Re-Upload of https://huggingface.co/dunzhang/stella_en_400M_v5 with :
- Only 1024 embeddings dimensions folder
- Max tokens lenght to 512 ( model has been trained on this sequence lenght )
- Padding strategy set to "BatchLongest" to avoid batching every line to a fixed length of 512 tokens, cause less than 512 is possible sometimes. Remove unecessary padding and save memory.
Parameters at the end of the file "config.json" has been set manually to false for CPU usage:
"unpad_inputs": false, "use_memory_efficient_attention": false
You can turn them back to "true" to enable GPU back again