cca14b0c63baf918c6151d0a13b34ea0775182297a000f6fa6c77b1362141c1f

base_results.json

@@ -0,0 +1,19 @@
+{
+    "current_gpu_type": "Tesla T4",
+    "current_gpu_total_memory": 15095.0625,
+    "perplexity": 2.2062017917633057,
+    "memory_inference_first": 7296.0,
+    "memory_inference": 7296.0,
+    "token_generation_latency_sync": 36.912679481506345,
+    "token_generation_latency_async": 36.8165610358119,
+    "token_generation_throughput_sync": 0.02709096207716405,
+    "token_generation_throughput_async": 0.027161689518673086,
+    "token_generation_CO2_emissions": 0.00017793332995351462,
+    "token_generation_energy_consumption": 0.0022603336588687913,
+    "inference_latency_sync": 1028.5993484497071,
+    "inference_latency_async": 132.02483654022217,
+    "inference_throughput_sync": 0.0009721958326215045,
+    "inference_throughput_async": 0.007574332422637341,
+    "inference_CO2_emissions": 0.00017765837549938682,
+    "inference_energy_consumption": 0.0006121215726133666
+}

smashed_results.json

@@ -0,0 +1,19 @@
+{
+    "current_gpu_type": "Tesla T4",
+    "current_gpu_total_memory": 15095.0625,
+    "perplexity": 2.3863821029663086,
+    "memory_inference_first": 1598.0,
+    "memory_inference": 1598.0,
+    "token_generation_latency_sync": 153.33177795410157,
+    "token_generation_latency_async": 152.51875221729279,
+    "token_generation_throughput_sync": 0.0065218052861771465,
+    "token_generation_throughput_async": 0.006556570818093925,
+    "token_generation_CO2_emissions": 5.481785809408624e-05,
+    "token_generation_energy_consumption": 0.008956993263524471,
+    "inference_latency_sync": 345.3065460205078,
+    "inference_latency_async": 190.85931777954102,
+    "inference_throughput_sync": 0.002895977535104735,
+    "inference_throughput_async": 0.005239461251533375,
+    "inference_CO2_emissions": 5.5271057652387143e-05,
+    "inference_energy_consumption": 0.00018832182125658322
+}