meta-llama/Llama-2-7b
converted to the HF format but maintaining the original precision it was traiend on (bf16) instead of converting to fp16 like meta-llama/Llama-2-7b-hf
does.
When benchmarked, it performs almost identically to the fp16 version.