[Kernel] Add fused MoE tuned config for E=256,N=512 on NVIDIA A100 80GB PCIe (#55511)

Signed-off-by: Baki Burak Ogun <63836730+bakiburakogun@users.noreply.github.com>

Baki Burak Öğün committed Sep 6, 2026 at 06:15 UTC f2e2936f91a794513af74d009286c69e3fb2592b
1 file changed +147
vllm/model_executor/layers/fused_moe/configs/E=256,N=512,device_name=NVIDIA_A100_80GB_PCIe.json new
+147
@@ -0,0 +1,147 @@
1 +{
2 + "triton_version": "3.7.1",
3 + "1": {
4 + "BLOCK_SIZE_M": 16,
5 + "BLOCK_SIZE_N": 128,
6 + "BLOCK_SIZE_K": 64,
7 + "GROUP_SIZE_M": 64,
8 + "num_warps": 8,
9 + "num_stages": 5
10 + },
11 + "2": {
12 + "BLOCK_SIZE_M": 16,
13 + "BLOCK_SIZE_N": 64,
14 + "BLOCK_SIZE_K": 64,
15 + "GROUP_SIZE_M": 16,
16 + "num_warps": 4,
17 + "num_stages": 4
18 + },
19 + "4": {
20 + "BLOCK_SIZE_M": 16,
21 + "BLOCK_SIZE_N": 64,
22 + "BLOCK_SIZE_K": 64,
23 + "GROUP_SIZE_M": 32,
24 + "num_warps": 4,
25 + "num_stages": 3
26 + },
27 + "8": {
28 + "BLOCK_SIZE_M": 16,
29 + "BLOCK_SIZE_N": 64,
30 + "BLOCK_SIZE_K": 256,
31 + "GROUP_SIZE_M": 64,
32 + "num_warps": 8,
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": 16,
40 + "num_warps": 8,
41 + "num_stages": 3
42 + },
43 + "24": {
44 + "BLOCK_SIZE_M": 16,
45 + "BLOCK_SIZE_N": 64,
46 + "BLOCK_SIZE_K": 128,
47 + "GROUP_SIZE_M": 32,
48 + "num_warps": 8,
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": 8,
57 + "num_stages": 2
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": 8,
65 + "num_stages": 2
66 + },
67 + "64": {
68 + "BLOCK_SIZE_M": 16,
69 + "BLOCK_SIZE_N": 64,
70 + "BLOCK_SIZE_K": 128,
71 + "GROUP_SIZE_M": 1,
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": 32,
80 + "num_warps": 8,
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": 32,
88 + "num_warps": 8,
89 + "num_stages": 2
90 + },
91 + "256": {
92 + "BLOCK_SIZE_M": 16,
93 + "BLOCK_SIZE_N": 128,
94 + "BLOCK_SIZE_K": 128,
95 + "GROUP_SIZE_M": 1,
96 + "num_warps": 8,
97 + "num_stages": 2
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": 2
106 + },
107 + "1024": {
108 + "BLOCK_SIZE_M": 64,
109 + "BLOCK_SIZE_N": 256,
110 + "BLOCK_SIZE_K": 128,
111 + "GROUP_SIZE_M": 16,
112 + "num_warps": 8,
113 + "num_stages": 3
114 + },
115 + "1536": {
116 + "BLOCK_SIZE_M": 64,
117 + "BLOCK_SIZE_N": 256,
118 + "BLOCK_SIZE_K": 128,
119 + "GROUP_SIZE_M": 16,
120 + "num_warps": 8,
121 + "num_stages": 3
122 + },
123 + "2048": {
124 + "BLOCK_SIZE_M": 64,
125 + "BLOCK_SIZE_N": 256,
126 + "BLOCK_SIZE_K": 64,
127 + "GROUP_SIZE_M": 1,
128 + "num_warps": 4,
129 + "num_stages": 3
130 + },
131 + "3072": {
132 + "BLOCK_SIZE_M": 128,
133 + "BLOCK_SIZE_N": 128,
134 + "BLOCK_SIZE_K": 64,
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": 256,
142 + "BLOCK_SIZE_K": 64,
143 + "GROUP_SIZE_M": 1,
144 + "num_warps": 4,
145 + "num_stages": 3
146 + }
147 +}