Upload folder using huggingface_hub (#3)
Browse files- 2cc334c719f5c032d43379df0c83fbddd275c7c7887f0eb39b9b323e7cfee4d0 (e66c95a8e7c53fc5bb5f402609a388ac49de90ec)
- base_results.json +14 -14
- plots.png +0 -0
- smashed_results.json +14 -14
base_results.json
CHANGED
@@ -2,18 +2,18 @@
|
|
2 |
"current_gpu_type": "Tesla T4",
|
3 |
"current_gpu_total_memory": 15095.0625,
|
4 |
"perplexity": 2.2062017917633057,
|
5 |
-
"memory_inference_first":
|
6 |
-
"memory_inference":
|
7 |
-
"token_generation_latency_sync":
|
8 |
-
"token_generation_latency_async": 36.
|
9 |
-
"token_generation_throughput_sync": 0.
|
10 |
-
"token_generation_throughput_async": 0.
|
11 |
-
"token_generation_CO2_emissions": 0.
|
12 |
-
"token_generation_energy_consumption": 0.
|
13 |
-
"inference_latency_sync":
|
14 |
-
"inference_latency_async":
|
15 |
-
"inference_throughput_sync": 0.
|
16 |
-
"inference_throughput_async": 0.
|
17 |
-
"inference_CO2_emissions": 0.
|
18 |
-
"inference_energy_consumption": 0.
|
19 |
}
|
|
|
2 |
"current_gpu_type": "Tesla T4",
|
3 |
"current_gpu_total_memory": 15095.0625,
|
4 |
"perplexity": 2.2062017917633057,
|
5 |
+
"memory_inference_first": 7296.0,
|
6 |
+
"memory_inference": 7296.0,
|
7 |
+
"token_generation_latency_sync": 36.99171943664551,
|
8 |
+
"token_generation_latency_async": 36.34188920259476,
|
9 |
+
"token_generation_throughput_sync": 0.027033077002886197,
|
10 |
+
"token_generation_throughput_async": 0.027516456132076962,
|
11 |
+
"token_generation_CO2_emissions": 0.00017776892139547917,
|
12 |
+
"token_generation_energy_consumption": 0.002222163814551259,
|
13 |
+
"inference_latency_sync": 1029.307325744629,
|
14 |
+
"inference_latency_async": 132.0033311843872,
|
15 |
+
"inference_throughput_sync": 0.0009715271377055175,
|
16 |
+
"inference_throughput_async": 0.00757556639690526,
|
17 |
+
"inference_CO2_emissions": 0.00017755345632530548,
|
18 |
+
"inference_energy_consumption": 0.000612054947249437
|
19 |
}
|
plots.png
CHANGED
smashed_results.json
CHANGED
@@ -2,18 +2,18 @@
|
|
2 |
"current_gpu_type": "Tesla T4",
|
3 |
"current_gpu_total_memory": 15095.0625,
|
4 |
"perplexity": 2.2114033699035645,
|
5 |
-
"memory_inference_first":
|
6 |
-
"memory_inference":
|
7 |
-
"token_generation_latency_sync": 132.
|
8 |
-
"token_generation_latency_async": 132.
|
9 |
-
"token_generation_throughput_sync": 0.
|
10 |
-
"token_generation_throughput_async": 0.
|
11 |
-
"token_generation_CO2_emissions": 5.
|
12 |
-
"token_generation_energy_consumption": 0.
|
13 |
-
"inference_latency_sync": 328.
|
14 |
-
"inference_latency_async":
|
15 |
-
"inference_throughput_sync": 0.
|
16 |
-
"inference_throughput_async": 0.
|
17 |
-
"inference_CO2_emissions": 5.
|
18 |
-
"inference_energy_consumption": 0.
|
19 |
}
|
|
|
2 |
"current_gpu_type": "Tesla T4",
|
3 |
"current_gpu_total_memory": 15095.0625,
|
4 |
"perplexity": 2.2114033699035645,
|
5 |
+
"memory_inference_first": 2354.0,
|
6 |
+
"memory_inference": 2354.0,
|
7 |
+
"token_generation_latency_sync": 132.0942596435547,
|
8 |
+
"token_generation_latency_async": 132.1555681526661,
|
9 |
+
"token_generation_throughput_sync": 0.0075703516768890365,
|
10 |
+
"token_generation_throughput_async": 0.0075668397024694425,
|
11 |
+
"token_generation_CO2_emissions": 5.2118064791246415e-05,
|
12 |
+
"token_generation_energy_consumption": 0.007773342725803954,
|
13 |
+
"inference_latency_sync": 328.54746704101564,
|
14 |
+
"inference_latency_async": 121.8498706817627,
|
15 |
+
"inference_throughput_sync": 0.0030437002269603883,
|
16 |
+
"inference_throughput_async": 0.008206820363492354,
|
17 |
+
"inference_CO2_emissions": 5.156530753032632e-05,
|
18 |
+
"inference_energy_consumption": 0.0001779811755834419
|
19 |
}
|