[Qwen3.8-Flash-Next] Tune FP8 TP2/TP4 Triton MoE on B200 (#55890)
Signed-off-by: Canlin <canlinguosdu@gmail.com> Co-authored-by: OpenAI Codex <noreply@openai.com>
Canlin Guo committed
Sep 8, 2026 at 22:45 UTC
8e1f97e70984192cc63c51a8af3e313cab8c5c73
2 files changed
+230
vllm/model_executor/layers/fused_moe/configs/E=512,N=160,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[32,32].json
new
+115
@@ -0,0 +1,115 @@
1
+{
2
+ "triton_version": "3.7.1",
3
+ "1": {
4
+ "BLOCK_SIZE_M": 16,
5
+ "BLOCK_SIZE_N": 32,
6
+ "BLOCK_SIZE_K": 32,
7
+ "GROUP_SIZE_M": 1,
8
+ "num_warps": 4,
9
+ "num_stages": 6
10
+ },
11
+ "2": {
12
+ "BLOCK_SIZE_M": 16,
13
+ "BLOCK_SIZE_N": 32,
14
+ "BLOCK_SIZE_K": 32,
15
+ "GROUP_SIZE_M": 1,
16
+ "num_warps": 2,
17
+ "num_stages": 5
18
+ },
19
+ "4": {
20
+ "BLOCK_SIZE_M": 16,
21
+ "BLOCK_SIZE_N": 64,
22
+ "BLOCK_SIZE_K": 32,
23
+ "GROUP_SIZE_M": 1,
24
+ "num_warps": 2,
25
+ "num_stages": 5
26
+ },
27
+ "8": {
28
+ "BLOCK_SIZE_M": 16,
29
+ "BLOCK_SIZE_N": 64,
30
+ "BLOCK_SIZE_K": 32,
31
+ "GROUP_SIZE_M": 1,
32
+ "num_warps": 2,
33
+ "num_stages": 5
34
+ },
35
+ "16": {
36
+ "BLOCK_SIZE_M": 16,
37
+ "BLOCK_SIZE_N": 128,
38
+ "BLOCK_SIZE_K": 32,
39
+ "GROUP_SIZE_M": 1,
40
+ "num_warps": 4,
41
+ "num_stages": 5
42
+ },
43
+ "32": {
44
+ "BLOCK_SIZE_M": 16,
45
+ "BLOCK_SIZE_N": 128,
46
+ "BLOCK_SIZE_K": 32,
47
+ "GROUP_SIZE_M": 1,
48
+ "num_warps": 2,
49
+ "num_stages": 5
50
+ },
51
+ "64": {
52
+ "BLOCK_SIZE_M": 16,
53
+ "BLOCK_SIZE_N": 128,
54
+ "BLOCK_SIZE_K": 32,
55
+ "GROUP_SIZE_M": 1,
56
+ "num_warps": 2,
57
+ "num_stages": 5
58
+ },
59
+ "128": {
60
+ "BLOCK_SIZE_M": 16,
61
+ "BLOCK_SIZE_N": 128,
62
+ "BLOCK_SIZE_K": 32,
63
+ "GROUP_SIZE_M": 1,
64
+ "num_warps": 4,
65
+ "num_stages": 5
66
+ },
67
+ "256": {
68
+ "BLOCK_SIZE_M": 16,
69
+ "BLOCK_SIZE_N": 128,
70
+ "BLOCK_SIZE_K": 32,
71
+ "GROUP_SIZE_M": 1,
72
+ "num_warps": 4,
73
+ "num_stages": 5
74
+ },
75
+ "512": {
76
+ "BLOCK_SIZE_M": 16,
77
+ "BLOCK_SIZE_N": 128,
78
+ "BLOCK_SIZE_K": 32,
79
+ "GROUP_SIZE_M": 1,
80
+ "num_warps": 4,
81
+ "num_stages": 5
82
+ },
83
+ "1024": {
84
+ "BLOCK_SIZE_M": 32,
85
+ "BLOCK_SIZE_N": 128,
86
+ "BLOCK_SIZE_K": 32,
87
+ "GROUP_SIZE_M": 32,
88
+ "num_warps": 4,
89
+ "num_stages": 4
90
+ },
91
+ "2048": {
92
+ "BLOCK_SIZE_M": 32,
93
+ "BLOCK_SIZE_N": 128,
94
+ "BLOCK_SIZE_K": 32,
95
+ "GROUP_SIZE_M": 8,
96
+ "num_warps": 2,
97
+ "num_stages": 4
98
+ },
99
+ "4096": {
100
+ "BLOCK_SIZE_M": 32,
101
+ "BLOCK_SIZE_N": 128,
102
+ "BLOCK_SIZE_K": 32,
103
+ "GROUP_SIZE_M": 8,
104
+ "num_warps": 2,
105
+ "num_stages": 4
106
+ },
107
+ "8192": {
108
+ "BLOCK_SIZE_M": 32,
109
+ "BLOCK_SIZE_N": 128,
110
+ "BLOCK_SIZE_K": 32,
111
+ "GROUP_SIZE_M": 8,
112
+ "num_warps": 2,
113
+ "num_stages": 4
114
+ }
115
+}
vllm/model_executor/layers/fused_moe/configs/E=512,N=320,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[64,64].json
new
+115
@@ -0,0 +1,115 @@
1
+{
2
+ "triton_version": "3.7.1",
3
+ "1": {
4
+ "BLOCK_SIZE_M": 16,
5
+ "BLOCK_SIZE_N": 32,
6
+ "BLOCK_SIZE_K": 64,
7
+ "GROUP_SIZE_M": 1,
8
+ "num_warps": 2,
9
+ "num_stages": 5
10
+ },
11
+ "2": {
12
+ "BLOCK_SIZE_M": 16,
13
+ "BLOCK_SIZE_N": 32,
14
+ "BLOCK_SIZE_K": 64,
15
+ "GROUP_SIZE_M": 1,
16
+ "num_warps": 2,
17
+ "num_stages": 5
18
+ },
19
+ "4": {
20
+ "BLOCK_SIZE_M": 16,
21
+ "BLOCK_SIZE_N": 32,
22
+ "BLOCK_SIZE_K": 64,
23
+ "GROUP_SIZE_M": 1,
24
+ "num_warps": 2,
25
+ "num_stages": 5
26
+ },
27
+ "8": {
28
+ "BLOCK_SIZE_M": 16,
29
+ "BLOCK_SIZE_N": 64,
30
+ "BLOCK_SIZE_K": 64,
31
+ "GROUP_SIZE_M": 1,
32
+ "num_warps": 2,
33
+ "num_stages": 5
34
+ },
35
+ "16": {
36
+ "BLOCK_SIZE_M": 16,
37
+ "BLOCK_SIZE_N": 128,
38
+ "BLOCK_SIZE_K": 64,
39
+ "GROUP_SIZE_M": 8,
40
+ "num_warps": 4,
41
+ "num_stages": 3
42
+ },
43
+ "32": {
44
+ "BLOCK_SIZE_M": 16,
45
+ "BLOCK_SIZE_N": 128,
46
+ "BLOCK_SIZE_K": 64,
47
+ "GROUP_SIZE_M": 1,
48
+ "num_warps": 4,
49
+ "num_stages": 4
50
+ },
51
+ "64": {
52
+ "BLOCK_SIZE_M": 16,
53
+ "BLOCK_SIZE_N": 128,
54
+ "BLOCK_SIZE_K": 64,
55
+ "GROUP_SIZE_M": 1,
56
+ "num_warps": 4,
57
+ "num_stages": 5
58
+ },
59
+ "128": {
60
+ "BLOCK_SIZE_M": 16,
61
+ "BLOCK_SIZE_N": 128,
62
+ "BLOCK_SIZE_K": 64,
63
+ "GROUP_SIZE_M": 1,
64
+ "num_warps": 4,
65
+ "num_stages": 5
66
+ },
67
+ "256": {
68
+ "BLOCK_SIZE_M": 16,
69
+ "BLOCK_SIZE_N": 128,
70
+ "BLOCK_SIZE_K": 64,
71
+ "GROUP_SIZE_M": 1,
72
+ "num_warps": 4,
73
+ "num_stages": 5
74
+ },
75
+ "512": {
76
+ "BLOCK_SIZE_M": 16,
77
+ "BLOCK_SIZE_N": 128,
78
+ "BLOCK_SIZE_K": 64,
79
+ "GROUP_SIZE_M": 1,
80
+ "num_warps": 2,
81
+ "num_stages": 3
82
+ },
83
+ "1024": {
84
+ "BLOCK_SIZE_M": 32,
85
+ "BLOCK_SIZE_N": 128,
86
+ "BLOCK_SIZE_K": 64,
87
+ "GROUP_SIZE_M": 8,
88
+ "num_warps": 4,
89
+ "num_stages": 5
90
+ },
91
+ "2048": {
92
+ "BLOCK_SIZE_M": 32,
93
+ "BLOCK_SIZE_N": 128,
94
+ "BLOCK_SIZE_K": 64,
95
+ "GROUP_SIZE_M": 32,
96
+ "num_warps": 4,
97
+ "num_stages": 5
98
+ },
99
+ "4096": {
100
+ "BLOCK_SIZE_M": 32,
101
+ "BLOCK_SIZE_N": 128,
102
+ "BLOCK_SIZE_K": 64,
103
+ "GROUP_SIZE_M": 32,
104
+ "num_warps": 4,
105
+ "num_stages": 5
106
+ },
107
+ "8192": {
108
+ "BLOCK_SIZE_M": 64,
109
+ "BLOCK_SIZE_N": 128,
110
+ "BLOCK_SIZE_K": 64,
111
+ "GROUP_SIZE_M": 32,
112
+ "num_warps": 4,
113
+ "num_stages": 6
114
+ }
115
+}