We read every piece of feedback, and take your input very seriously.
To see all available qualifiers, see our documentation.
Similar to #8633, we'll need to refactor pallas kernels needed by Mixtral (e.g. GMM) into PyTorch custom ops in order to use scan in Mixtral.
