Jack Gallagher
Jack Gallagher
This issue is to figure out what size the stack should be by default - the current one is small enough that it's really easy to run into stack overflow...
Currently, functions like `match_term` return substitutions represented as a list of pairs. Why don't all these functions use maps instead of lists? ## Want to back this issue? **[Post a...
It'd be nice to be able to have an `Option` which didn't take up more space than `Key` in memory.
it would be really nice to be able to eg take models implemented in jax with `jax.lax.scan` and port them over to torch without having to unroll scans over modules
I was trying to implement a fused `abs_angle` kernel for complex tensors, and needed to use inverse trigonometric ops (specifically `atan`), and noticed that the available math instructions are very...
I'd like to use equinox for some fairly large-scale training runs, but the state for those models is often too large to fit on a single accelerator, so gathering all...
right now, there's clear documentation re how to use `custom_jvp` and `custom_vjp`, but the automatic transposition of `custom_jvp` isn't necessarily going to be very good for pallas kernels (is it...
it would be nice to have some fields be pytree nodes and others not, which would make this a full replacement for `flax.struct`
currently impossible to use `flash_attention` within a function that will use gradient checkpointing minimal example to reproduce: ```py b = 3 lq = 16 lkv = 17 h = 5...