The Plan 0.4

rabroldan

Ronald R

Posted on November 29, 2023

The Plan 0.4

Release 0.4 is taking on the biggest problem, the biggest task, and the largest project we can for now. It's crazy to think that we are in the final legs of our term.

We have three things.

  1. What I'll do
  2. How I'll do it
  3. The goals.

Let's start with the What

What I need to do is find a bug that I want to work with. I have no clue where to find it. so I placed in a few labels on github

archived:false is:open language:Python label:bug
Enter fullscreen mode Exit fullscreen mode

pretty simple but straight forward I am always looking for python projects to work on, I had a few chances to take a look on some of the repositories I worked but it was not enough, well for me though. I sent a few comments on a few issues what I wanted to work on but there is really one issue I want to work on. Which was funccol collectives rewrite in dynamo does not work w/ kwargs #114179 which is an issue from pytorch.

Just a quick rundown on the bug we need to fix a bug. So this code is working it's fine

dist.all_gather_into_tensor(
                output_tensor=flat_tensor, input_tensor=input_flat_tensor_shard, group=group
            )
Enter fullscreen mode Exit fullscreen mode

BUT

When we start rewritting it with dynamoDB an start using kwargs or keyword arguments it's not working.

dist.all_gather_into_tensor(
                flat_tensor, input_flat_tensor_shard, group=group
            )
Enter fullscreen mode Exit fullscreen mode

Now the How

Let's discuss the how, I need to learn how dynamo rewriting works.they mentioned dynamically rewriting it with dynamo, which in all honestly I have no clue how dynamically re writing works so ill start there. I have nothing much to say here yet, just because the how is something I need to work on first.

they asked me to check out

def _traceable_collective_remaps():
    # We can't rely on importing from distributed, since it's not always built
    if torch.distributed.is_available():
        from torch.distributed._functional_collectives import (
            traceable_collective_remaps,
        )

        return traceable_collective_remaps
    return {}
Enter fullscreen mode Exit fullscreen mode

So that's what im doing here first.

The goal

Well the goal is find a way to fix it and hopyfully merge it. It's not as easy as it sounds to be honest. I have to back track a bit and work my way up. I'm guessing I'm starting at -200 right now and not even at 0 just because I need to learn dynamo and being able to dynamically re write with it.

The goal hopefully I can submit a PR by next week but we will see.
For now let's get going and learn dynamo!

Hopefully I can give good update by next week.

okay See you!

💖 💪 🙅 🚩
rabroldan
Ronald R

Posted on November 29, 2023

Join Our Newsletter. No Spam, Only the good stuff.

Sign up to receive the latest update from our blog.

Related