harshithsaiv commited on
Commit
c1bcd73
·
1 Parent(s): c7426b8

feat: Mistral results for long context benchmark

Browse files
results/mistral-7b/long_context_results.json ADDED
@@ -0,0 +1,75 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "mistral-7b",
3
+ "results": [
4
+ {
5
+ "context_len": 512,
6
+ "peak_memory_gb": 14.63,
7
+ "fp16_mb": 67.11,
8
+ "uniform8_mb": 33.55,
9
+ "mixed_precision_mb": 29.17,
10
+ "compression_vs_fp16": 2.3,
11
+ "compression_vs_8bit": 1.15,
12
+ "prefill_ms": 57.0
13
+ },
14
+ {
15
+ "context_len": 1024,
16
+ "peak_memory_gb": 14.76,
17
+ "fp16_mb": 134.22,
18
+ "uniform8_mb": 67.11,
19
+ "mixed_precision_mb": 58.33,
20
+ "compression_vs_fp16": 2.3,
21
+ "compression_vs_8bit": 1.15,
22
+ "prefill_ms": 85.1
23
+ },
24
+ {
25
+ "context_len": 2048,
26
+ "peak_memory_gb": 15.02,
27
+ "fp16_mb": 268.44,
28
+ "uniform8_mb": 134.22,
29
+ "mixed_precision_mb": 116.66,
30
+ "compression_vs_fp16": 2.3,
31
+ "compression_vs_8bit": 1.15,
32
+ "prefill_ms": 165.6
33
+ },
34
+ {
35
+ "context_len": 4096,
36
+ "peak_memory_gb": 15.53,
37
+ "fp16_mb": 536.87,
38
+ "uniform8_mb": 268.44,
39
+ "mixed_precision_mb": 233.31,
40
+ "compression_vs_fp16": 2.3,
41
+ "compression_vs_8bit": 1.15,
42
+ "prefill_ms": 333.1
43
+ },
44
+ {
45
+ "context_len": 8192,
46
+ "peak_memory_gb": 16.56,
47
+ "fp16_mb": 1073.74,
48
+ "uniform8_mb": 536.87,
49
+ "mixed_precision_mb": 466.62,
50
+ "compression_vs_fp16": 2.3,
51
+ "compression_vs_8bit": 1.15,
52
+ "prefill_ms": 700.6
53
+ },
54
+ {
55
+ "context_len": 16384,
56
+ "peak_memory_gb": 18.61,
57
+ "fp16_mb": 2147.48,
58
+ "uniform8_mb": 1073.74,
59
+ "mixed_precision_mb": 933.24,
60
+ "compression_vs_fp16": 2.3,
61
+ "compression_vs_8bit": 1.15,
62
+ "prefill_ms": 1554.1
63
+ },
64
+ {
65
+ "context_len": 32768,
66
+ "peak_memory_gb": 22.71,
67
+ "fp16_mb": 4294.97,
68
+ "uniform8_mb": 2147.48,
69
+ "mixed_precision_mb": 1866.47,
70
+ "compression_vs_fp16": 2.3,
71
+ "compression_vs_8bit": 1.15,
72
+ "prefill_ms": 3807.8
73
+ }
74
+ ]
75
+ }