Benchmarks / results_arcee_supernova.py
Julien Simon
Add SuperNova on g6e.12xlarge
5a0445e
raw
history blame
6.66 kB
"""Module containing performance results for the Arcee-SuperNova model."""
results_arcee_supernova = {
"name": "Arcee-SuperNova",
"modelType": "Llama 3.1 70B",
"configurations": [
{
"instanceType": "c7g.16xlarge",
"quantization": "Q4_0_8_8",
"container": "llama.cpp 9/19/24",
"status": "OK",
"tokensPerSecond": "6.5",
"notes": "",
},
{
"instanceType": "r8g.16xlarge",
"quantization": "Q4_0_4_8",
"container": "llama.cpp 9/19/24",
"status": "OK",
"tokensPerSecond": "9.8",
"notes": "With Flash Attention",
},
{
"instanceType": "g5.12xlarge",
"quantization": "awq",
"container": "TGI 2.2.0",
"status": "OK",
"tokensPerSecond": "33",
"notes": "MAX_INPUT_TOKENS: 8192, MAX_TOTAL_TOKENS: 16384",
},
{
"instanceType": "g6e.2xlarge",
"quantization": "awq (w4 g128)",
"container": "vLLM 0.6.2",
"status": "OK",
"tokensPerSecond": "18",
"notes": "--max-model-len 10000 --max-num-seqs 16 --enforce-eager",
},
{
"instanceType": "g6e.2xlarge",
"quantization": "Q4_K_M",
"container": "llama.cpp 10/2/24",
"status": "OK",
"tokensPerSecond": "16",
"notes": "-ngl 81 -c 13000 -fa -t 8",
},
{
"instanceType": "g6e.12xlarge",
"quantization": "None",
"container": "vLLM 0.6.3",
"status": "OK",
"tokensPerSecond": "18.6",
"notes": "--max-model-len 16384",
},
{
"instanceType": "p4d.24xlarge",
"quantization": "awq",
"container": "TGI 2.2.0",
"status": "OK",
"tokensPerSecond": "58",
"notes": "MAX_INPUT_TOKENS: 16384, MAX_TOTAL_TOKENS: 32768",
},
{
"instanceType": "p5.48xlarge",
"quantization": "awq",
"container": "TGI 2.2.0",
"status": "OK",
"tokensPerSecond": "73",
"notes": "MAX_INPUT_TOKENS: 16384, MAX_TOTAL_TOKENS: 32768",
},
{
"instanceType": "inf2.24xlarge",
"configurations": [
{
"quantization": "none",
"container": "LMI 0.29+transformers-neuronx 0.11.351",
"status": "KO",
"tokensPerSecond": "-",
"notes": "OOM bs=2,seqlen=4096 - SDK 2.19.1",
},
{
"quantization": "none",
"container": "LMI 0.29+transformers-neuronx 0.11.351",
"status": "KO",
"tokensPerSecond": "-",
"notes": "OOM bs=2,seqlen=2048 - SDK 2.19.1",
},
{
"quantization": "8-bit",
"container": "LMI 0.29+transformers-neuronx 0.11.351",
"status": "???",
"tokensPerSecond": "???",
"notes": "bs=2,seqlen=8192 - SDK 2.19.1 - OPTION_LOAD_IN_8BIT=True",
},
],
},
{
"instanceType": "inf2.48xlarge",
"configurations": [
{
"quantization": "none",
"container": "LMI 0.29+transformers-neuronx 0.11.351",
"status": "OK",
"tokensPerSecond": "28",
"notes": "bs=4,seqlen=4096 - SDK 2.19.1",
},
{
"quantization": "none",
"container": "LMI 0.29+transformers-neuronx 0.11.351",
"status": "OK",
"tokensPerSecond": "24",
"notes": "bs=2,seqlen=8192 - SDK 2.19.1",
},
{
"quantization": "none",
"container": "LMI 0.29+transformers-neuronx 0.11.351",
"status": "KO",
"tokensPerSecond": "-",
"notes": "OOM bs=2,seqlen=16384 - SDK 2.19.1",
},
],
},
{
"instanceType": "trn1.32xlarge",
"configurations": [
{
"quantization": "none",
"container": "LMI 0.29+transformers-neuronx 0.11.351",
"status": "OK",
"tokensPerSecond": "32",
"notes": "bs=2,seqlen=8192 - SDK 2.19.1",
},
{
"quantization": "none",
"container": "LMI 0.30rc1",
"status": "OK",
"tokensPerSecond": "34",
"notes": "bs=2,seqlen=8192 - SDK 2.20",
},
],
},
{
"instanceType": "p4d.24xlarge",
"configurations": [
{
"quantization": "none",
"container": "TGI 2.2.0",
"status": "OK",
"tokensPerSecond": "30",
"notes": "",
},
{
"quantization": "none",
"container": "LMI 0.29+vLLM 0.5.5",
"status": "OK",
"tokensPerSecond": "45",
"notes": "OPTION_MAX_MODEL_LEN 64k",
},
],
},
{
"instanceType": "p5.48xlarge",
"configurations": [
{
"quantization": "none",
"container": "TGI 2.2.0",
"status": "OK",
"tokensPerSecond": "58",
"notes": "MAX_INPUT_TOKENS: 16384, MAX_TOTAL_TOKENS: 32768",
},
{
"quantization": "none",
"container": "LMI 0.29+vLLM 0.5.5",
"status": "OK",
"tokensPerSecond": "70",
"notes": "OPTION_MAX_MODEL_LEN 128k",
},
{
"quantization": "none",
"container": "LMI 0.29+vLLM 0.5.5",
"status": "OK",
"tokensPerSecond": "70",
"notes": "OPTION_ENFORCE_EAGER=True",
},
],
},
],
}