update cuda op to support fp16/bf16 absmax

Signed-off-by: jiqing-feng <[email protected]>

Author: jiqing-feng

bitsandbytes/backends/cuda/ops.py

@@ -237,7 +237,7 @@ def _(A: torch.Tensor, code: torch.Tensor, blocksize: int) -> tuple[torch.Tensor
         else:
             raise ValueError(f"Blockwise quantization only supports 16/32-bit floats, but got {A.dtype}")
 
-    return out, absmax
+    return out, absmax.to(A.dtype)
 
 
 @register_kernel("bitsandbytes::dequantize_blockwise", "cuda")
@@ -332,7 +332,7 @@ def _(
             else:
                 lib.cquantize_blockwise_fp32_nf4(*args)
 
-    return out, absmax
+    return out, absmax.to(A.dtype)
 
 
 @register_kernel("bitsandbytes::dequantize_4bit", "cuda")