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 13, 03:15 UTC
[Kernel][Moe Configs] llama4 maverick fp8 moe config tp8 on mi325 (#28709)
Signed-off-by: Zhewen Li <zhewenli@meta.com>
Zhewen Li committed
Nov 15, 2025 at 01:10 UTC
1ec978c209391286d4cee968426900e9a4d256a5
1 file changed
+164
vllm/model_executor/layers/fused_moe/configs/E=128,N=1024,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json
new
+164
@@ -0,0 +1,164 @@
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
+ },
11
+ "2": {
12
+ "BLOCK_SIZE_M": 16,
13
+ "BLOCK_SIZE_N": 16,
14
+ "BLOCK_SIZE_K": 256,
15
+ "GROUP_SIZE_M": 1,
16
+ "num_warps": 8,
17
+ "num_stages": 2,
18
+ "waves_per_eu": 0
19
+ },
20
+ "4": {
21
+ "BLOCK_SIZE_M": 16,
22
+ "BLOCK_SIZE_N": 64,
23
+ "BLOCK_SIZE_K": 256,
24
+ "GROUP_SIZE_M": 1,
25
+ "num_warps": 4,
26
+ "num_stages": 2,
27
+ "waves_per_eu": 0
28
+ },
29
+ "8": {
30
+ "BLOCK_SIZE_M": 16,
31
+ "BLOCK_SIZE_N": 16,
32
+ "BLOCK_SIZE_K": 256,
33
+ "GROUP_SIZE_M": 1,
34
+ "num_warps": 2,
35
+ "num_stages": 2,
36
+ "waves_per_eu": 0
37
+ },
38
+ "16": {
39
+ "BLOCK_SIZE_M": 16,
40
+ "BLOCK_SIZE_N": 64,
41
+ "BLOCK_SIZE_K": 128,
42
+ "GROUP_SIZE_M": 1,
43
+ "num_warps": 4,
44
+ "num_stages": 2,
45
+ "waves_per_eu": 0
46
+ },
47
+ "24": {
48
+ "BLOCK_SIZE_M": 16,
49
+ "BLOCK_SIZE_N": 64,
50
+ "BLOCK_SIZE_K": 256,
51
+ "GROUP_SIZE_M": 1,
52
+ "num_warps": 2,
53
+ "num_stages": 2,
54
+ "waves_per_eu": 0
55
+ },
56
+ "32": {
57
+ "BLOCK_SIZE_M": 16,
58
+ "BLOCK_SIZE_N": 64,
59
+ "BLOCK_SIZE_K": 256,
60
+ "GROUP_SIZE_M": 1,
61
+ "num_warps": 4,
62
+ "num_stages": 2,
63
+ "waves_per_eu": 0
64
+ },
65
+ "48": {
66
+ "BLOCK_SIZE_M": 16,
67
+ "BLOCK_SIZE_N": 64,
68
+ "BLOCK_SIZE_K": 256,
69
+ "GROUP_SIZE_M": 1,
70
+ "num_warps": 2,
71
+ "num_stages": 2,
72
+ "waves_per_eu": 0
73
+ },
74
+ "64": {
75
+ "BLOCK_SIZE_M": 16,
76
+ "BLOCK_SIZE_N": 128,
77
+ "BLOCK_SIZE_K": 256,
78
+ "GROUP_SIZE_M": 1,
79
+ "num_warps": 4,
80
+ "num_stages": 2,
81
+ "waves_per_eu": 0
82
+ },
83
+ "96": {
84
+ "BLOCK_SIZE_M": 16,
85
+ "BLOCK_SIZE_N": 64,
86
+ "BLOCK_SIZE_K": 256,
87
+ "GROUP_SIZE_M": 1,
88
+ "num_warps": 4,
89
+ "num_stages": 2,
90
+ "waves_per_eu": 0
91
+ },
92
+ "128": {
93
+ "BLOCK_SIZE_M": 16,
94
+ "BLOCK_SIZE_N": 64,
95
+ "BLOCK_SIZE_K": 256,
96
+ "GROUP_SIZE_M": 4,
97
+ "num_warps": 4,
98
+ "num_stages": 2,
99
+ "waves_per_eu": 0
100
+ },
101
+ "256": {
102
+ "BLOCK_SIZE_M": 16,
103
+ "BLOCK_SIZE_N": 64,
104
+ "BLOCK_SIZE_K": 256,
105
+ "GROUP_SIZE_M": 16,
106
+ "num_warps": 4,
107
+ "num_stages": 2,
108
+ "waves_per_eu": 0
109
+ },
110
+ "512": {
111
+ "BLOCK_SIZE_M": 16,
112
+ "BLOCK_SIZE_N": 64,
113
+ "BLOCK_SIZE_K": 256,
114
+ "GROUP_SIZE_M": 4,
115
+ "num_warps": 4,
116
+ "num_stages": 2,
117
+ "waves_per_eu": 0
118
+ },
119
+ "1024": {
120
+ "BLOCK_SIZE_M": 16,
121
+ "BLOCK_SIZE_N": 128,
122
+ "BLOCK_SIZE_K": 256,
123
+ "GROUP_SIZE_M": 4,
124
+ "num_warps": 8,
125
+ "num_stages": 2,
126
+ "waves_per_eu": 0
127
+ },
128
+ "1536": {
129
+ "BLOCK_SIZE_M": 16,
130
+ "BLOCK_SIZE_N": 64,
131
+ "BLOCK_SIZE_K": 256,
132
+ "GROUP_SIZE_M": 1,
133
+ "num_warps": 2,
134
+ "num_stages": 2,
135
+ "waves_per_eu": 0
136
+ },
137
+ "2048": {
138
+ "BLOCK_SIZE_M": 64,
139
+ "BLOCK_SIZE_N": 128,
140
+ "BLOCK_SIZE_K": 256,
141
+ "GROUP_SIZE_M": 4,
142
+ "num_warps": 8,
143
+ "num_stages": 2,
144
+ "waves_per_eu": 0
145
+ },
146
+ "3072": {
147
+ "BLOCK_SIZE_M": 64,
148
+ "BLOCK_SIZE_N": 128,
149
+ "BLOCK_SIZE_K": 128,
150
+ "GROUP_SIZE_M": 16,
151
+ "num_warps": 8,
152
+ "num_stages": 2,
153
+ "waves_per_eu": 0
154
+ },
155
+ "4096": {
156
+ "BLOCK_SIZE_M": 64,
157
+ "BLOCK_SIZE_N": 256,
158
+ "BLOCK_SIZE_K": 128,
159
+ "GROUP_SIZE_M": 1,
160
+ "num_warps": 8,
161
+ "num_stages": 2,
162
+ "waves_per_eu": 0
163
+ }
164
+}