brrr cuda dtypes float8 inference llama mx offloading optimizer pytorch quantization sparsity training transformer