You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Currently we don't check whether a QuantTensor is per channel or per tensor when we do certain ops like flatten or shuffle/unshuffle that are sensitive to it.
The text was updated successfully, but these errors were encountered:
PyTorch Tensor has the qscheme attribute to specify how to quantize a tensor, which includes:
torch.per_tensor_affine
torch.per_tensor_symmetric
torch.per_channel_affine
torch.per_channel_symmetric
In Brevitas, if granularity is our only concern so far, maybe we can assess it by examining the shape of the quantization parameters (scale, zero point):
Currently we don't check whether a QuantTensor is per channel or per tensor when we do certain ops like flatten or shuffle/unshuffle that are sensitive to it.
The text was updated successfully, but these errors were encountered: