[Kernel][Perf] Tune fused_moe FP8 config for Qwen3.5 on L40S (+7%) (#53819)

Signed-off-by: nizhang1 <nizhang1@coupang.com> Co-authored-by: nizhang1 <nizhang1@coupang.com> Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>

nicholaskh-ai committed Aug 26, 2026 at 12:02 UTC 044b05220ee9eb42ae3b90b984d0c7e9eb442f0e
1 file changed +171
vllm/model_executor/layers/fused_moe/configs/E=256,N=512,device_name=NVIDIA_L40S,dtype=fp8_w8a8.json new
+171
@@ -0,0 +1,171 @@
1 +{
2 + "triton_version": "3.6.0",
3 + "1": {
4 + "BLOCK_SIZE_M": 16,
5 + "BLOCK_SIZE_N": 64,
6 + "BLOCK_SIZE_K": 128,
7 + "GROUP_SIZE_M": 1,
8 + "num_warps": 4,
9 + "num_stages": 4
10 + },
11 + "2": {
12 + "BLOCK_SIZE_M": 16,
13 + "BLOCK_SIZE_N": 64,
14 + "BLOCK_SIZE_K": 128,
15 + "GROUP_SIZE_M": 1,
16 + "num_warps": 4,
17 + "num_stages": 3
18 + },
19 + "4": {
20 + "BLOCK_SIZE_M": 16,
21 + "BLOCK_SIZE_N": 128,
22 + "BLOCK_SIZE_K": 256,
23 + "GROUP_SIZE_M": 16,
24 + "num_warps": 8,
25 + "num_stages": 3
26 + },
27 + "8": {
28 + "BLOCK_SIZE_M": 32,
29 + "BLOCK_SIZE_N": 32,
30 + "BLOCK_SIZE_K": 256,
31 + "GROUP_SIZE_M": 16,
32 + "num_warps": 4,
33 + "num_stages": 4
34 + },
35 + "16": {
36 + "BLOCK_SIZE_M": 16,
37 + "BLOCK_SIZE_N": 32,
38 + "BLOCK_SIZE_K": 256,
39 + "GROUP_SIZE_M": 64,
40 + "num_warps": 4,
41 + "num_stages": 4
42 + },
43 + "24": {
44 + "BLOCK_SIZE_M": 16,
45 + "BLOCK_SIZE_N": 32,
46 + "BLOCK_SIZE_K": 256,
47 + "GROUP_SIZE_M": 1,
48 + "num_warps": 4,
49 + "num_stages": 4
50 + },
51 + "32": {
52 + "BLOCK_SIZE_M": 32,
53 + "BLOCK_SIZE_N": 32,
54 + "BLOCK_SIZE_K": 256,
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": 64,
62 + "BLOCK_SIZE_K": 256,
63 + "GROUP_SIZE_M": 1,
64 + "num_warps": 4,
65 + "num_stages": 5
66 + },
67 + "64": {
68 + "BLOCK_SIZE_M": 16,
69 + "BLOCK_SIZE_N": 32,
70 + "BLOCK_SIZE_K": 256,
71 + "GROUP_SIZE_M": 1,
72 + "num_warps": 4,
73 + "num_stages": 5
74 + },
75 + "96": {
76 + "BLOCK_SIZE_M": 16,
77 + "BLOCK_SIZE_N": 32,
78 + "BLOCK_SIZE_K": 256,
79 + "GROUP_SIZE_M": 1,
80 + "num_warps": 4,
81 + "num_stages": 5
82 + },
83 + "128": {
84 + "BLOCK_SIZE_M": 32,
85 + "BLOCK_SIZE_N": 64,
86 + "BLOCK_SIZE_K": 256,
87 + "GROUP_SIZE_M": 1,
88 + "num_warps": 8,
89 + "num_stages": 5
90 + },
91 + "256": {
92 + "BLOCK_SIZE_M": 16,
93 + "BLOCK_SIZE_N": 32,
94 + "BLOCK_SIZE_K": 256,
95 + "GROUP_SIZE_M": 16,
96 + "num_warps": 4,
97 + "num_stages": 5
98 + },
99 + "512": {
100 + "BLOCK_SIZE_M": 16,
101 + "BLOCK_SIZE_N": 64,
102 + "BLOCK_SIZE_K": 256,
103 + "GROUP_SIZE_M": 16,
104 + "num_warps": 4,
105 + "num_stages": 5
106 + },
107 + "1024": {
108 + "BLOCK_SIZE_M": 16,
109 + "BLOCK_SIZE_N": 64,
110 + "BLOCK_SIZE_K": 256,
111 + "GROUP_SIZE_M": 64,
112 + "num_warps": 4,
113 + "num_stages": 5
114 + },
115 + "1536": {
116 + "BLOCK_SIZE_M": 32,
117 + "BLOCK_SIZE_N": 64,
118 + "BLOCK_SIZE_K": 256,
119 + "GROUP_SIZE_M": 32,
120 + "num_warps": 8,
121 + "num_stages": 5
122 + },
123 + "2048": {
124 + "BLOCK_SIZE_M": 32,
125 + "BLOCK_SIZE_N": 128,
126 + "BLOCK_SIZE_K": 256,
127 + "GROUP_SIZE_M": 64,
128 + "num_warps": 8,
129 + "num_stages": 3
130 + },
131 + "3072": {
132 + "BLOCK_SIZE_M": 32,
133 + "BLOCK_SIZE_N": 128,
134 + "BLOCK_SIZE_K": 256,
135 + "GROUP_SIZE_M": 64,
136 + "num_warps": 8,
137 + "num_stages": 3
138 + },
139 + "4096": {
140 + "BLOCK_SIZE_M": 64,
141 + "BLOCK_SIZE_N": 128,
142 + "BLOCK_SIZE_K": 256,
143 + "GROUP_SIZE_M": 16,
144 + "num_warps": 4,
145 + "num_stages": 3
146 + },
147 + "8192": {
148 + "BLOCK_SIZE_M": 128,
149 + "BLOCK_SIZE_N": 128,
150 + "BLOCK_SIZE_K": 256,
151 + "GROUP_SIZE_M": 1,
152 + "num_warps": 8,
153 + "num_stages": 2
154 + },
155 + "16384": {
156 + "BLOCK_SIZE_M": 128,
157 + "BLOCK_SIZE_N": 128,
158 + "BLOCK_SIZE_K": 256,
159 + "GROUP_SIZE_M": 16,
160 + "num_warps": 8,
161 + "num_stages": 2
162 + },
163 + "32768": {
164 + "BLOCK_SIZE_M": 256,
165 + "BLOCK_SIZE_N": 128,
166 + "BLOCK_SIZE_K": 128,
167 + "GROUP_SIZE_M": 1,
168 + "num_warps": 8,
169 + "num_stages": 2
170 + }
171 +}
\ No newline at end of file