opt-30b-deepspeed-inference-fp16-shard-4

This is a copy of the original OPT weights that is more efficient to use with the DeepSpeed-MII and DeepSpeed-Inference. In this repo the original tensors are split into 4 shards to target 4 GPUs, this allows the user to run the model with DeepSpeed-inference Tensor Parallelism.

For specific details about the OPT model itself, please see the original OPT model card.

For examples on using this repo please see the following:

• https://github.com/huggingface/transformers-bloom-inference
• https://github.com/microsoft/DeepSpeed-MII