Sam

Can't like songs

5

I can't like songs with the `s` key, in either the playbar or in search screens. Is there a location I can look for logs? Or a CLI option like...

Only wrap modules in DDP if they require grad

1

Like the title says. Fixes #760. There is [just such a check](https://github.com/pytorch/pytorch/blob/13cff2ee8ea1d7aea2ad201cbd77ebe2b9a29d25/torch/nn/parallel/distributed.py#L550-L555) in DistributedDataParallel, which I do the opposite of. I would like to add a test so that this...

RuntimeError: DistributedDataParallel is not needed when a module doesn't have any parameter that requires a gradient.

1

I normally just throw everything into `accelerator.prepare` that has any tensor buffers or parameters. However, if an object subclasses `torch.nn.Module`, has no parameters and we are in a multi-GPU context,...

I am using ffcv loaders with [huggingface's accelerate](https://github.com/huggingface/accelerate) for single-node multi-GPU training using 8x A100 with 16GB memory each. When using ffcv, my training loop is slower and I need...

- improves support for #24 (hex). Design - Adds a struct to represent the config right after loading from the `.ini` file using primitive types (`string`, `int`, `bool`, etc.) -...

The `ml_tag_molting_waterfowl` task has two text labels: `has_red_eyes` and `not_red_eyes`. I think these should be `molting` and `not_molting`. Can you confirm?

Given a model compiled with: ```py model = torch.compile(model, mode="reduce-overhead", fullgraph=True, dynamic=True) ``` where the bulk of the task is computing next token logits for different prompts (MMLU), memory usage...

Non-default `__init__` with `default_factory` throws error.

2

```py import equinox as eqx class Module(eqx.Module): stuff: object default_int: 2 default_lst: list[object] = eqx.field(default_factory=list) def __init__(self, stuff: object): self.stuff = fancy_calculation(stuff) if __name__ == "__main__": module = Module(1) ```...

PyTorch Hooks in Equinox?

2

I would like to record some model activations in an architecture-invariant way. In PyTorch, we can use [forward hooks](https://pytorch.org/docs/stable/generated/torch.nn.modules.module.register_module_forward_hook.html) to do this, by registering a hook on modules that match...

cloudpickle + weakref + index_variadic

4

Cloudpickle (latest, 3.3.1 at time of writing) fails to pickle jaxtyped functions because of the weakref (introduced in 0.2.35 of jaxtyping). I have a MWE using uv with inline packages:...