROUP_SIZE_M": 2,
177
+ "SPLIT_K": 1,
178
+ "num_warps": 4,
179
+ "num_stages": 2,
180
+ "matrix_instr_nonkdim": 16
181
+ },
182
+ "8192": {
183
+ "BLOCK_SIZE_M": 256,
184
+ "BLOCK_SIZE_N": 256,
185
+ "BLOCK_SIZE_K": 32,
186
+ "GROUP_SIZE_M": 2,
187
+ "SPLIT_K": 1,
188
+ "num_warps": 8,
189
+ "num_stages": 2,
190
+ "matrix_instr_nonkdim": 32
191
+ }
192
+}
\ No newline at end of file