Can Relax frontend convert bf16 scalar value?

myeongbo · August 7, 2025, 3:08pm

I’m trying to convert LLM models including bf16 scalar value and it seems that relax.constant try to convert it to numpy which doesn’t natively support bf16 type.

github.com/apache/tvm

python/tvm/relax/expr.py

3c189f015


      
          
          Note
          ----
          When dtype is None, we use the following rule:
          
          - int maps to "int32"
          - float maps to "float32"
          - bool maps to "bool"
          - other using the same default rule as numpy.
          """
          if isinstance(value, (Number, (bool, list))):
              value = _np.array(value, dtype=dtype)
          
          if not dtype:
              # when dtype is None: int maps to "int32", float maps to "float32"
              dtype = {  # type: ignore
                  _np.dtype("int64"): _np.int32,  # type: ignore
                  _np.dtype("float64"): _np.float32,  # type: ignore
              }.get(
                  value.dtype, None  # type: ignore
              )

Can anyone explain how relax frontend create a constant node from bf16 number?