Lukasz Stafiniak issues

Results 191 issues of


                                            Lukasz Stafiniak

Don't require the `ocannl_` prefix for commandline arguments, and warn / err on arguments missing from `ocannl_config.example`

Probably also rename `ocannl_config.example`. Double-validate: must be a superset of both commandline args, and of get_global_arg calls. Important for usability.

Refactor use_padding from a global shape inference setting to a field of the `Conv_input` constructor

Good first step on the road to padding inference.

Shape, projections, and padding inference for convolutions and affine einsum specs more generally

Adapt the FUN OCaml 2025 RL Workshop contents to idiomatic OCANNL

https://github.com/raven-ml/funocaml-2025-rl Originally at: https://github.com/lukstafi/raven-sokoban/tree/feat/sokoban-rl-environment/dev/fehu split between docs/ and examples/

RoPE and other non-learned position embeddings

Get rid of the `array` field of `Tnode.t`, of the "hosted" memory mode, and ideally also of the `Ndarray` module

We will anyway be moving towards a migration model where buffers can appear and disappear as needed, eventually, from the current static / indefinite lifetime model. So this change will...

enhancement

Blacklist primitive operator names when selecting identifier names

This isn't a bug strictly speaking but it's confusing. Then update the introductory slides also.

Remove remaining unnecessary zeroing-out in backend code

I noticed this when updating the introductory slides: ```c void scalar_loss_forward_and_gradient_then_gd_update( const uint32_t i1, const uint32_t i2, float *b1, float *b2, float *learning_rate, float *moons_classes, float *moons_flat, float *scalar_loss, float...

Implement tensor saving, loading and restoring

`load ?prefix_namespace path` loads all tensors from a file ensuring that they are disjoint from any pre-existing tensors; it optionally prefixes the namespaces of all the tensors stored in a...

Introduce namespaces for tensor node IDs

This is critical for "model surgery": being able to load weights of independently developed models, or weights of two snapshots of the same model. We add a type `Tnode.id =...