GPT-2 models
8 GPT-2 models on Hugging Face, from 110M to 1.6B parameters. At the precision each one is published in, the smallest needs about 0.5 GB of VRAM (vietnamese-gpt2, cheapest live fit: V100) and the largest about 7.2 GB (gpt2-xl, cheapest live fit: V100). The cheapest way to run vietnamese-gpt2 is $0.088/hr.
GPT-2 models
| Model | Parameters | Published as | VRAM needed | Live GPU fit | Est. $/hr |
|---|---|---|---|---|---|
| vietnamese-gpt2 | 110M | F32 | 0.5 GB | V100 | $0.088/hr |
| turkish-gpt2 | 124M | F32 | 0.6 GB | V100 | $0.088/hr |
| gpt2-small-dutch | 129M | F32 | 0.6 GB | V100 | $0.088/hr |
| gpt2 | 137M | F32 | 0.6 GB | V100 | $0.088/hr |
| gpt2-medium | 380M | F32 | 1.7 GB | V100 | $0.088/hr |
| megatron-gpt2-345m | 380M | F16 | 0.8 GB | V100 | $0.088/hr |
| gpt2-large | 812M | F32 | 3.6 GB | V100 | $0.088/hr |
| gpt2-xl | 1.6B | F32 | 7.2 GB | V100 | $0.088/hr |
VRAM is for the precision the model is published in, with the same overhead and an 8,192-token context assumed on every page; see the methodology. The fit shown is the lowest-priced single GPU type that holds the model at that precision, or the lowest-priced multi-GPU set (up to 8) when none does.
