functorch

functorch copied to clipboard

[POC] vmap for TensorDict

1

Makes the necessary changes to vmap.py to make [Using TensorDict for functorch](https://github.com/facebookresearch/rl/pull/364) work

vmoens

autograd.Function is silently incorrect on functorch transforms

11

We should fix it. Not quite sure how though. https://gist.github.com/zou3519/73a6189e21561f6ef5b42874e8a4826f

zou3519

AOT Autograd - LSTM - grads not generated (model tts_angular)

1

This is a subgraph from `tts_angular` model The generated backward pass has many `None` outputs, suggesting that that `requires_grad` is somehow not passed correctly when LSTM cell is used. ~~~...

anijain2305

Hello, I saw that here is a place to leave suggestions and use cases. I am working on solving PDEs (numerical method) using PyTorch, I cannot avoid Jacobian (Newton iteration,...

caowen1510957

No Batching rules for aten::_linalg_solve_ex, aten::linalg_solve, aten::linalg_solve_ex, and aten::_linalg_slogdet causes significant slowdown for per-sample gradients with torch.linalg.slogdet

16

TL;DR - `torch.linalg.slogdet` is over one order of magnitude slower in computing per-sample gradients in the latest nightly version of PyTorch/FuncTorch (`1.13.0.dev20220721` / ` 0.3.0a0+e8a68f4`) than a previous version of...

AlphaBetaGamma96

MPS Support

1

Hello, I am using functorch for my project. It's a good substitute for Jax and makes life easy. However it does not have GPU support for M1 chips which is...

Mahnoo06

[Feature Request] Support `different` randomness settings to train an ensemble of models with TorchOpt

3

## Motivation We recently used [TorchOpt](https://github.com/metaopt/TorchOpt) as a functional optimizer API mentioned in [functorch parallel training example](https://github.com/pytorch/functorch/blob/main/examples/ensembling/parallel_train.py#L86) to achieve [batchable optimization training small neural networks on one GPU](https://github.com/metaopt/TorchOpt/blob/main/examples/FuncTorch/parallel_train_torchopt.py) with `functorch.vmap`....

Benjamin-eecs

vmap and forward-mode AD fail sometimes on in-place views

9

## The Problem ```py import torch from functorch import jvp, vmap from functools import partial B = 2 def f(x, y): x = x.clone() view = x[0] x.copy_(y) return view,...

zou3519

Batching rule for GRU cell

3

Hi, I have a GRU model and want to calculate the jacobian of the model with functorch. But there is a performance drop because we have not yet implemented the...

0aqz0

This happens a lot to us, we should add a test to prevent it from happening. Test probably needs to shell out

zou3519