Skip to content

Commit ba41cc9

Browse files
mgoinsimon-mo
mgoin
authored and
simon-mo
committed
[Model] Add tuned triton fused_moe configs for Qwen3Moe (#17328)
Signed-off-by: mgoin <mgoin64@gmail.com>
1 parent dcbac4c commit ba41cc9

15 files changed

+1904
-5
lines changed

‎benchmarks/kernels/benchmark_moe.py

Lines changed: 5 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -527,7 +527,7 @@ def get_weight_block_size_safety(config, default_value=None):
527527

528528
def main(args: argparse.Namespace):
529529
print(args)
530-
block_quant_shape = None
530+
531531
config = AutoConfig.from_pretrained(
532532
args.model, trust_remote_code=args.trust_remote_code)
533533
if config.architectures[0] == "DbrxForCausalLM":
@@ -546,8 +546,9 @@ def main(args: argparse.Namespace):
546546
topk = config.num_experts_per_tok
547547
intermediate_size = config.moe_intermediate_size
548548
shard_intermediate_size = 2 * intermediate_size // args.tp_size
549-
block_quant_shape = get_weight_block_size_safety(config)
550-
elif config.architectures[0] == "Qwen2MoeForCausalLM":
549+
elif config.architectures[0] in [
550+
"Qwen2MoeForCausalLM", "Qwen3MoeForCausalLM"
551+
]:
551552
E = config.num_experts
552553
topk = config.num_experts_per_tok
553554
intermediate_size = config.moe_intermediate_size
@@ -565,6 +566,7 @@ def main(args: argparse.Namespace):
565566
dtype = torch.float16 if current_platform.is_rocm() else config.torch_dtype
566567
use_fp8_w8a8 = args.dtype == "fp8_w8a8"
567568
use_int8_w8a16 = args.dtype == "int8_w8a16"
569+
block_quant_shape = get_weight_block_size_safety(config)
568570

569571
if args.batch_size is None:
570572
batch_sizes = [

‎vllm/model_executor/layers/fused_moe/configs/E=128,N=192,device_name=NVIDIA_H100_80GB_HBM3.json

Lines changed: 146 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,146 @@
1+
{
2+
"1": {
3+
"BLOCK_SIZE_M": 16,
4+
"BLOCK_SIZE_N": 64,
5+
"BLOCK_SIZE_K": 64,
6+
"GROUP_SIZE_M": 1,
7+
"num_warps": 4,
8+
"num_stages": 5
9+
},
10+
"2": {
11+
"BLOCK_SIZE_M": 16,
12+
"BLOCK_SIZE_N": 64,
13+
"BLOCK_SIZE_K": 64,
14+
"GROUP_SIZE_M": 64,
15+
"num_warps": 4,
16+
"num_stages": 5
17+
},
18+
"4": {
19+
"BLOCK_SIZE_M": 16,
20+
"BLOCK_SIZE_N": 64,
21+
"BLOCK_SIZE_K": 64,
22+
"GROUP_SIZE_M": 32,
23+
"num_warps": 4,
24+
"num_stages": 5
25+
},
26+
"8": {
27+
"BLOCK_SIZE_M": 16,
28+
"BLOCK_SIZE_N": 128,
29+
"BLOCK_SIZE_K": 64,
30+
"GROUP_SIZE_M": 16,
31+
"num_warps": 4,
32+
"num_stages": 5
33+
},
34+
"16": {
35+
"BLOCK_SIZE_M": 16,
36+
"BLOCK_SIZE_N": 128,
37+
"BLOCK_SIZE_K": 64,
38+
"GROUP_SIZE_M": 1,
39+
"num_warps": 8,
40+
"num_stages": 3
41+
},
42+
"24": {
43+
"BLOCK_SIZE_M": 16,
44+
"BLOCK_SIZE_N": 128,
45+
"BLOCK_SIZE_K": 64,
46+
"GROUP_SIZE_M": 1,
47+
"num_warps": 8,
48+
"num_stages": 3
49+
},
50+
"32": {
51+
"BLOCK_SIZE_M": 16,
52+
"BLOCK_SIZE_N": 128,
53+
"BLOCK_SIZE_K": 128,
54+
"GROUP_SIZE_M": 16,
55+
"num_warps": 4,
56+
"num_stages": 3
57+
},
58+
"48": {
59+
"BLOCK_SIZE_M": 16,
60+
"BLOCK_SIZE_N": 128,
61+
"BLOCK_SIZE_K": 64,
62+
"GROUP_SIZE_M": 64,
63+
"num_warps": 4,
64+
"num_stages": 4
65+
},
66+
"64": {
67+
"BLOCK_SIZE_M": 16,
68+
"BLOCK_SIZE_N": 64,
69+
"BLOCK_SIZE_K": 64,
70+
"GROUP_SIZE_M": 32,
71+
"num_warps": 4,
72+
"num_stages": 3
73+
},
74+
"96": {
75+
"BLOCK_SIZE_M": 16,
76+
"BLOCK_SIZE_N": 128,
77+
"BLOCK_SIZE_K": 128,
78+
"GROUP_SIZE_M": 16,
79+
"num_warps": 4,
80+
"num_stages": 2
81+
},
82+
"128": {
83+
"BLOCK_SIZE_M": 16,
84+
"BLOCK_SIZE_N": 128,
85+
"BLOCK_SIZE_K": 128,
86+
"GROUP_SIZE_M": 32,
87+
"num_warps": 4,
88+
"num_stages": 3
89+
},
90+
"256": {
91+
"BLOCK_SIZE_M": 32,
92+
"BLOCK_SIZE_N": 128,
93+
"BLOCK_SIZE_K": 128,
94+
"GROUP_SIZE_M": 32,
95+
"num_warps": 4,
96+
"num_stages": 2
97+
},
98+
"512": {
99+
"BLOCK_SIZE_M": 64,
100+
"BLOCK_SIZE_N": 128,
101+
"BLOCK_SIZE_K": 64,
102+
"GROUP_SIZE_M": 16,
103+
"num_warps": 4,
104+
"num_stages": 3
105+
},
106+
"1024": {
107+
"BLOCK_SIZE_M": 64,
108+
"BLOCK_SIZE_N": 128,
109+
"BLOCK_SIZE_K": 64,
110+
"GROUP_SIZE_M": 32,
111+
"num_warps": 4,
112+
"num_stages": 3
113+
},
114+
"1536": {
115+
"BLOCK_SIZE_M": 128,
116+
"BLOCK_SIZE_N": 256,
117+
"BLOCK_SIZE_K": 64,
118+
"GROUP_SIZE_M": 16,
119+
"num_warps": 8,
120+
"num_stages": 4
121+
},
122+
"2048": {
123+
"BLOCK_SIZE_M": 128,
124+
"BLOCK_SIZE_N": 256,
125+
"BLOCK_SIZE_K": 64,
126+
"GROUP_SIZE_M": 16,
127+
"num_warps": 8,
128+
"num_stages": 4
129+
},
130+
"3072": {
131+
"BLOCK_SIZE_M": 128,
132+
"BLOCK_SIZE_N": 128,
133+
"BLOCK_SIZE_K": 64,
134+
"GROUP_SIZE_M": 16,
135+
"num_warps": 8,
136+
"num_stages": 3
137+
},
138+
"4096": {
139+
"BLOCK_SIZE_M": 128,
140+
"BLOCK_SIZE_N": 128,
141+
"BLOCK_SIZE_K": 64,
142+
"GROUP_SIZE_M": 16,
143+
"num_warps": 4,
144+
"num_stages": 3
145+
}
146+
}

‎vllm/model_executor/layers/fused_moe/configs/E=128,N=192,device_name=NVIDIA_H20.json

Lines changed: 146 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,146 @@
1+
{
2+
"1": {
3+
"BLOCK_SIZE_M": 16,
4+
"BLOCK_SIZE_N": 64,
5+
"BLOCK_SIZE_K": 64,
6+
"GROUP_SIZE_M": 1,
7+
"num_warps": 4,
8+
"num_stages": 5
9+
},
10+
"2": {
11+
"BLOCK_SIZE_M": 16,
12+
"BLOCK_SIZE_N": 128,
13+
"BLOCK_SIZE_K": 64,
14+
"GROUP_SIZE_M": 16,
15+
"num_warps": 4,
16+
"num_stages": 5
17+
},
18+
"4": {
19+
"BLOCK_SIZE_M": 16,
20+
"BLOCK_SIZE_N": 64,
21+
"BLOCK_SIZE_K": 64,
22+
"GROUP_SIZE_M": 32,
23+
"num_warps": 4,
24+
"num_stages": 4
25+
},
26+
"8": {
27+
"BLOCK_SIZE_M": 16,
28+
"BLOCK_SIZE_N": 128,
29+
"BLOCK_SIZE_K": 64,
30+
"GROUP_SIZE_M": 1,
31+
"num_warps": 4,
32+
"num_stages": 4
33+
},
34+
"16": {
35+
"BLOCK_SIZE_M": 16,
36+
"BLOCK_SIZE_N": 128,
37+
"BLOCK_SIZE_K": 64,
38+
"GROUP_SIZE_M": 16,
39+
"num_warps": 4,
40+
"num_stages": 3
41+
},
42+
"24": {
43+
"BLOCK_SIZE_M": 16,
44+
"BLOCK_SIZE_N": 128,
45+
"BLOCK_SIZE_K": 64,
46+
"GROUP_SIZE_M": 32,
47+
"num_warps": 4,
48+
"num_stages": 3
49+
},
50+
"32": {
51+
"BLOCK_SIZE_M": 16,
52+
"BLOCK_SIZE_N": 128,
53+
"BLOCK_SIZE_K": 64,
54+
"GROUP_SIZE_M": 64,
55+
"num_warps": 4,
56+
"num_stages": 3
57+
},
58+
"48": {
59+
"BLOCK_SIZE_M": 16,
60+
"BLOCK_SIZE_N": 128,
61+
"BLOCK_SIZE_K": 64,
62+
"GROUP_SIZE_M": 32,
63+
"num_warps": 4,
64+
"num_stages": 3
65+
},
66+
"64": {
67+
"BLOCK_SIZE_M": 16,
68+
"BLOCK_SIZE_N": 128,
69+
"BLOCK_SIZE_K": 64,
70+
"GROUP_SIZE_M": 1,
71+
"num_warps": 4,
72+
"num_stages": 3
73+
},
74+
"96": {
75+
"BLOCK_SIZE_M": 16,
76+
"BLOCK_SIZE_N": 128,
77+
"BLOCK_SIZE_K": 64,
78+
"GROUP_SIZE_M": 1,
79+
"num_warps": 4,
80+
"num_stages": 3
81+
},
82+
"128": {
83+
"BLOCK_SIZE_M": 16,
84+
"BLOCK_SIZE_N": 128,
85+
"BLOCK_SIZE_K": 64,
86+
"GROUP_SIZE_M": 1,
87+
"num_warps": 4,
88+
"num_stages": 3
89+
},
90+
"256": {
91+
"BLOCK_SIZE_M": 16,
92+
"BLOCK_SIZE_N": 128,
93+
"BLOCK_SIZE_K": 64,
94+
"GROUP_SIZE_M": 32,
95+
"num_warps": 4,
96+
"num_stages": 3
97+
},
98+
"512": {
99+
"BLOCK_SIZE_M": 64,
100+
"BLOCK_SIZE_N": 128,
101+
"BLOCK_SIZE_K": 64,
102+
"GROUP_SIZE_M": 1,
103+
"num_warps": 8,
104+
"num_stages": 3
105+
},
106+
"1024": {
107+
"BLOCK_SIZE_M": 64,
108+
"BLOCK_SIZE_N": 128,
109+
"BLOCK_SIZE_K": 64,
110+
"GROUP_SIZE_M": 16,
111+
"num_warps": 4,
112+
"num_stages": 3
113+
},
114+
"1536": {
115+
"BLOCK_SIZE_M": 64,
116+
"BLOCK_SIZE_N": 128,
117+
"BLOCK_SIZE_K": 64,
118+
"GROUP_SIZE_M": 32,
119+
"num_warps": 4,
120+
"num_stages": 3
121+
},
122+
"2048": {
123+
"BLOCK_SIZE_M": 64,
124+
"BLOCK_SIZE_N": 128,
125+
"BLOCK_SIZE_K": 64,
126+
"GROUP_SIZE_M": 1,
127+
"num_warps": 4,
128+
"num_stages": 3
129+
},
130+
"3072": {
131+
"BLOCK_SIZE_M": 64,
132+
"BLOCK_SIZE_N": 64,
133+
"BLOCK_SIZE_K": 64,
134+
"GROUP_SIZE_M": 16,
135+
"num_warps": 4,
136+
"num_stages": 2
137+
},
138+
"4096": {
139+
"BLOCK_SIZE_M": 64,
140+
"BLOCK_SIZE_N": 128,
141+
"BLOCK_SIZE_K": 64,
142+
"GROUP_SIZE_M": 1,
143+
"num_warps": 4,
144+
"num_stages": 2
145+
}
146+
}

0 commit comments

Comments
 (0)