Hugging Face's logo

callensxavier
/
s20-attention-kernel

attention
positional-encoding
efficient-attention
number-theory
calabi-yau
negative-result
Model card Files
xet
 Community
callensxavier commited on
Commit
51243c0
·
verified ·
1 Parent(s): 6491ed4

Upload benchmark_results_full.json with huggingface_hub

Browse files
Files changed (1) hide show
  1. benchmark_results_full.json +106 -0
benchmark_results_full.json ADDED
@@ -0,0 +1,106 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "benchmark": [
3
+ {
4
+ "seq_len": 64,
5
+ "hardware": "CPU",
6
+ "device": "cpu",
7
+ "batch": 1,
8
+ "heads": 8,
9
+ "head_dim": 64,
10
+ "fp16_sdpa_ms": 0.296,
11
+ "alix_v1_legacy_ms": null,
12
+ "alix_v2_vectorized_ms": 0.218,
13
+ "lia_v2_vectorized_ms": 0.206,
14
+ "speedup_v1_to_v2": null,
15
+ "fp16_vs_s20_ratio": 0.737,
16
+ "alix_v2_correct": true,
17
+ "lia_v2_correct": true,
18
+ "alix_v1_correct": true
19
+ },
20
+ {
21
+ "seq_len": 128,
22
+ "hardware": "CPU",
23
+ "device": "cpu",
24
+ "batch": 1,
25
+ "heads": 8,
26
+ "head_dim": 64,
27
+ "fp16_sdpa_ms": 1.04,
28
+ "alix_v1_legacy_ms": null,
29
+ "alix_v2_vectorized_ms": 0.37,
30
+ "lia_v2_vectorized_ms": 0.362,
31
+ "speedup_v1_to_v2": null,
32
+ "fp16_vs_s20_ratio": 0.355,
33
+ "alix_v2_correct": true,
34
+ "lia_v2_correct": true,
35
+ "alix_v1_correct": true
36
+ },
37
+ {
38
+ "seq_len": 256,
39
+ "hardware": "CPU",
40
+ "device": "cpu",
41
+ "batch": 1,
42
+ "heads": 8,
43
+ "head_dim": 64,
44
+ "fp16_sdpa_ms": 3.644,
45
+ "alix_v1_legacy_ms": null,
46
+ "alix_v2_vectorized_ms": 0.776,
47
+ "lia_v2_vectorized_ms": 0.733,
48
+ "speedup_v1_to_v2": null,
49
+ "fp16_vs_s20_ratio": 0.213,
50
+ "alix_v2_correct": true,
51
+ "lia_v2_correct": true,
52
+ "alix_v1_correct": true
53
+ },
54
+ {
55
+ "seq_len": 512,
56
+ "hardware": "CPU",
57
+ "device": "cpu",
58
+ "batch": 1,
59
+ "heads": 8,
60
+ "head_dim": 64,
61
+ "fp16_sdpa_ms": 12.892,
62
+ "alix_v1_legacy_ms": null,
63
+ "alix_v2_vectorized_ms": 2.106,
64
+ "lia_v2_vectorized_ms": 2.083,
65
+ "speedup_v1_to_v2": null,
66
+ "fp16_vs_s20_ratio": 0.163,
67
+ "alix_v2_correct": true,
68
+ "lia_v2_correct": true,
69
+ "alix_v1_correct": true
70
+ },
71
+ {
72
+ "seq_len": 1024,
73
+ "hardware": "CPU",
74
+ "device": "cpu",
75
+ "batch": 1,
76
+ "heads": 8,
77
+ "head_dim": 64,
78
+ "fp16_sdpa_ms": 44.105,
79
+ "alix_v1_legacy_ms": null,
80
+ "alix_v2_vectorized_ms": 8.382,
81
+ "lia_v2_vectorized_ms": 8.424,
82
+ "speedup_v1_to_v2": null,
83
+ "fp16_vs_s20_ratio": 0.19,
84
+ "alix_v2_correct": true,
85
+ "lia_v2_correct": true,
86
+ "alix_v1_correct": true
87
+ },
88
+ {
89
+ "seq_len": 2048,
90
+ "hardware": "CPU",
91
+ "device": "cpu",
92
+ "batch": 1,
93
+ "heads": 8,
94
+ "head_dim": 64,
95
+ "fp16_sdpa_ms": 132.415,
96
+ "alix_v1_legacy_ms": null,
97
+ "alix_v2_vectorized_ms": 26.687,
98
+ "lia_v2_vectorized_ms": 26.639,
99
+ "speedup_v1_to_v2": null,
100
+ "fp16_vs_s20_ratio": 0.202,
101
+ "alix_v2_correct": true,
102
+ "lia_v2_correct": true,
103
+ "alix_v1_correct": true
104
+ }
105
+ ]
106
+ }