The base and exponent of pow() should be allowed to have different data types

[miaobin]

Summary

pow() currently inherits the element-wise binary contract that requires both inputs to share the same data type. Unlike add/sub/mul — where both operands are the same kind of quantity — a power's base and exponent are conceptually distinct and are commonly different types (e.g. an integer or float32 exponent applied to a float16/int base). The same-type requirement forces redundant casts and rejects valid models.

Current behavior

We are currently encountering the following situation with the model (pow(int64_base, float32_exponent) is rejected at
graph-build time):

Backends status

• ONNX Pow-15 constrains X (and output) and Y with independent type sets — they need not match. Output type follows X (the base), consistent with WebNN's "output = a's type".
• TFLite tfl.pow and CoreML MIL elementwise_binary.pow likewise take base/exponent as separate operands.

Proposal and questions

• Give pow() its own tensor limits
  • Should the b (exponent) be restricted to a small set (e.g. float32 + uint32, TFLite only support tensor of 32-bit float or 32-bit signless integer values), or allowed broadly？
  • Should we accept the output based on the input a (with integer truncation)?
• Give pow() its own data-type validation