Thanks - the current behavior seems intended: https://onnx.ai/onnx/technical/float8.html#cast

Is there an example from a different framework that will turn NaNs into 0s for reference?

Why would the original value be NaN in the first place? Can that be fixed?

I have fix it, add an arg.

I still don’t think this is the right fix. Does this behavior (NaN to 0) exist in other frameworks? Should the quantization tool itself be fixed?

I am also still wondering where did the NaN values come from

If it’s due to division by zero, then the actual computation (in the call site, not in onnx) needs to be fixed

Sorry, I can't understand the meaning, I think it's difficult to see why NaN happens. And other kinds of packages also has the feature to make NaN to 0.

Sorry, I can't understand the meaning, I think it's difficult to see why NaN happens. And other kinds of packages also has the feature to make NaN to 0.

Out of interest, which package does this?

NaN should be kept as NaN in down-casting. NaN indicates previous computing errors, hiding NaN by 0 makes no sense. The current behaviour is the same as numpy.

import numpy as np


# Casting NaN to another float type
arr_float64 = np.array([1.0, np.nan], dtype=np.float64)
arr_float32 = arr_float64.astype(np.float32)
print(f"Original float64 array with NaN: {arr_float64}")
print(f"Casting to float32: {arr_float32}")  # NaN is preserved