A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
Tuned H100/H200 triton fp8 block configs for fused_qkv_a_proj (#23939)
Signed-off-by: mgoin <mgoin64@gmail.com>
Michael Goin committed
Aug 29, 2025 at 13:28 UTC
b7adf94c4a6c7290dd8765819da68a801008f5a1
4 files changed
+271
-3
benchmarks/kernels/bench_block_fp8_gemm.py
+1
@@ -16,6 +16,7 @@ assert current_platform.is_cuda(), (
16
# DeepSeek-V3 weight shapes
17
DEEPSEEK_V3_SHAPES = [
18
(512 + 64, 7168),
19
+ (2112, 7168),
20
((128 + 64) * 128, 7168),
21
(128 * (128 + 128), 512),
22
(7168, 16384),
benchmarks/kernels/benchmark_w8a8_block_fp8.py
+1
@@ -141,6 +141,7 @@ def get_weight_shapes(tp_size):
141
# cannot TP
142
total = [
143
(512 + 64, 7168),
144
+ (2112, 7168),
145
((128 + 64) * 128, 7168),
146
(128 * (128 + 128), 512),
147
(7168, 16384),
vllm/model_executor/layers/quantization/utils/configs/N=2112,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json
new
+146
@@ -0,0 +1,146 @@
1
+{
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 32,
5
+ "BLOCK_SIZE_K": 128,
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": 32,
13
+ "BLOCK_SIZE_K": 128,
14
+ "GROUP_SIZE_M": 1,
15
+ "num_warps": 4,
16
+ "num_stages": 5
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 32,
21
+ "BLOCK_SIZE_K": 128,
22
+ "GROUP_SIZE_M": 64,
23
+ "num_warps": 4,
24
+ "num_stages": 5
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 32,
29
+ "BLOCK_SIZE_K": 128,
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": 32,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 1,
39
+ "num_warps": 4,
40
+ "num_stages": 3
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 32,
45
+ "BLOCK_SIZE_K": 128,
46
+ "GROUP_SIZE_M": 32,
47
+ "num_warps": 4,
48
+ "num_stages": 4
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 16,
52
+ "BLOCK_SIZE_N": 32,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 16,
55
+ "num_warps": 4,
56
+ "num_stages": 5
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 16,
60
+ "BLOCK_SIZE_N": 64,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 16,
63
+ "num_warps": 4,
64
+ "num_stages": 3
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 16,
68
+ "BLOCK_SIZE_N": 64,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 16,
71
+ "num_warps": 4,
72
+ "num_stages": 5
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 64,
76
+ "BLOCK_SIZE_N": 64,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 32,
79
+ "num_warps": 4,
80
+ "num_stages": 5
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 64,
84
+ "BLOCK_SIZE_N": 64,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 16,
87
+ "num_warps": 4,
88
+ "num_stages": 3
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 64,
92
+ "BLOCK_SIZE_N": 64,
93
+ "BLOCK_SIZE_K": 128,
94
+ "GROUP_SIZE_M": 16,
95
+ "num_warps": 4,
96
+ "num_stages": 4
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 128,
100
+ "BLOCK_SIZE_N": 64,
101
+ "BLOCK_SIZE_K": 128,
102
+ "GROUP_SIZE_M": 16,
103
+ "num_warps": 4,
104
+ "num_stages": 5
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 256,
108
+ "BLOCK_SIZE_N": 64,
109
+ "BLOCK_SIZE_K": 128,
110
+ "GROUP_SIZE_M": 1,
111
+ "num_warps": 8,
112
+ "num_stages": 5
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 64,
116
+ "BLOCK_SIZE_N": 128,
117
+ "BLOCK_SIZE_K": 128,
118
+ "GROUP_SIZE_M": 16,
119
+ "num_warps": 4,
120
+ "num_stages": 4
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 128,
124
+ "BLOCK_SIZE_N": 64,
125
+ "BLOCK_SIZE_K": 128,
126
+ "GROUP_SIZE_M": 1,
127
+ "num_warps": 4,
128
+ "num_stages": 4
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 128,
132
+ "BLOCK_SIZE_N": 64,
133
+ "BLOCK_SIZE_K": 128,
134
+ "GROUP_SIZE_M": 1,
135
+ "num_warps": 4,
136
+ "num_stages": 3
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 128,
140
+ "BLOCK_SIZE_N": 64,
141
+ "BLOCK_SIZE_K": 128,
142
+ "GROUP_SIZE_M": 1,
143
+ "num_warps": 4,
144
+ "num_stages": 4
145
+ }
146
+}
vllm/model_executor/layers/quantization/utils/configs/N=2112,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json
+123
-3
@@ -1,10 +1,130 @@
1
{
2
- "2048": {
3
- "BLOCK_SIZE_M": 256,
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 32,
5
+ "BLOCK_SIZE_K": 128,
6
+ "GROUP_SIZE_M": 64,
7
+ "num_warps": 4,
8
+ "num_stages": 4
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 32,
13
+ "BLOCK_SIZE_K": 128,
14
+ "GROUP_SIZE_M": 32,
15
+ "num_warps": 4,
16
+ "num_stages": 4
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 32,
21
+ "BLOCK_SIZE_K": 128,
22
+ "GROUP_SIZE_M": 1,
23
+ "num_warps": 4,
24
+ "num_stages": 4
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 32,
29
+ "BLOCK_SIZE_K": 128,
30
+ "GROUP_SIZE_M": 16,
31
+ "num_warps": 4,
32
+ "num_stages": 3
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 32,
37
+ "BLOCK_SIZE_K": 128,
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": 32,
45
+ "BLOCK_SIZE_K": 128,
46
+ "GROUP_SIZE_M": 64,
47
+ "num_warps": 4,
48
+ "num_stages": 4
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 16,
52
+ "BLOCK_SIZE_N": 32,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 64,
55
+ "num_warps": 4,
56
+ "num_stages": 4
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 16,
60
+ "BLOCK_SIZE_N": 64,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 16,
63
+ "num_warps": 4,
64
+ "num_stages": 3
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 16,
68
+ "BLOCK_SIZE_N": 64,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 64,
71
+ "num_warps": 4,
72
+ "num_stages": 3
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 64,
76
+ "BLOCK_SIZE_N": 64,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 32,
79
+ "num_warps": 4,
80
+ "num_stages": 4
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 64,
84
+ "BLOCK_SIZE_N": 64,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 64,
87
+ "num_warps": 4,
88
+ "num_stages": 5
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 64,
92
"BLOCK_SIZE_N": 64,
93
"BLOCK_SIZE_K": 128,
94
"GROUP_SIZE_M": 32,
7
- "num_warps": 8,
95
+ "num_warps": 4,
96
+ "num_stages": 3
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 64,
100
+ "BLOCK_SIZE_N": 64,
101
+ "BLOCK_SIZE_K": 128,
102
+ "GROUP_SIZE_M": 64,
103
+ "num_warps": 4,
104
+ "num_stages": 3
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 128,
108
+ "BLOCK_SIZE_N": 64,
109
+ "BLOCK_SIZE_K": 128,
110
+ "GROUP_SIZE_M": 64,
111
+ "num_warps": 4,
112
+ "num_stages": 4
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 64,
116
+ "BLOCK_SIZE_N": 128,
117
+ "BLOCK_SIZE_K": 128,
118
+ "GROUP_SIZE_M": 64,
119
+ "num_warps": 4,
120
+ "num_stages": 3
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 128,
124
+ "BLOCK_SIZE_N": 64,
125
+ "BLOCK_SIZE_K": 128,
126
+ "GROUP_SIZE_M": 1,
127
+ "num_warps": 4,
128
"num_stages": 3
129
},
130
"3072": {