deepspeedai / deepspeedai/DeepSpeed

[request] more intutive api for GatheredParameters read_only vs. write

Open
#1,046 2 comments 1 reaction 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
43.1k
Forks
5k
Avg merge
4d 15h
Merged PRs (30d)
112

Description

Currently one has to remember to add modifier_rank=0 to modify params.

deepspeed.zero.GatheredParameters(params, modifier_rank=0)

this is very unintuitive API IMHO and it won't fail if you forget to add it. Often that leads to subtle errors.

I already started using modifier_rank=None for read-only when I don't need to set it - just to remind me that it's a crucial arg to think about, when copy-n-pasting some code.

Could we change this API to something more human meaningful and more difficult to forget to set?

Here are a few possible ideas:

  1. add a required named flag, e.g. modify, so now we will have:
# just going to use the data 
deepspeed.zero.GatheredParameters(params, modify=False, modifier_rank=None)
# going to use and modify data
deepspeed.zero.GatheredParameters(params, modify=True, modifier_rank=0)

and if modify==True it will check that modifier_rank is enforced? Actually it can set it automatically

  • modify=True => modifier_rank=0 (unless it was set explicitly)
  • modify=False => modifier_rank=None ( asserts if non-None was passed

since that's the case most of the time.

So the API becomes a very simple and easy to use:

# just going to use the data 
deepspeed.zero.GatheredParameters(params, modify=False)
# going to use and modify data
deepspeed.zero.GatheredParameters(params, modify=True)

and of course one can override modifier_rank for the modifying case.

I'm totally open to perhaps even more intuitive names.

  1. Perhaps read_only is a better one?
# just going to use the data 
deepspeed.zero.GatheredParameters(params, read_only=True)
# going to use and modify data
deepspeed.zero.GatheredParameters(params, read_only=False)

It won't break any back-compat API if we make it optional, but will be helpful to always use in the code. But of course best to make it required - though it'd not be back-compat.

  1. Yet another idea, adding 2 wrappers:
# just going to use the data 
deepspeed.zero.GatheredParametersRead(params)
# going to use and modify data
deepspeed.zero.GatheredParametersWrite(params)

Even simpler and more straightforward.

I like this last idea the most as it's the most readable, and one can still use the current API if they want to.

@samyam, @jeffra, @ShadenSmith

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by locating the GatheredParameters entry point and reviewing how modifier_rank distinguishes read-only from modifying use. Compare the proposed modify, read_only, and wrapper APIs against existing callers and backward-compatibility needs. Done means an agreed API direction with clear behavior for defaults and modifier_rank overrides.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
distributed-systems, machine-learning
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Mostly clear
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.