3 ms·
The hard part of embedding is that the smallest 124M GPT-2 model itself is huge at 500MB, which would be unreasonable for performance/storage on the user end (a
by minimaxir 6y ago
The hard part of embedding is that the smallest 124M GPT-2 model itself is huge at 500MB, which would be unreasonable for performance/storage on the user end (and quantization/tracing can't save that much space).
Hence why I'm looking into smaller models, which has been difficult, but releasing aitextgen was a necessary first step.