@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / ac243886b0
Fork 0

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 14, 05:30 UTC

[Kernel] MI-300X triton moe configs (#23445)

Signed-off-by: Sara Kokkila Schumacher <saraks@ibm.com>

Sara-KS committed Sep 22, 2025 at 09:29 UTC ac243886b02399d327314b2d8bbf9989fd07f247
6 files changed +1200
vllm/model_executor/layers/fused_moe/configs/E=62,N=128,device_name=AMD_Instinct_MI300X.json new
+200
@@ -0,0 +1,200 @@
1 +{
2 + "1": {
3 + "BLOCK_SIZE_M": 16,
4 + "BLOCK_SIZE_N": 16,
5 + "BLOCK_SIZE_K": 128,
6 + "GROUP_SIZE_M": 1,
7 + "num_warps": 4,
8 + "num_stages": 2,
9 + "waves_per_eu": 0,
10 + "matrix_instr_nonkdim": 16,
11 + "kpack": 1
12 + },
13 + "2": {
14 + "BLOCK_SIZE_M": 16,
15 + "BLOCK_SIZE_N": 16,
16 + "BLOCK_SIZE_K": 128,
17 + "GROUP_SIZE_M": 1,
18 + "num_warps": 4,
19 + "num_stages": 2,
20 + "waves_per_eu": 0,
21 + "matrix_instr_nonkdim": 16,
22 + "kpack": 1
23 + },
24 + "4": {
25 + "BLOCK_SIZE_M": 16,
26 + "BLOCK_SIZE_N": 16,
27 + "BLOCK_SIZE_K": 128,
28 + "GROUP_SIZE_M": 1,
29 + "num_warps": 4,
30 + "num_stages": 2,
31 + "waves_per_eu": 0,
32 + "matrix_instr_nonkdim": 16,
33 + "kpack": 1
34 + },
35 + "8": {
36 + "BLOCK_SIZE_M": 16,
37 + "BLOCK_SIZE_N": 16,
38 + "BLOCK_SIZE_K": 64,
39 + "GROUP_SIZE_M": 1,
40 + "num_warps": 4,
41 + "num_stages": 2,
42 + "waves_per_eu": 0,
43 + "matrix_instr_nonkdim": 16,
44 + "kpack": 2
45 + },
46 + "16": {
47 + "BLOCK_SIZE_M": 16,
48 + "BLOCK_SIZE_N": 64,
49 + "BLOCK_SIZE_K": 64,
50 + "GROUP_SIZE_M": 1,
51 + "num_warps": 4,
52 + "num_stages": 2,
53 + "waves_per_eu": 0,
54 + "matrix_instr_nonkdim": 16,
55 + "kpack": 2
56 + },
57 + "24": {
58 + "BLOCK_SIZE_M": 16,
59 + "BLOCK_SIZE_N": 64,
60 + "BLOCK_SIZE_K": 64,
61 + "GROUP_SIZE_M": 1,
62 + "num_warps": 4,
63 + "num_stages": 2,
64 + "waves_per_eu": 0,
65 + "matrix_instr_nonkdim": 16,
66 + "kpack": 2
67 + },
68 + "32": {
69 + "BLOCK_SIZE_M": 16,
70 + "BLOCK_SIZE_N": 64,
71 + "BLOCK_SIZE_K": 64,
72 + "GROUP_SIZE_M": 1,
73 + "num_warps": 4,
74 + "num_stages": 2,
75 + "waves_per_eu": 0,
76 + "matrix_instr_nonkdim": 16,
77 + "kpack": 2
78 + },
79 + "48": {
80 + "BLOCK_SIZE_M": 16,
81 + "BLOCK_SIZE_N": 64,
82 + "BLOCK_SIZE_K": 64,
83 + "GROUP_SIZE_M": 1,
84 + "num_warps": 4,
85 + "num_stages": 2,
86 + "waves_per_eu": 0,
87 + "matrix_instr_nonkdim": 16,
88 + "kpack": 2
89 + },
90 + "64": {
91 + "BLOCK_SIZE_M": 16,
92 + "BLOCK_SIZE_N": 64,
93 + "BLOCK_SIZE_K": 64,
94 + "GROUP_SIZE_M": 1,
95 + "num_warps": 4,
96 + "num_stages": 2,
97 + "waves_per_eu": 0,
98 + "matrix_instr_nonkdim": 16,
99 + "kpack": 2
100 + },
101 + "96": {
102 + "BLOCK_SIZE_M": 16,
103 + "BLOCK_SIZE_N": 64,
104 + "BLOCK_SIZE_K": 64,
105 + "GROUP_SIZE_M": 1,
106 + "num_warps": 4,
107 + "num_stages": 2,
108 + "waves_per_eu": 0,
109 + "matrix_instr_nonkdim": 16,
110 + "kpack": 2
111 + },
112 + "128": {
113 + "BLOCK_SIZE_M": 16,
114 + "BLOCK_SIZE_N": 64,
115 + "BLOCK_SIZE_K": 128,
116 + "GROUP_SIZE_M": 8,
117 + "num_warps": 4,
118 + "num_stages": 2,
119 + "waves_per_eu": 0,
120 + "matrix_instr_nonkdim": 16,
121 + "kpack": 2
122 + },
123 + "256": {
124 + "BLOCK_SIZE_M": 32,
125 + "BLOCK_SIZE_N": 64,
126 + "BLOCK_SIZE_K": 128,
127 + "GROUP_SIZE_M": 4,
128 + "num_warps": 8,
129 + "num_stages": 2,
130 + "waves_per_eu": 0,
131 + "matrix_instr_nonkdim": 16,
132 + "kpack": 1
133 + },
134 + "512": {
135 + "BLOCK_SIZE_M": 32,
136 + "BLOCK_SIZE_N": 128,
137 + "BLOCK_SIZE_K": 64,
138 + "GROUP_SIZE_M": 4,
139 + "num_warps": 8,
140 + "num_stages": 2,
141 + "waves_per_eu": 0,
142 + "matrix_instr_nonkdim": 16,
143 + "kpack": 2
144 + },
145 + "1024": {
146 + "BLOCK_SIZE_M": 16,
147 + "BLOCK_SIZE_N": 128,
148 + "BLOCK_SIZE_K": 64,
149 + "GROUP_SIZE_M": 1,
150 + "num_warps": 4,
151 + "num_stages": 2,
152 + "waves_per_eu": 0,
153 + "matrix_instr_nonkdim": 16,
154 + "kpack": 2
155 + },
156 + "1536": {
157 + "BLOCK_SIZE_M": 16,
158 + "BLOCK_SIZE_N": 256,
159 + "BLOCK_SIZE_K": 64,
160 + "GROUP_SIZE_M": 1,
161 + "num_warps": 4,
162 + "num_stages": 2,
163 + "waves_per_eu": 0,
164 + "matrix_instr_nonkdim": 16,
165 + "kpack": 2
166 + },
167 + "2048": {
168 + "BLOCK_SIZE_M": 32,
169 + "BLOCK_SIZE_N": 256,
170 + "BLOCK_SIZE_K": 32,
171 + "GROUP_SIZE_M": 4,
172 + "num_warps": 8,
173 + "num_stages": 2,
174 + "waves_per_eu": 0,
175 + "matrix_instr_nonkdim": 16,
176 + "kpack": 1
177 + },
178 + "3072": {
179 + "BLOCK_SIZE_M": 32,
180 + "BLOCK_SIZE_N": 256,
181 + "BLOCK_SIZE_K": 32,
182 + "GROUP_SIZE_M": 1,
183 + "num_warps": 8,
184 + "num_stages": 2,
185 + "waves_per_eu": 0,
186 + "matrix_instr_nonkdim": 16,
187 + "kpack": 1
188 + },
189 + "4096": {
190 + "BLOCK_SIZE_M": 32,
191 + "BLOCK_SIZE_N": 256,
192 + "BLOCK_SIZE_K": 32,
193 + "GROUP_SIZE_M": 4,
194 + "num_warps": 8,
195 + "num_stages": 2,
196 + "waves_per_eu": 0,
197 + "matrix_instr_nonkdim": 16,
198 + "kpack": 1
199 + }
200 +}
vllm/model_executor/layers/fused_moe/configs/E=62,N=256,device_name=AMD_Instinct_MI300X.json new
+200
@@ -0,0 +1,200 @@
1 +{
2 + "1": {
3 + "BLOCK_SIZE_M": 16,
4 + "BLOCK_SIZE_N": 16,
5 + "BLOCK_SIZE_K": 256,
6 + "GROUP_SIZE_M": 1,
7 + "num_warps": 2,
8 + "num_stages": 2,
9 + "waves_per_eu": 0,
10 + "matrix_instr_nonkdim": 16,
11 + "kpack": 2
12 + },
13 + "2": {
14 + "BLOCK_SIZE_M": 16,
15 + "BLOCK_SIZE_N": 16,
16 + "BLOCK_SIZE_K": 256,
17 + "GROUP_SIZE_M": 1,
18 + "num_warps": 2,
19 + "num_stages": 2,
20 + "waves_per_eu": 0,
21 + "matrix_instr_nonkdim": 16,
22 + "kpack": 2
23 + },
24 + "4": {
25 + "BLOCK_SIZE_M": 16,
26 + "BLOCK_SIZE_N": 16,
27 + "BLOCK_SIZE_K": 128,
28 + "GROUP_SIZE_M": 1,
29 + "num_warps": 2,
30 + "num_stages": 2,
31 + "waves_per_eu": 0,
32 + "matrix_instr_nonkdim": 16,
33 + "kpack": 2
34 + },
35 + "8": {
36 + "BLOCK_SIZE_M": 16,
37 + "BLOCK_SIZE_N": 32,
38 + "BLOCK_SIZE_K": 64,
39 + "GROUP_SIZE_M": 1,
40 + "num_warps": 2,
41 + "num_stages": 2,
42 + "waves_per_eu": 0,
43 + "matrix_instr_nonkdim": 16,
44 + "kpack": 1
45 + },
46 + "16": {
47 + "BLOCK_SIZE_M": 16,
48 + "BLOCK_SIZE_N": 32,
49 + "BLOCK_SIZE_K": 64,
50 + "GROUP_SIZE_M": 4,
51 + "num_warps": 2,
52 + "num_stages": 2,
53 + "waves_per_eu": 0,
54 + "matrix_instr_nonkdim": 16,
55 + "kpack": 2
56 + },
57 + "24": {
58 + "BLOCK_SIZE_M": 16,
59 + "BLOCK_SIZE_N": 64,
60 + "BLOCK_SIZE_K": 64,
61 + "GROUP_SIZE_M": 1,
62 + "num_warps": 4,
63 + "num_stages": 2,
64 + "waves_per_eu": 0,
65 + "matrix_instr_nonkdim": 16,
66 + "kpack": 2
67 + },
68 + "32": {
69 + "BLOCK_SIZE_M": 16,
70 + "BLOCK_SIZE_N": 64,
71 + "BLOCK_SIZE_K": 64,
72 + "GROUP_SIZE_M": 1,
73 + "num_warps": 4,
74 + "num_stages": 2,
75 + "waves_per_eu": 0,
76 + "matrix_instr_nonkdim": 16,
77 + "kpack": 2
78 + },
79 + "48": {
80 + "BLOCK_SIZE_M": 16,
81 + "BLOCK_SIZE_N": 64,
82 + "BLOCK_SIZE_K": 64,
83 + "GROUP_SIZE_M": 8,
84 + "num_warps": 4,
85 + "num_stages": 2,
86 + "waves_per_eu": 0,
87 + "matrix_instr_nonkdim": 16,
88 + "kpack": 2
89 + },
90 + "64": {
91 + "BLOCK_SIZE_M": 16,
92 + "BLOCK_SIZE_N": 64,
93 + "BLOCK_SIZE_K": 64,
94 + "GROUP_SIZE_M": 8,
95 + "num_warps": 4,
96 + "num_stages": 2,
97 + "waves_per_eu": 0,
98 + "matrix_instr_nonkdim": 16,
99 + "kpack": 2
100 + },
101 + "96": {
102 + "BLOCK_SIZE_M": 16,
103 + "BLOCK_SIZE_N": 64,
104 + "BLOCK_SIZE_K": 64,
105 + "GROUP_SIZE_M": 8,
106 + "num_warps": 4,
107 + "num_stages": 2,
108 + "waves_per_eu": 0,
109 + "matrix_instr_nonkdim": 16,
110 + "kpack": 1
111 + },
112 + "128": {
113 + "BLOCK_SIZE_M": 16,
114 + "BLOCK_SIZE_N": 64,
115 + "BLOCK_SIZE_K": 64,
116 + "GROUP_SIZE_M": 1,
117 + "num_warps": 4,
118 + "num_stages": 2,
119 + "waves_per_eu": 0,
120 + "matrix_instr_nonkdim": 16,
121 + "kpack": 2
122 + },
123 + "256": {
124 + "BLOCK_SIZE_M": 32,
125 + "BLOCK_SIZE_N": 128,
126 + "BLOCK_SIZE_K": 64,
127 + "GROUP_SIZE_M": 8,
128 + "num_warps": 8,
129 + "num_stages": 2,
130 + "waves_per_eu": 0,
131 + "matrix_instr_nonkdim": 16,
132 + "kpack": 2
133 + },
134 + "512": {
135 + "BLOCK_SIZE_M": 64,
136 + "BLOCK_SIZE_N": 128,
137 + "BLOCK_SIZE_K": 64,
138 + "GROUP_SIZE_M": 8,
139 + "num_warps": 8,
140 + "num_stages": 2,
141 + "waves_per_eu": 0,
142 + "matrix_instr_nonkdim": 16,
143 + "kpack": 2
144 + },
145 + "1024": {
146 + "BLOCK_SIZE_M": 128,
147 + "BLOCK_SIZE_N": 64,
148 + "BLOCK_SIZE_K": 64,
149 + "GROUP_SIZE_M": 8,
150 + "num_warps": 4,
151 + "num_stages": 2,
152 + "waves_per_eu": 0,
153 + "matrix_instr_nonkdim": 16,
154 + "kpack": 2
155 + },
156 + "1536": {
157 + "BLOCK_SIZE_M": 256,
158 + "BLOCK_SIZE_N": 128,
159 + "BLOCK_SIZE_K": 64,
160 + "GROUP_SIZE_M": 4,
161 + "num_warps": 8,
162 + "num_stages": 2,
163 + "waves_per_eu": 0,
164 + "matrix_instr_nonkdim": 16,
165 + "kpack": 2
166 + },
167 + "2048": {
168 + "BLOCK_SIZE_M": 256,
169 + "BLOCK_SIZE_N": 128,
170 + "BLOCK_SIZE_K": 64,
171 + "GROUP_SIZE_M": 16,
172 + "num_warps": 8,
173 + "num_stages": 2,
174 + "waves_per_eu": 0,
175 + "matrix_instr_nonkdim": 16,
176 + "kpack": 2
177 + },
178 + "3072": {
179 + "BLOCK_SIZE_M": 128,
180 + "BLOCK_SIZE_N": 128,
181 + "BLOCK_SIZE_K": 64,
182 + "GROUP_SIZE_M": 8,
183 + "num_warps": 4,
184 + "num_stages": 2,
185 + "waves_per_eu": 0,
186 + "matrix_instr_nonkdim": 16,
187 + "kpack": 2
188 + },
189 + "4096": {
190 + "BLOCK_SIZE_M": 128,
191 + "BLOCK_SIZE_N": 128,
192 + "BLOCK_SIZE_K": 64,
193 + "GROUP_SIZE_M": 1,
194 + "num_warps": 4,
195 + "num_stages": 2,
196 + "waves_per_eu": 0,
197 + "matrix_instr_nonkdim": 16,
198 + "kpack": 2
199 + }
200 +}
vllm/model_executor/layers/fused_moe/configs/E=62,N=512,device_name=AMD_Instinct_MI300X.json new
+200
@@ -0,0 +1,200 @@
1 +{
2 + "1": {
3 + "BLOCK_SIZE_M": 16,
4 + "BLOCK_SIZE_N": 16,
5 + "BLOCK_SIZE_K": 256,
6 + "GROUP_SIZE_M": 1,
7 + "num_warps": 2,
8 + "num_stages": 2,
9 + "waves_per_eu": 0,
10 + "matrix_instr_nonkdim": 16,
11 + "kpack": 2
12 + },
13 + "2": {
14 + "BLOCK_SIZE_M": 16,
15 + "BLOCK_SIZE_N": 16,
16 + "BLOCK_SIZE_K": 128,
17 + "GROUP_SIZE_M": 1,
18 + "num_warps": 2,
19 + "num_stages": 2,
20 + "waves_per_eu": 0,
21 + "matrix_instr_nonkdim": 16,
22 + "kpack": 2
23 + },
24 + "4": {
25 + "BLOCK_SIZE_M": 16,
26 + "BLOCK_SIZE_N": 32,
27 + "BLOCK_SIZE_K": 64,
28 + "GROUP_SIZE_M": 1,
29 + "num_warps": 2,
30 + "num_stages": 2,
31 + "waves_per_eu": 0,
32 + "matrix_instr_nonkdim": 16,
33 + "kpack": 2
34 + },
35 + "8": {
36 + "BLOCK_SIZE_M": 32,
37 + "BLOCK_SIZE_N": 64,
38 + "BLOCK_SIZE_K": 64,
39 + "GROUP_SIZE_M": 1,
40 + "num_warps": 4,
41 + "num_stages": 2,
42 + "waves_per_eu": 0,
43 + "matrix_instr_nonkdim": 16,
44 + "kpack": 2
45 + },
46 + "16": {
47 + "BLOCK_SIZE_M": 32,
48 + "BLOCK_SIZE_N": 64,
49 + "BLOCK_SIZE_K": 64,
50 + "GROUP_SIZE_M": 1,
51 + "num_warps": 4,
52 + "num_stages": 2,
53 + "waves_per_eu": 0,
54 + "matrix_instr_nonkdim": 16,
55 + "kpack": 1
56 + },
57 + "24": {
58 + "BLOCK_SIZE_M": 16,
59 + "BLOCK_SIZE_N": 64,
60 + "BLOCK_SIZE_K": 64,
61 + "GROUP_SIZE_M": 1,
62 + "num_warps": 4,
63 + "num_stages": 2,
64 + "waves_per_eu": 0,
65 + "matrix_instr_nonkdim": 16,
66 + "kpack": 2
67 + },
68 + "32": {
69 + "BLOCK_SIZE_M": 16,
70 + "BLOCK_SIZE_N": 32,
71 + "BLOCK_SIZE_K": 64,
72 + "GROUP_SIZE_M": 1,
73 + "num_warps": 2,
74 + "num_stages": 2,
75 + "waves_per_eu": 0,
76 + "matrix_instr_nonkdim": 16,
77 + "kpack": 2
78 + },
79 + "48": {
80 + "BLOCK_SIZE_M": 16,
81 + "BLOCK_SIZE_N": 32,
82 + "BLOCK_SIZE_K": 64,
83 + "GROUP_SIZE_M": 16,
84 + "num_warps": 2,
85 + "num_stages": 2,
86 + "waves_per_eu": 0,
87 + "matrix_instr_nonkdim": 16,
88 + "kpack": 1
89 + },
90 + "64": {
91 + "BLOCK_SIZE_M": 16,
92 + "BLOCK_SIZE_N": 32,
93 + "BLOCK_SIZE_K": 64,
94 + "GROUP_SIZE_M": 16,
95 + "num_warps": 2,
96 + "num_stages": 2,
97 + "waves_per_eu": 0,
98 + "matrix_instr_nonkdim": 16,
99 + "kpack": 1
100 + },
101 + "96": {
102 + "BLOCK_SIZE_M": 16,
103 + "BLOCK_SIZE_N": 128,
104 + "BLOCK_SIZE_K": 128,
105 + "GROUP_SIZE_M": 8,
106 + "num_warps": 8,
107 + "num_stages": 2,
108 + "waves_per_eu": 0,
109 + "matrix_instr_nonkdim": 16,
110 + "kpack": 2
111 + },
112 + "128": {
113 + "BLOCK_SIZE_M": 16,
114 + "BLOCK_SIZE_N": 128,
115 + "BLOCK_SIZE_K": 64,
116 + "GROUP_SIZE_M": 4,
117 + "num_warps": 8,
118 + "num_stages": 2,
119 + "waves_per_eu": 0,
120 + "matrix_instr_nonkdim": 16,
121 + "kpack": 2
122 + },
123 + "256": {
124 + "BLOCK_SIZE_M": 32,
125 + "BLOCK_SIZE_N": 64,
126 + "BLOCK_SIZE_K": 128,
127 + "GROUP_SIZE_M": 4,
128 + "num_warps": 8,
129 + "num_stages": 2,
130 + "waves_per_eu": 0,
131 + "matrix_instr_nonkdim": 16,
132 + "kpack": 2
133 + },
134 + "512": {
135 + "BLOCK_SIZE_M": 64,
136 + "BLOCK_SIZE_N": 128,
137 + "BLOCK_SIZE_K": 128,
138 + "GROUP_SIZE_M": 8,
139 + "num_warps": 8,
140 + "num_stages": 2,
141 + "waves_per_eu": 0,
142 + "matrix_instr_nonkdim": 16,
143 + "kpack": 2
144 + },
145 + "1024": {
146 + "BLOCK_SIZE_M": 128,
147 + "BLOCK_SIZE_N": 128,
148 + "BLOCK_SIZE_K": 64,
149 + "GROUP_SIZE_M": 16,
150 + "num_warps": 4,
151 + "num_stages": 2,
152 + "waves_per_eu": 0,
153 + "matrix_instr_nonkdim": 16,
154 + "kpack": 2
155 + },
156 + "1536": {
157 + "BLOCK_SIZE_M": 256,
158 + "BLOCK_SIZE_N": 128,
159 + "BLOCK_SIZE_K": 64,
160 + "GROUP_SIZE_M": 16,
161 + "num_warps": 8,
162 + "num_stages": 2,
163 + "waves_per_eu": 0,
164 + "matrix_instr_nonkdim": 16,
165 + "kpack": 2
166 + },
167 + "2048": {
168 + "BLOCK_SIZE_M": 256,
169 + "BLOCK_SIZE_N": 128,
170 + "BLOCK_SIZE_K": 64,
171 + "GROUP_SIZE_M": 4,
172 + "num_warps": 8,
173 + "num_stages": 2,
174 + "waves_per_eu": 0,
175 + "matrix_instr_nonkdim": 16,
176 + "kpack": 2
177 + },
178 + "3072": {
179 + "BLOCK_SIZE_M": 128,
180 + "BLOCK_SIZE_N": 128,
181 + "BLOCK_SIZE_K": 64,
182 + "GROUP_SIZE_M": 1,
183 + "num_warps": 4,
184 + "num_stages": 2,
185 + "waves_per_eu": 0,
186 + "matrix_instr_nonkdim": 16,
187 + "kpack": 2
188 + },
189 + "4096": {
190 + "BLOCK_SIZE_M": 128,
191 + "BLOCK_SIZE_N": 128,
192 + "BLOCK_SIZE_K": 64,
193 + "GROUP_SIZE_M": 1,
194 + "num_warps": 4,
195 + "num_stages": 2,
196 + "waves_per_eu": 0,
197 + "matrix_instr_nonkdim": 16,
198 + "kpack": 2
199 + }
200 +}
vllm/model_executor/layers/fused_moe/configs/E=72,N=192,device_name=AMD_Instinct_MI300X.json new
+200
@@ -0,0 +1,200 @@
1 +{
2 + "1": {
3 + "BLOCK_SIZE_M": 16,
4 + "BLOCK_SIZE_N": 16,
5 + "BLOCK_SIZE_K": 256,
6 + "GROUP_SIZE_M": 1,
7 + "num_warps": 4,
8 + "num_stages": 2,
9 + "waves_per_eu": 0,
10 + "matrix_instr_nonkdim": 16,
11 + "kpack": 1
12 + },
13 + "2": {
14 + "BLOCK_SIZE_M": 16,
15 + "BLOCK_SIZE_N": 16,
16 + "BLOCK_SIZE_K": 128,
17 + "GROUP_SIZE_M": 1,
18 + "num_warps": 4,
19 + "num_stages": 2,
20 + "waves_per_eu": 0,
21 + "matrix_instr_nonkdim": 16,
22 + "kpack": 1
23 + },
24 + "4": {
25 + "BLOCK_SIZE_M": 16,
26 + "BLOCK_SIZE_N": 16,
27 + "BLOCK_SIZE_K": 64,
28 + "GROUP_SIZE_M": 1,
29 + "num_warps": 4,
30 + "num_stages": 2,
31 + "waves_per_eu": 0,
32 + "matrix_instr_nonkdim": 16,
33 + "kpack": 2
34 + },
35 + "8": {
36 + "BLOCK_SIZE_M": 16,
37 + "BLOCK_SIZE_N": 32,
38 + "BLOCK_SIZE_K": 64,
39 + "GROUP_SIZE_M": 1,
40 + "num_warps": 4,
41 + "num_stages": 2,
42 + "waves_per_eu": 0,
43 + "matrix_instr_nonkdim": 16,
44 + "kpack": 1
45 + },
46 + "16": {
47 + "BLOCK_SIZE_M": 16,
48 + "BLOCK_SIZE_N": 64,
49 + "BLOCK_SIZE_K": 64,
50 + "GROUP_SIZE_M": 1,
51 + "num_warps": 4,
52 + "num_stages": 2,
53 + "waves_per_eu": 0,
54 + "matrix_instr_nonkdim": 16,
55 + "kpack": 2
56 + },
57 + "24": {
58 + "BLOCK_SIZE_M": 16,
59 + "BLOCK_SIZE_N": 64,
60 + "BLOCK_SIZE_K": 64,
61 + "GROUP_SIZE_M": 1,
62 + "num_warps": 4,
63 + "num_stages": 2,
64 + "waves_per_eu": 0,
65 + "matrix_instr_nonkdim": 16,
66 + "kpack": 2
67 + },
68 + "32": {
69 + "BLOCK_SIZE_M": 16,
70 + "BLOCK_SIZE_N": 64,
71 + "BLOCK_SIZE_K": 64,
72 + "GROUP_SIZE_M": 1,
73 + "num_warps": 4,
74 + "num_stages": 2,
75 + "waves_per_eu": 0,
76 + "matrix_instr_nonkdim": 16,
77 + "kpack": 2
78 + },
79 + "48": {
80 + "BLOCK_SIZE_M": 16,
81 + "BLOCK_SIZE_N": 64,
82 + "BLOCK_SIZE_K": 64,
83 + "GROUP_SIZE_M": 1,
84 + "num_warps": 4,
85 + "num_stages": 2,
86 + "waves_per_eu": 0,
87 + "matrix_instr_nonkdim": 16,
88 + "kpack": 2
89 + },
90 + "64": {
91 + "BLOCK_SIZE_M": 16,
92 + "BLOCK_SIZE_N": 64,
93 + "BLOCK_SIZE_K": 64,
94 + "GROUP_SIZE_M": 1,
95 + "num_warps": 4,
96 + "num_stages": 2,
97 + "waves_per_eu": 0,
98 + "matrix_instr_nonkdim": 16,
99 + "kpack": 2
100 + },
101 + "96": {
102 + "BLOCK_SIZE_M": 16,
103 + "BLOCK_SIZE_N": 64,
104 + "BLOCK_SIZE_K": 64,
105 + "GROUP_SIZE_M": 8,
106 + "num_warps": 4,
107 + "num_stages": 2,
108 + "waves_per_eu": 0,
109 + "matrix_instr_nonkdim": 16,
110 + "kpack": 2
111 + },
112 + "128": {
113 + "BLOCK_SIZE_M": 16,
114 + "BLOCK_SIZE_N": 128,
115 + "BLOCK_SIZE_K": 64,
116 + "GROUP_SIZE_M": 4,
117 + "num_warps": 8,
118 + "num_stages": 2,
119 + "waves_per_eu": 0,
120 + "matrix_instr_nonkdim": 16,
121 + "kpack": 2
122 + },
123 + "256": {
124 + "BLOCK_SIZE_M": 32,
125 + "BLOCK_SIZE_N": 128,
126 + "BLOCK_SIZE_K": 64,
127 + "GROUP_SIZE_M": 8,
128 + "num_warps": 8,
129 + "num_stages": 2,
130 + "waves_per_eu": 0,
131 + "matrix_instr_nonkdim": 16,
132 + "kpack": 2
133 + },
134 + "512": {
135 + "BLOCK_SIZE_M": 16,
136 + "BLOCK_SIZE_N": 128,
137 + "BLOCK_SIZE_K": 64,
138 + "GROUP_SIZE_M": 1,
139 + "num_warps": 4,
140 + "num_stages": 2,
141 + "waves_per_eu": 0,
142 + "matrix_instr_nonkdim": 16,
143 + "kpack": 2
144 + },
145 + "1024": {
146 + "BLOCK_SIZE_M": 32,
147 + "BLOCK_SIZE_N": 256,
148 + "BLOCK_SIZE_K": 32,
149 + "GROUP_SIZE_M": 8,
150 + "num_warps": 8,
151 + "num_stages": 2,
152 + "waves_per_eu": 0,
153 + "matrix_instr_nonkdim": 16,
154 + "kpack": 1
155 + },
156 + "1536": {
157 + "BLOCK_SIZE_M": 32,
158 + "BLOCK_SIZE_N": 128,
159 + "BLOCK_SIZE_K": 64,
160 + "GROUP_SIZE_M": 1,
161 + "num_warps": 8,
162 + "num_stages": 2,
163 + "waves_per_eu": 0,
164 + "matrix_instr_nonkdim": 16,
165 + "kpack": 1
166 + },
167 + "2048": {
168 + "BLOCK_SIZE_M": 32,
169 + "BLOCK_SIZE_N": 128,
170 + "BLOCK_SIZE_K": 64,
171 + "GROUP_SIZE_M": 1,
172 + "num_warps": 8,
173 + "num_stages": 2,
174 + "waves_per_eu": 0,
175 + "matrix_instr_nonkdim": 16,
176 + "kpack": 1
177 + },
178 + "3072": {
179 + "BLOCK_SIZE_M": 32,
180 + "BLOCK_SIZE_N": 128,
181 + "BLOCK_SIZE_K": 64,
182 + "GROUP_SIZE_M": 1,
183 + "num_warps": 8,
184 + "num_stages": 2,
185 + "waves_per_eu": 0,
186 + "matrix_instr_nonkdim": 16,
187 + "kpack": 1
188 + },
189 + "4096": {
190 + "BLOCK_SIZE_M": 32,
191 + "BLOCK_SIZE_N": 256,
192 + "BLOCK_SIZE_K": 32,
193 + "GROUP_SIZE_M": 1,
194 + "num_warps": 8,
195 + "num_stages": 2,
196 + "waves_per_eu": 0,
197 + "matrix_instr_nonkdim": 16,
198 + "kpack": 1
199 + }
200 +}
vllm/model_executor/layers/fused_moe/configs/E=72,N=384,device_name=AMD_Instinct_MI300X.json new
+200
@@ -0,0 +1,200 @@
1 +{
2 + "1": {
3 + "BLOCK_SIZE_M": 16,
4 + "BLOCK_SIZE_N": 16,
5 + "BLOCK_SIZE_K": 256,
6 + "GROUP_SIZE_M": 1,
7 + "num_warps": 4,
8 + "num_stages": 2,
9 + "waves_per_eu": 0,
10 + "matrix_instr_nonkdim": 16,
11 + "kpack": 1
12 + },
13 + "2": {
14 + "BLOCK_SIZE_M": 16,
15 + "BLOCK_SIZE_N": 16,
16 + "BLOCK_SIZE_K": 128,
17 + "GROUP_SIZE_M": 1,
18 + "num_warps": 4,
19 + "num_stages": 2,
20 + "waves_per_eu": 0,
21 + "matrix_instr_nonkdim": 16,
22 + "kpack": 1
23 + },
24 + "4": {
25 + "BLOCK_SIZE_M": 16,
26 + "BLOCK_SIZE_N": 16,
27 + "BLOCK_SIZE_K": 64,
28 + "GROUP_SIZE_M": 1,
29 + "num_warps": 2,
30 + "num_stages": 2,
31 + "waves_per_eu": 0,
32 + "matrix_instr_nonkdim": 16,
33 + "kpack": 1
34 + },
35 + "8": {
36 + "BLOCK_SIZE_M": 16,
37 + "BLOCK_SIZE_N": 128,
38 + "BLOCK_SIZE_K": 64,
39 + "GROUP_SIZE_M": 1,
40 + "num_warps": 8,
41 + "num_stages": 2,
42 + "waves_per_eu": 0,
43 + "matrix_instr_nonkdim": 16,
44 + "kpack": 2
45 + },
46 + "16": {
47 + "BLOCK_SIZE_M": 16,
48 + "BLOCK_SIZE_N": 32,
49 + "BLOCK_SIZE_K": 64,
50 + "GROUP_SIZE_M": 1,
51 + "num_warps": 1,
52 + "num_stages": 2,
53 + "waves_per_eu": 0,
54 + "matrix_instr_nonkdim": 16,
55 + "kpack": 2
56 + },
57 + "24": {
58 + "BLOCK_SIZE_M": 16,
59 + "BLOCK_SIZE_N": 64,
60 + "BLOCK_SIZE_K": 64,
61 + "GROUP_SIZE_M": 4,
62 + "num_warps": 2,
63 + "num_stages": 2,
64 + "waves_per_eu": 0,
65 + "matrix_instr_nonkdim": 16,
66 + "kpack": 2
67 + },
68 + "32": {
69 + "BLOCK_SIZE_M": 16,
70 + "BLOCK_SIZE_N": 64,
71 + "BLOCK_SIZE_K": 64,
72 + "GROUP_SIZE_M": 1,
73 + "num_warps": 2,
74 + "num_stages": 2,
75 + "waves_per_eu": 0,
76 + "matrix_instr_nonkdim": 16,
77 + "kpack": 2
78 + },
79 + "48": {
80 + "BLOCK_SIZE_M": 16,
81 + "BLOCK_SIZE_N": 64,
82 + "BLOCK_SIZE_K": 64,
83 + "GROUP_SIZE_M": 8,
84 + "num_warps": 4,
85 + "num_stages": 2,
86 + "waves_per_eu": 0,
87 + "matrix_instr_nonkdim": 16,
88 + "kpack": 2
89 + },
90 + "64": {
91 + "BLOCK_SIZE_M": 16,
92 + "BLOCK_SIZE_N": 64,
93 + "BLOCK_SIZE_K": 64,
94 + "GROUP_SIZE_M": 8,
95 + "num_warps": 4,
96 + "num_stages": 2,
97 + "waves_per_eu": 0,
98 + "matrix_instr_nonkdim": 16,
99 + "kpack": 2
100 + },
101 + "96": {
102 + "BLOCK_SIZE_M": 16,
103 + "BLOCK_SIZE_N": 64,
104 + "BLOCK_SIZE_K": 64,
105 + "GROUP_SIZE_M": 4,
106 + "num_warps": 2,
107 + "num_stages": 2,
108 + "waves_per_eu": 0,
109 + "matrix_instr_nonkdim": 16,
110 + "kpack": 2
111 + },
112 + "128": {
113 + "BLOCK_SIZE_M": 16,
114 + "BLOCK_SIZE_N": 64,
115 + "BLOCK_SIZE_K": 64,
116 + "GROUP_SIZE_M": 1,
117 + "num_warps": 4,
118 + "num_stages": 2,
119 + "waves_per_eu": 0,
120 + "matrix_instr_nonkdim": 16,
121 + "kpack": 2
122 + },
123 + "256": {
124 + "BLOCK_SIZE_M": 32,
125 + "BLOCK_SIZE_N": 64,
126 + "BLOCK_SIZE_K": 64,
127 + "GROUP_SIZE_M": 4,
128 + "num_warps": 8,
129 + "num_stages": 2,
130 + "waves_per_eu": 0,
131 + "matrix_instr_nonkdim": 16,
132 + "kpack": 1
133 + },
134 + "512": {
135 + "BLOCK_SIZE_M": 64,
136 + "BLOCK_SIZE_N": 64,
137 + "BLOCK_SIZE_K": 64,
138 + "GROUP_SIZE_M": 8,
139 + "num_warps": 4,
140 + "num_stages": 2,
141 + "waves_per_eu": 0,
142 + "matrix_instr_nonkdim": 16,
143 + "kpack": 2
144 + },
145 + "1024": {
146 + "BLOCK_SIZE_M": 128,
147 + "BLOCK_SIZE_N": 128,
148 + "BLOCK_SIZE_K": 64,
149 + "GROUP_SIZE_M": 4,
150 + "num_warps": 8,
151 + "num_stages": 2,
152 + "waves_per_eu": 0,
153 + "matrix_instr_nonkdim": 16,
154 + "kpack": 2
155 + },
156 + "1536": {
157 + "BLOCK_SIZE_M": 128,
158 + "BLOCK_SIZE_N": 128,
159 + "BLOCK_SIZE_K": 64,
160 + "GROUP_SIZE_M": 4,
161 + "num_warps": 8,
162 + "num_stages": 2,
163 + "waves_per_eu": 0,
164 + "matrix_instr_nonkdim": 16,
165 + "kpack": 2
166 + },
167 + "2048": {
168 + "BLOCK_SIZE_M": 256,
169 + "BLOCK_SIZE_N": 256,
170 + "BLOCK_SIZE_K": 32,
171 + "GROUP_SIZE_M": 16,
172 + "num_warps": 8,
173 + "num_stages": 2,
174 + "waves_per_eu": 0,
175 + "matrix_instr_nonkdim": 16,
176 + "kpack": 1
177 + },
178 + "3072": {
179 + "BLOCK_SIZE_M": 128,
180 + "BLOCK_SIZE_N": 128,
181 + "BLOCK_SIZE_K": 64,
182 + "GROUP_SIZE_M": 4,
183 + "num_warps": 8,
184 + "num_stages": 2,
185 + "waves_per_eu": 0,
186 + "matrix_instr_nonkdim": 16,
187 + "kpack": 2
188 + },
189 + "4096": {
190 + "BLOCK_SIZE_M": 128,
191 + "BLOCK_SIZE_N": 128,
192 + "BLOCK_SIZE_K": 64,
193 + "GROUP_SIZE_M": 8,
194 + "num_warps": 8,
195 + "num_stages": 2,
196 + "waves_per_eu": 0,
197 + "matrix_instr_nonkdim": 16,
198 + "kpack": 2
199 + }
200 +}
vllm/model_executor/layers/fused_moe/configs/E=72,N=768,device_name=AMD_Instinct_MI300X.json new
+200
@@ -0,0 +1,200 @@
1 +{
2 + "1": {
3 + "BLOCK_SIZE_M": 16,
4 + "BLOCK_SIZE_N": 16,
5 + "BLOCK_SIZE_K": 128,
6 + "GROUP_SIZE_M": 1,
7 + "num_warps": 4,
8 + "num_stages": 2,
9 + "waves_per_eu": 0,
10 + "matrix_instr_nonkdim": 16,
11 + "kpack": 2
12 + },
13 + "2": {
14 + "BLOCK_SIZE_M": 16,
15 + "BLOCK_SIZE_N": 64,
16 + "BLOCK_SIZE_K": 128,
17 + "GROUP_SIZE_M": 1,
18 + "num_warps": 4,
19 + "num_stages": 2,
20 + "waves_per_eu": 0,
21 + "matrix_instr_nonkdim": 16,
22 + "kpack": 2
23 + },
24 + "4": {
25 + "BLOCK_SIZE_M": 16,
26 + "BLOCK_SIZE_N": 128,
27 + "BLOCK_SIZE_K": 64,
28 + "GROUP_SIZE_M": 1,
29 + "num_warps": 8,
30 + "num_stages": 2,
31 + "waves_per_eu": 0,
32 + "matrix_instr_nonkdim": 16,
33 + "kpack": 2
34 + },
35 + "8": {
36 + "BLOCK_SIZE_M": 32,
37 + "BLOCK_SIZE_N": 32,
38 + "BLOCK_SIZE_K": 64,
39 + "GROUP_SIZE_M": 1,
40 + "num_warps": 2,
41 + "num_stages": 2,
42 + "waves_per_eu": 0,
43 + "matrix_instr_nonkdim": 16,
44 + "kpack": 1
45 + },
46 + "16": {
47 + "BLOCK_SIZE_M": 16,
48 + "BLOCK_SIZE_N": 32,
49 + "BLOCK_SIZE_K": 64,
50 + "GROUP_SIZE_M": 4,
51 + "num_warps": 2,
52 + "num_stages": 2,
53 + "waves_per_eu": 0,
54 + "matrix_instr_nonkdim": 16,
55 + "kpack": 2
56 + },
57 + "24": {
58 + "BLOCK_SIZE_M": 16,
59 + "BLOCK_SIZE_N": 64,
60 + "BLOCK_SIZE_K": 128,
61 + "GROUP_SIZE_M": 1,
62 + "num_warps": 4,
63 + "num_stages": 2,
64 + "waves_per_eu": 0,
65 + "matrix_instr_nonkdim": 16,
66 + "kpack": 2
67 + },
68 + "32": {
69 + "BLOCK_SIZE_M": 16,
70 + "BLOCK_SIZE_N": 64,
71 + "BLOCK_SIZE_K": 128,
72 + "GROUP_SIZE_M": 1,
73 + "num_warps": 4,
74 + "num_stages": 2,
75 + "waves_per_eu": 0,
76 + "matrix_instr_nonkdim": 16,
77 + "kpack": 2
78 + },
79 + "48": {
80 + "BLOCK_SIZE_M": 16,
81 + "BLOCK_SIZE_N": 128,
82 + "BLOCK_SIZE_K": 128,
83 + "GROUP_SIZE_M": 4,
84 + "num_warps": 8,
85 + "num_stages": 2,
86 + "waves_per_eu": 0,
87 + "matrix_instr_nonkdim": 16,
88 + "kpack": 1
89 + },
90 + "64": {
91 + "BLOCK_SIZE_M": 16,
92 + "BLOCK_SIZE_N": 128,
93 + "BLOCK_SIZE_K": 128,
94 + "GROUP_SIZE_M": 1,
95 + "num_warps": 8,
96 + "num_stages": 2,
97 + "waves_per_eu": 0,
98 + "matrix_instr_nonkdim": 16,
99 + "kpack": 1
100 + },
101 + "96": {
102 + "BLOCK_SIZE_M": 16,
103 + "BLOCK_SIZE_N": 128,
104 + "BLOCK_SIZE_K": 128,
105 + "GROUP_SIZE_M": 4,
106 + "num_warps": 8,
107 + "num_stages": 2,
108 + "waves_per_eu": 0,
109 + "matrix_instr_nonkdim": 16,
110 + "kpack": 2
111 + },
112 + "128": {
113 + "BLOCK_SIZE_M": 32,
114 + "BLOCK_SIZE_N": 128,
115 + "BLOCK_SIZE_K": 128,
116 + "GROUP_SIZE_M": 1,
117 + "num_warps": 8,
118 + "num_stages": 2,
119 + "waves_per_eu": 0,
120 + "matrix_instr_nonkdim": 16,
121 + "kpack": 2
122 + },
123 + "256": {
124 + "BLOCK_SIZE_M": 32,
125 + "BLOCK_SIZE_N": 64,
126 + "BLOCK_SIZE_K": 128,
127 + "GROUP_SIZE_M": 1,
128 + "num_warps": 8,
129 + "num_stages": 2,
130 + "waves_per_eu": 0,
131 + "matrix_instr_nonkdim": 16,
132 + "kpack": 2
133 + },
134 + "512": {
135 + "BLOCK_SIZE_M": 64,
136 + "BLOCK_SIZE_N": 128,
137 + "BLOCK_SIZE_K": 128,
138 + "GROUP_SIZE_M": 4,
139 + "num_warps": 8,
140 + "num_stages": 2,
141 + "waves_per_eu": 0,
142 + "matrix_instr_nonkdim": 16,
143 + "kpack": 2
144 + },
145 + "1024": {
146 + "BLOCK_SIZE_M": 128,
147 + "BLOCK_SIZE_N": 128,
148 + "BLOCK_SIZE_K": 64,
149 + "GROUP_SIZE_M": 8,
150 + "num_warps": 8,
151 + "num_stages": 2,
152 + "waves_per_eu": 0,
153 + "matrix_instr_nonkdim": 16,
154 + "kpack": 2
155 + },
156 + "1536": {
157 + "BLOCK_SIZE_M": 64,
158 + "BLOCK_SIZE_N": 64,
159 + "BLOCK_SIZE_K": 128,
160 + "GROUP_SIZE_M": 1,
161 + "num_warps": 8,
162 + "num_stages": 2,
163 + "waves_per_eu": 0,
164 + "matrix_instr_nonkdim": 16,
165 + "kpack": 2
166 + },
167 + "2048": {
168 + "BLOCK_SIZE_M": 256,
169 + "BLOCK_SIZE_N": 128,
170 + "BLOCK_SIZE_K": 64,
171 + "GROUP_SIZE_M": 4,
172 + "num_warps": 8,
173 + "num_stages": 2,
174 + "waves_per_eu": 0,
175 + "matrix_instr_nonkdim": 16,
176 + "kpack": 2
177 + },
178 + "3072": {
179 + "BLOCK_SIZE_M": 128,
180 + "BLOCK_SIZE_N": 128,
181 + "BLOCK_SIZE_K": 64,
182 + "GROUP_SIZE_M": 16,
183 + "num_warps": 8,
184 + "num_stages": 2,
185 + "waves_per_eu": 0,
186 + "matrix_instr_nonkdim": 16,
187 + "kpack": 2
188 + },
189 + "4096": {
190 + "BLOCK_SIZE_M": 128,
191 + "BLOCK_SIZE_N": 128,
192 + "BLOCK_SIZE_K": 64,
193 + "GROUP_SIZE_M": 1,
194 + "num_warps": 8,
195 + "num_stages": 2,
196 + "waves_per_eu": 0,
197 + "matrix_instr_nonkdim": 16,
198 + "kpack": 2
199 + }
200 +}