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
[Hardware][AMD][Model] Triton MoE tuning configs for GLM-4.6 for MI300X (#27323)
Signed-off-by: minatoaquaMK2 <jiacheng.yue@foxmail.com>
Eric Yue committed
Oct 28, 2025 at 13:58 UTC
bdb01a38fe20d9eb97a2d219dcc5506cf24cccad
1 file changed
+201
vllm/model_executor/layers/fused_moe/configs/E=160,N=192,device_name=AMD_Instinct_MI300X.json
new
+201
@@ -0,0 +1,201 @@
1
+{
2
+ "triton_version": "3.4.0",
3
+ "1": {
4
+ "BLOCK_SIZE_M": 16,
5
+ "BLOCK_SIZE_N": 16,
6
+ "BLOCK_SIZE_K": 256,
7
+ "GROUP_SIZE_M": 1,
8
+ "num_warps": 4,
9
+ "num_stages": 2,
10
+ "waves_per_eu": 0,
11
+ "matrix_instr_nonkdim": 16,
12
+ "kpack": 2
13
+ },
14
+ "2": {
15
+ "BLOCK_SIZE_M": 16,
16
+ "BLOCK_SIZE_N": 64,
17
+ "BLOCK_SIZE_K": 128,
18
+ "GROUP_SIZE_M": 1,
19
+ "num_warps": 4,
20
+ "num_stages": 2,
21
+ "waves_per_eu": 0,
22
+ "matrix_instr_nonkdim": 16,
23
+ "kpack": 2
24
+ },
25
+ "4": {
26
+ "BLOCK_SIZE_M": 16,
27
+ "BLOCK_SIZE_N": 64,
28
+ "BLOCK_SIZE_K": 128,
29
+ "GROUP_SIZE_M": 1,
30
+ "num_warps": 4,
31
+ "num_stages": 2,
32
+ "waves_per_eu": 0,
33
+ "matrix_instr_nonkdim": 16,
34
+ "kpack": 2
35
+ },
36
+ "8": {
37
+ "BLOCK_SIZE_M": 16,
38
+ "BLOCK_SIZE_N": 64,
39
+ "BLOCK_SIZE_K": 128,
40
+ "GROUP_SIZE_M": 1,
41
+ "num_warps": 4,
42
+ "num_stages": 2,
43
+ "waves_per_eu": 0,
44
+ "matrix_instr_nonkdim": 16,
45
+ "kpack": 1
46
+ },
47
+ "16": {
48
+ "BLOCK_SIZE_M": 16,
49
+ "BLOCK_SIZE_N": 64,
50
+ "BLOCK_SIZE_K": 64,
51
+ "GROUP_SIZE_M": 4,
52
+ "num_warps": 4,
53
+ "num_stages": 2,
54
+ "waves_per_eu": 0,
55
+ "matrix_instr_nonkdim": 16,
56
+ "kpack": 2
57
+ },
58
+ "24": {
59
+ "BLOCK_SIZE_M": 16,
60
+ "BLOCK_SIZE_N": 32,
61
+ "BLOCK_SIZE_K": 64,
62
+ "GROUP_SIZE_M": 1,
63
+ "num_warps": 1,
64
+ "num_stages": 2,
65
+ "waves_per_eu": 0,
66
+ "matrix_instr_nonkdim": 16,
67
+ "kpack": 2
68
+ },
69
+ "32": {
70
+ "BLOCK_SIZE_M": 16,
71
+ "BLOCK_SIZE_N": 64,
72
+ "BLOCK_SIZE_K": 128,
73
+ "GROUP_SIZE_M": 4,
74
+ "num_warps": 2,
75
+ "num_stages": 2,
76
+ "waves_per_eu": 0,
77
+ "matrix_instr_nonkdim": 16,
78
+ "kpack": 2
79
+ },
80
+ "48": {
81
+ "BLOCK_SIZE_M": 16,
82
+ "BLOCK_SIZE_N": 64,
83
+ "BLOCK_SIZE_K": 64,
84
+ "GROUP_SIZE_M": 8,
85
+ "num_warps": 1,
86
+ "num_stages": 2,
87
+ "waves_per_eu": 0,
88
+ "matrix_instr_nonkdim": 16,
89
+ "kpack": 2
90
+ },
91
+ "64": {
92
+ "BLOCK_SIZE_M": 16,
93
+ "BLOCK_SIZE_N": 128,
94
+ "BLOCK_SIZE_K": 64,
95
+ "GROUP_SIZE_M": 4,
96
+ "num_warps": 8,
97
+ "num_stages": 2,
98
+ "waves_per_eu": 0,
99
+ "matrix_instr_nonkdim": 16,
100
+ "kpack": 2
101
+ },
102
+ "96": {
103
+ "BLOCK_SIZE_M": 16,
104
+ "BLOCK_SIZE_N": 128,
105
+ "BLOCK_SIZE_K": 64,
106
+ "GROUP_SIZE_M": 32,
107
+ "num_warps": 8,
108
+ "num_stages": 2,
109
+ "waves_per_eu": 0,
110
+ "matrix_instr_nonkdim": 16,
111
+ "kpack": 1
112
+ },
113
+ "128": {
114
+ "BLOCK_SIZE_M": 16,
115
+ "BLOCK_SIZE_N": 128,
116
+ "BLOCK_SIZE_K": 64,
117
+ "GROUP_SIZE_M": 16,
118
+ "num_warps": 8,
119
+ "num_stages": 2,
120
+ "waves_per_eu": 0,
121
+ "matrix_instr_nonkdim": 16,
122
+ "kpack": 2
123
+ },
124
+ "256": {
125
+ "BLOCK_SIZE_M": 16,
126
+ "BLOCK_SIZE_N": 128,
127
+ "BLOCK_SIZE_K": 64,
128
+ "GROUP_SIZE_M": 32,
129
+ "num_warps": 8,
130
+ "num_stages": 2,
131
+ "waves_per_eu": 0,
132
+ "matrix_instr_nonkdim": 16,
133
+ "kpack": 2
134
+ },
135
+ "512": {
136
+ "BLOCK_SIZE_M": 32,
137
+ "BLOCK_SIZE_N": 128,
138
+ "BLOCK_SIZE_K": 64,
139
+ "GROUP_SIZE_M": 32,
140
+ "num_warps": 4,
141
+ "num_stages": 2,
142
+ "waves_per_eu": 0,
143
+ "matrix_instr_nonkdim": 16,
144
+ "kpack": 1
145
+ },
146
+ "1024": {
147
+ "BLOCK_SIZE_M": 64,
148
+ "BLOCK_SIZE_N": 128,
149
+ "BLOCK_SIZE_K": 64,
150
+ "GROUP_SIZE_M": 4,
151
+ "num_warps": 8,
152
+ "num_stages": 2,
153
+ "waves_per_eu": 0,
154
+ "matrix_instr_nonkdim": 16,
155
+ "kpack": 2
156
+ },
157
+ "1536": {
158
+ "BLOCK_SIZE_M": 128,
159
+ "BLOCK_SIZE_N": 128,
160
+ "BLOCK_SIZE_K": 64,
161
+ "GROUP_SIZE_M": 16,
162
+ "num_warps": 4,
163
+ "num_stages": 2,
164
+ "waves_per_eu": 0,
165
+ "matrix_instr_nonkdim": 16,
166
+ "kpack": 2
167
+ },
168
+ "2048": {
169
+ "BLOCK_SIZE_M": 128,
170
+ "BLOCK_SIZE_N": 128,
171
+ "BLOCK_SIZE_K": 64,
172
+ "GROUP_SIZE_M": 4,
173
+ "num_warps": 8,
174
+ "num_stages": 2,
175
+ "waves_per_eu": 0,
176
+ "matrix_instr_nonkdim": 16,
177
+ "kpack": 1
178
+ },
179
+ "3072": {
180
+ "BLOCK_SIZE_M": 256,
181
+ "BLOCK_SIZE_N": 128,
182
+ "BLOCK_SIZE_K": 64,
183
+ "GROUP_SIZE_M": 32,
184
+ "num_warps": 8,
185
+ "num_stages": 2,
186
+ "waves_per_eu": 0,
187
+ "matrix_instr_nonkdim": 16,
188
+ "kpack": 2
189
+ },
190
+ "4096": {
191
+ "BLOCK_SIZE_M": 256,
192
+ "BLOCK_SIZE_N": 128,
193
+ "BLOCK_SIZE_K": 64,
194
+ "GROUP_SIZE_M": 16,
195
+ "num_warps": 8,
196
+ "num_stages": 2,
197
+ "waves_per_eu": 0,
198
+ "matrix_instr_nonkdim": 16,
199
+ "kpack": 2
200
+ }
201
+}