Fp32 vs fp16
Thank you for contributing to this excellent model.
I have a question, the model is trained using float32 data type, but due to resource constraints, I am performing inference with fp16. Does this significantly impact the performance of the model?
Currently, I find it not performing as well as Orca 1 when inferring with fp16.
The model is trained with bfloat16. With fp16 inference you might see a loss, but overall that affects both Orca 1 and Orca 2. You can see the inference code here: https://huggingface.co./spaces/ari9dam/Orca-2-13B
(imp : use slow version of the tokenizer)
The model is trained with bfloat16. With fp16 inference you might see a loss, but overall that affects both Orca 1 and Orca 2. You can see the inference code here: https://huggingface.co./spaces/ari9dam/Orca-2-13B
(imp : use slow version of the tokenizer)
Thank you very much for your response.
As I see in the config.json file, Orca 2 used "torch_dtype": "float32", in the other hand, Orca 1 used "torch_dtype": "bfloat16". Adding one more thing, the total weight file size of Orca 1 is 26GB, while that of Orca 2 is 53GB. It implies that Orca 2 is storing weights in a data type that is twice the size of Orca 1. So I still do not quite understand your statement 'The model is trained with bfloat16.'.