functorch
functorch copied to clipboard
pytorch
functorch is JAX-like composable function transforms for PyTorch.
`jvp(f, x, tangents)` computes both f(x) and the Jacobian-vector product. The goal of `linearize(f)` is to return a function that can compute forward-mode AD multiple times without re-evaluating f(x).
I am trying to calculate the jacobian of a matrix exponential. However, it yields the following error ``` RuntimeError: vmap: inplace arithmetic(self, *extra_args) is not possible because there exists a...
Looking at a lot of jax code they seem to freely mix together `jit` and `vmap` for efficient vectorization. So I tried to do the same in functorch but found...
Functions like `jacrev` and `vmap` strip the type of an object down to `torch.Tensor`. - PyTorch version: 1.11.0+cu113 - functorch version: 0.1.0 - On Ubuntu 20.04, with Python 3.8.10 Let...
This decomposition: ``` @register_decomposition([aten.transpose], decompositions) def transpose(x, dim0: int, dim1: int): ndim = x.ndimension() dims = list(range(ndim)) dims[dim0], dims[dim1] = dims[dim1], dims[dim0] return torch.permute(x, dims) ``` Gives this error: ```...
My OS: Edition Windows 11 Pro Version 21H2 Installed on 12/23/2021 OS build 22000.556 Experience Windows Feature Experience Pack 1000.22000.556.0 ``` >>nvcc --version nvcc: NVIDIA (R) Cuda compiler driver Copyright...
Surprisingly this is the only TestCase that has tests failures (and nnc tests are passing). Repro: comment out https://github.com/pytorch/functorch/blob/52d9c91d7915c6af5cf23740fb1910887ca2975e/test/test_compile_cache.py#L188-L189 and run CI cc @Chillee @anijain2305
## 🚀 Feature Explore the existence of control flow operators: e.g. `functorch.cond`, `functorch.while_loop`, etc. ## Motivation (1) vmap over data-dependent control flow (2) lowering data-dependent control flow to backends. ##...
## What We should offer nightly binaries that match PyTorch's nightly binaries ## The difficulty I don't know what time we should build our nightly binaries. For example, PyTorch releases...
As per the title, I find that updating the tensors pointed by the `params` returned by `make_functional` does not update the real parameters in the original model. Is there a...
