Skip to content

Commit 0b8229e

Browse files
shihaobaishihaobai
committed
235B H20 configs
1 parent 7708b30 commit 0b8229e

6 files changed

+702
-0
lines changed

‎lightllm/common/triton_utils/autotune_kernel_configs/triton_3.4.0/NVIDIA_H20/grouped_matmul:v1/{K=192,N=4096,expert_num=128,mul_routed_weight=true,out_dtype=torch.bfloat16,topk_num=1,use_fp8_w8a8=false}_NVIDIA_H20.json‎

Lines changed: 110 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,110 @@
1+
{
2+
"1024": {
3+
"BLOCK_SIZE_K": 32,
4+
"BLOCK_SIZE_M": 16,
5+
"BLOCK_SIZE_N": 128,
6+
"GROUP_SIZE_M": 1,
7+
"NEED_TRANS": false,
8+
"num_stages": 4,
9+
"num_warps": 4
10+
},
11+
"128": {
12+
"BLOCK_SIZE_K": 32,
13+
"BLOCK_SIZE_M": 16,
14+
"BLOCK_SIZE_N": 128,
15+
"GROUP_SIZE_M": 1,
16+
"NEED_TRANS": false,
17+
"num_stages": 4,
18+
"num_warps": 4
19+
},
20+
"16384": {
21+
"BLOCK_SIZE_K": 32,
22+
"BLOCK_SIZE_M": 64,
23+
"BLOCK_SIZE_N": 128,
24+
"GROUP_SIZE_M": 16,
25+
"NEED_TRANS": false,
26+
"num_stages": 3,
27+
"num_warps": 4
28+
},
29+
"2048": {
30+
"BLOCK_SIZE_K": 32,
31+
"BLOCK_SIZE_M": 16,
32+
"BLOCK_SIZE_N": 128,
33+
"GROUP_SIZE_M": 1,
34+
"NEED_TRANS": false,
35+
"num_stages": 4,
36+
"num_warps": 4
37+
},
38+
"256": {
39+
"BLOCK_SIZE_K": 32,
40+
"BLOCK_SIZE_M": 16,
41+
"BLOCK_SIZE_N": 128,
42+
"GROUP_SIZE_M": 1,
43+
"NEED_TRANS": false,
44+
"num_stages": 4,
45+
"num_warps": 4
46+
},
47+
"32768": {
48+
"BLOCK_SIZE_K": 32,
49+
"BLOCK_SIZE_M": 64,
50+
"BLOCK_SIZE_N": 128,
51+
"GROUP_SIZE_M": 16,
52+
"NEED_TRANS": false,
53+
"num_stages": 3,
54+
"num_warps": 4
55+
},
56+
"512": {
57+
"BLOCK_SIZE_K": 32,
58+
"BLOCK_SIZE_M": 16,
59+
"BLOCK_SIZE_N": 128,
60+
"GROUP_SIZE_M": 1,
61+
"NEED_TRANS": false,
62+
"num_stages": 4,
63+
"num_warps": 4
64+
},
65+
"64": {
66+
"BLOCK_SIZE_K": 32,
67+
"BLOCK_SIZE_M": 16,
68+
"BLOCK_SIZE_N": 128,
69+
"GROUP_SIZE_M": 1,
70+
"NEED_TRANS": false,
71+
"num_stages": 3,
72+
"num_warps": 4
73+
},
74+
"65536": {
75+
"BLOCK_SIZE_K": 32,
76+
"BLOCK_SIZE_M": 64,
77+
"BLOCK_SIZE_N": 128,
78+
"GROUP_SIZE_M": 32,
79+
"NEED_TRANS": false,
80+
"num_stages": 3,
81+
"num_warps": 4
82+
},
83+
"8": {
84+
"BLOCK_SIZE_K": 32,
85+
"BLOCK_SIZE_M": 16,
86+
"BLOCK_SIZE_N": 128,
87+
"GROUP_SIZE_M": 1,
88+
"NEED_TRANS": false,
89+
"num_stages": 3,
90+
"num_warps": 4
91+
},
92+
"800": {
93+
"BLOCK_SIZE_K": 32,
94+
"BLOCK_SIZE_M": 16,
95+
"BLOCK_SIZE_N": 128,
96+
"GROUP_SIZE_M": 1,
97+
"NEED_TRANS": false,
98+
"num_stages": 4,
99+
"num_warps": 4
100+
},
101+
"8192": {
102+
"BLOCK_SIZE_K": 32,
103+
"BLOCK_SIZE_M": 64,
104+
"BLOCK_SIZE_N": 128,
105+
"GROUP_SIZE_M": 16,
106+
"NEED_TRANS": false,
107+
"num_stages": 3,
108+
"num_warps": 4
109+
}
110+
}

‎lightllm/common/triton_utils/autotune_kernel_configs/triton_3.4.0/NVIDIA_H20/grouped_matmul:v1/{K=4096,N=384,expert_num=128,mul_routed_weight=false,out_dtype=torch.bfloat16,topk_num=8,use_fp8_w8a8=false}_NVIDIA_H20.json‎

Lines changed: 110 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,110 @@
1+
{
2+
"1": {
3+
"BLOCK_SIZE_K": 128,
4+
"BLOCK_SIZE_M": 16,
5+
"BLOCK_SIZE_N": 64,
6+
"GROUP_SIZE_M": 1,
7+
"NEED_TRANS": false,
8+
"num_stages": 4,
9+
"num_warps": 4
10+
},
11+
"100": {
12+
"BLOCK_SIZE_K": 128,
13+
"BLOCK_SIZE_M": 16,
14+
"BLOCK_SIZE_N": 64,
15+
"GROUP_SIZE_M": 1,
16+
"NEED_TRANS": false,
17+
"num_stages": 3,
18+
"num_warps": 4
19+
},
20+
"1024": {
21+
"BLOCK_SIZE_K": 64,
22+
"BLOCK_SIZE_M": 64,
23+
"BLOCK_SIZE_N": 64,
24+
"GROUP_SIZE_M": 64,
25+
"NEED_TRANS": false,
26+
"num_stages": 4,
27+
"num_warps": 4
28+
},
29+
"128": {
30+
"BLOCK_SIZE_K": 128,
31+
"BLOCK_SIZE_M": 16,
32+
"BLOCK_SIZE_N": 64,
33+
"GROUP_SIZE_M": 1,
34+
"NEED_TRANS": false,
35+
"num_stages": 3,
36+
"num_warps": 4
37+
},
38+
"16": {
39+
"BLOCK_SIZE_K": 128,
40+
"BLOCK_SIZE_M": 16,
41+
"BLOCK_SIZE_N": 32,
42+
"GROUP_SIZE_M": 1,
43+
"NEED_TRANS": false,
44+
"num_stages": 3,
45+
"num_warps": 4
46+
},
47+
"2048": {
48+
"BLOCK_SIZE_K": 64,
49+
"BLOCK_SIZE_M": 64,
50+
"BLOCK_SIZE_N": 64,
51+
"GROUP_SIZE_M": 64,
52+
"NEED_TRANS": false,
53+
"num_stages": 4,
54+
"num_warps": 4
55+
},
56+
"256": {
57+
"BLOCK_SIZE_K": 128,
58+
"BLOCK_SIZE_M": 16,
59+
"BLOCK_SIZE_N": 64,
60+
"GROUP_SIZE_M": 64,
61+
"NEED_TRANS": false,
62+
"num_stages": 3,
63+
"num_warps": 4
64+
},
65+
"32": {
66+
"BLOCK_SIZE_K": 128,
67+
"BLOCK_SIZE_M": 16,
68+
"BLOCK_SIZE_N": 64,
69+
"GROUP_SIZE_M": 1,
70+
"NEED_TRANS": false,
71+
"num_stages": 4,
72+
"num_warps": 4
73+
},
74+
"4096": {
75+
"BLOCK_SIZE_K": 64,
76+
"BLOCK_SIZE_M": 64,
77+
"BLOCK_SIZE_N": 64,
78+
"GROUP_SIZE_M": 64,
79+
"NEED_TRANS": false,
80+
"num_stages": 4,
81+
"num_warps": 4
82+
},
83+
"64": {
84+
"BLOCK_SIZE_K": 128,
85+
"BLOCK_SIZE_M": 16,
86+
"BLOCK_SIZE_N": 64,
87+
"GROUP_SIZE_M": 1,
88+
"NEED_TRANS": false,
89+
"num_stages": 3,
90+
"num_warps": 4
91+
},
92+
"8": {
93+
"BLOCK_SIZE_K": 128,
94+
"BLOCK_SIZE_M": 16,
95+
"BLOCK_SIZE_N": 64,
96+
"GROUP_SIZE_M": 32,
97+
"NEED_TRANS": false,
98+
"num_stages": 3,
99+
"num_warps": 4
100+
},
101+
"8192": {
102+
"BLOCK_SIZE_K": 64,
103+
"BLOCK_SIZE_M": 64,
104+
"BLOCK_SIZE_N": 64,
105+
"GROUP_SIZE_M": 64,
106+
"NEED_TRANS": false,
107+
"num_stages": 4,
108+
"num_warps": 4
109+
}
110+
}

‎lightllm/common/triton_utils/autotune_kernel_configs/triton_3.4.0/NVIDIA_H20/moe_align_fused:v1/{topk_num=8}_NVIDIA_H20.json‎

Lines changed: 50 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,50 @@
1+
{
2+
"1": {
3+
"BLOCK_SIZE": 256,
4+
"num_warps": 8
5+
},
6+
"100": {
7+
"BLOCK_SIZE": 128,
8+
"num_warps": 8
9+
},
10+
"1024": {
11+
"BLOCK_SIZE": 128,
12+
"num_warps": 8
13+
},
14+
"128": {
15+
"BLOCK_SIZE": 128,
16+
"num_warps": 8
17+
},
18+
"16": {
19+
"BLOCK_SIZE": 256,
20+
"num_warps": 8
21+
},
22+
"2048": {
23+
"BLOCK_SIZE": 128,
24+
"num_warps": 4
25+
},
26+
"256": {
27+
"BLOCK_SIZE": 128,
28+
"num_warps": 8
29+
},
30+
"32": {
31+
"BLOCK_SIZE": 512,
32+
"num_warps": 8
33+
},
34+
"4096": {
35+
"BLOCK_SIZE": 128,
36+
"num_warps": 8
37+
},
38+
"64": {
39+
"BLOCK_SIZE": 128,
40+
"num_warps": 8
41+
},
42+
"8": {
43+
"BLOCK_SIZE": 256,
44+
"num_warps": 8
45+
},
46+
"8192": {
47+
"BLOCK_SIZE": 128,
48+
"num_warps": 8
49+
}
50+
}

‎lightllm/common/triton_utils/autotune_kernel_configs/triton_3.4.0/NVIDIA_H20/moe_sum_reduce:v1/{hidden_dim=4096,out_dtype=torch.bfloat16,topk_num=8}_NVIDIA_H20.json‎

Lines changed: 74 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,74 @@
1+
{
2+
"1": {
3+
"BLOCK_DIM": 64,
4+
"BLOCK_M": 4,
5+
"NUM_STAGE": 2,
6+
"num_warps": 8
7+
},
8+
"100": {
9+
"BLOCK_DIM": 1024,
10+
"BLOCK_M": 1,
11+
"NUM_STAGE": 1,
12+
"num_warps": 8
13+
},
14+
"1024": {
15+
"BLOCK_DIM": 256,
16+
"BLOCK_M": 1,
17+
"NUM_STAGE": 1,
18+
"num_warps": 2
19+
},
20+
"128": {
21+
"BLOCK_DIM": 1024,
22+
"BLOCK_M": 1,
23+
"NUM_STAGE": 1,
24+
"num_warps": 8
25+
},
26+
"16": {
27+
"BLOCK_DIM": 512,
28+
"BLOCK_M": 1,
29+
"NUM_STAGE": 1,
30+
"num_warps": 8
31+
},
32+
"2048": {
33+
"BLOCK_DIM": 1024,
34+
"BLOCK_M": 1,
35+
"NUM_STAGE": 4,
36+
"num_warps": 2
37+
},
38+
"256": {
39+
"BLOCK_DIM": 512,
40+
"BLOCK_M": 1,
41+
"NUM_STAGE": 1,
42+
"num_warps": 2
43+
},
44+
"32": {
45+
"BLOCK_DIM": 1024,
46+
"BLOCK_M": 1,
47+
"NUM_STAGE": 1,
48+
"num_warps": 8
49+
},
50+
"4096": {
51+
"BLOCK_DIM": 1024,
52+
"BLOCK_M": 1,
53+
"NUM_STAGE": 4,
54+
"num_warps": 2
55+
},
56+
"64": {
57+
"BLOCK_DIM": 1024,
58+
"BLOCK_M": 1,
59+
"NUM_STAGE": 1,
60+
"num_warps": 8
61+
},
62+
"8": {
63+
"BLOCK_DIM": 512,
64+
"BLOCK_M": 1,
65+
"NUM_STAGE": 4,
66+
"num_warps": 16
67+
},
68+
"8192": {
69+
"BLOCK_DIM": 1024,
70+
"BLOCK_M": 1,
71+
"NUM_STAGE": 4,
72+
"num_warps": 1
73+
}
74+
}

0 commit comments

Comments
 (0)