temprature as numpy.float64 is captured as a fake tensor during torch…

[umechand-amd]

Description:

What

speech_transformer fails the --export accuracy benchmark with fail_accuracy.

Why

The attention temperature is set with np.power(d_k, 0.5), which returns a numpy.float64. Under torch.export(strict=True), that numpy scalar gets captured as a fake (meta) constant, so the exported module returns a FakeTensor instead of a real output — and the accuracy check rejects it.

Fix

Cast the temperature to a plain Python float:

temperature=float(np.power(d_k, 0.5))
                                                                                                                                                                                                                                                                                                                                                                                   
Value is unchanged; it just traces as a normal scalar.    
                                                                                                                                                                                                                                                                                                                                                                                   
Testing
                                                                                                                                                                                                                                                                                                                                                                                   
- --export accuracy config now passes (previously fail_accuracy).                                                                                                                                                                                                                                                                                                                  
- python test.py -v -k speech_transformer — all pass.
- Not device-specific: confirmed on NVIDIA A100, AMD MI300/MI350X, and CPU.