et-nivard

Interpreter is reporting FullyConnected operators with quantized activations and float bias as supported: see https://github.com/pytorch/glow/blob/master/lib/Backends/Interpreter/Interpreter.cpp#L192. But when trying it I'm hitting the assert here https://github.com/pytorch/glow/blob/master/lib/Backends/Interpreter/InterpreterNodes.cpp#L315 because the implementation relies on...

Quantized types with scale <= 0

2

I'm trying to enable a protection against invalid/malformed test cases and I found that all Scratch operands are being generated as type Int8QTy with scale = 0.0 (see https://github.com/pytorch/glow/blob/master/tools/ClassGen/InstrBuilder.cpp#L97 and...

Make allowedError / isEqual() independent of Tensor dynamic range

2

Tensor method isEqual() allows specifying a threshold of acceptable absolute error when comparing tensors: https://github.com/pytorch/glow/blob/master/include/glow/Base/Tensor.h#L871 This approach has a strong dependency on the dynamic range of the tensor under check....