-
## 🚀 Feature
Add support for PyTorch Callable -> Thunder Callable translation in Thunder JIT.
### Motivation
Several PyTorch operators accept a Python function with PyTorch operations inside …
-
Exists in TF: https://www.tensorflow.org/api_docs/python/tf/scan (or some variant inspired by scan), JAX/LAX: https://jax.readthedocs.io/en/latest/_autosummary/jax.lax.scan.html, (old theano: https://…
-
### 🚀 The feature, motivation and pitch
It would be great to have a general parallel prefix sum (associative scan) operation in PyTorch, something like [associative_scan](https://jax.readthedocs.io…
-
### 🐛 Describe the bug
My networks rely on varying shapes during training as well as during inference. Thus, I tried to use `torch.compile(... dynamic=True)` as well as the `torch._dynamo.optimize(..…
-
# Remember TorchRL: the state of memory in TorchRL
Hello! This is a discussion post to recap the state of memory models in TorchRL: what's doable, what's not doable, what is the way to do things, a…
-
We all know RNNs have this problem. While the paper "Were RNNs All We Needed?" focuses on parallelism, does it also lay down any changes to handle vanishing gradients?
Just curious.
-
Hello, thanks for your work. I wonder what is the difference between the proposed algorithm and sect.1.4.1 in https://www.cs.cmu.edu/~guyb/papers/Ble93.pdf.
-
# Failing Tests
> Please see the failing tests divided into sections below. Click on each section to expand. Feel free to get assigned to an issue by following the instructions [here](https://unify.ai…
-
Dear, how to use prev_state in apply_ssm function since I see it is now purely forward?
I would ideally want to:
x, states = s5(x, states), where apply_ssm carries state such that I can train with m…
-
Hi! Just a quick question. Isn't it possible to use a parallel scan already existing in jax? https://jax.readthedocs.io/en/latest/_autosummary/jax.lax.associative_scan.html (with binary operator from …