Upload cpu_benchmark_results.json with huggingface_hub
Browse files- cpu_benchmark_results.json +68 -0
cpu_benchmark_results.json
ADDED
|
@@ -0,0 +1,68 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"kernel_benchmark": [
|
| 3 |
+
{
|
| 4 |
+
"model_id": "openai-community/gpt2",
|
| 5 |
+
"model_name": "GPT-2 (124M)",
|
| 6 |
+
"n_params_M": 124.4,
|
| 7 |
+
"seq_len": 56,
|
| 8 |
+
"device": "cpu",
|
| 9 |
+
"baseline_ms": 39.176,
|
| 10 |
+
"s20_injected_ms": 41.556,
|
| 11 |
+
"s20_overhead": 1.061,
|
| 12 |
+
"avg_perplexity": 175.7,
|
| 13 |
+
"perplexity_per_prompt": [
|
| 14 |
+
89.63,
|
| 15 |
+
202.44,
|
| 16 |
+
235.02
|
| 17 |
+
]
|
| 18 |
+
},
|
| 19 |
+
{
|
| 20 |
+
"model_id": "distilbert/distilgpt2",
|
| 21 |
+
"model_name": "DistilGPT-2 (82M)",
|
| 22 |
+
"n_params_M": 81.9,
|
| 23 |
+
"seq_len": 56,
|
| 24 |
+
"device": "cpu",
|
| 25 |
+
"baseline_ms": 21.623,
|
| 26 |
+
"s20_injected_ms": 21.331,
|
| 27 |
+
"s20_overhead": 0.987,
|
| 28 |
+
"avg_perplexity": 302.02,
|
| 29 |
+
"perplexity_per_prompt": [
|
| 30 |
+
223.34,
|
| 31 |
+
339.5,
|
| 32 |
+
343.22
|
| 33 |
+
]
|
| 34 |
+
},
|
| 35 |
+
{
|
| 36 |
+
"model_id": "facebook/opt-125m",
|
| 37 |
+
"model_name": "OPT-125M",
|
| 38 |
+
"n_params_M": 125.2,
|
| 39 |
+
"seq_len": 57,
|
| 40 |
+
"device": "cpu",
|
| 41 |
+
"baseline_ms": 29.008,
|
| 42 |
+
"s20_injected_ms": 29.36,
|
| 43 |
+
"s20_overhead": 1.012,
|
| 44 |
+
"avg_perplexity": 199.73,
|
| 45 |
+
"perplexity_per_prompt": [
|
| 46 |
+
148.67,
|
| 47 |
+
265.95,
|
| 48 |
+
184.57
|
| 49 |
+
]
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"model_id": "bigscience/bloom-560m",
|
| 53 |
+
"model_name": "BLOOM-560M",
|
| 54 |
+
"n_params_M": 559.2,
|
| 55 |
+
"seq_len": 48,
|
| 56 |
+
"device": "cpu",
|
| 57 |
+
"baseline_ms": 122.361,
|
| 58 |
+
"s20_injected_ms": 117.968,
|
| 59 |
+
"s20_overhead": 0.964,
|
| 60 |
+
"avg_perplexity": 138.99,
|
| 61 |
+
"perplexity_per_prompt": [
|
| 62 |
+
177.74,
|
| 63 |
+
209.88,
|
| 64 |
+
29.35
|
| 65 |
+
]
|
| 66 |
+
}
|
| 67 |
+
]
|
| 68 |
+
}
|