This repository was archived by the owner on Aug 7, 2024. It is now read-only.
[1/x]: Make Float8Linear support dynamic scaling #290
Closed
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Summary: At a high level, we need to make dynamic vs delayed scaling configurable separately for activations, weights and gradients. The way I am approaching this is as follows: * PR 1 (this PR): add basic support for dynamic scaling, configurable by tensor, to `Float8Linear` * PRs 2..n: one by one, add features implemented in `Float8DynamicLinear` to `Float8Linear`, as necessary * last PR: delete `Float8DynamicLinear` Test Plan: ``` ./test/test_everything.sh ``` Reviewers: Subscribers: Tasks: Tags: [ghstack-poisoned]
vkuzo
added a commit
that referenced
this pull request
Jun 28, 2024
Summary: At a high level, we need to make dynamic vs delayed scaling configurable separately for activations, weights and gradients. The way I am approaching this is as follows: * PR 1 (this PR): add basic support for dynamic scaling, configurable by tensor, to `Float8Linear` * PRs 2..n: one by one, add features implemented in `Float8DynamicLinear` to `Float8Linear`, as necessary * last PR: delete `Float8DynamicLinear` Test Plan: ``` ./test/test_everything.sh ``` Reviewers: Subscribers: Tasks: Tags: ghstack-source-id: 7020d5e88e29710c77baaf8c95ff1d859c91bebd Pull Request resolved: #290
facebook-github-bot
added
the
CLA Signed
vkuzo
commented
Jun 28, 2024
Summary: At a high level, we need to make dynamic vs delayed scaling configurable separately for activations, weights and gradients. The way I am approaching this is as follows: * PR 1 (this PR): add basic support for dynamic scaling, configurable by tensor, to `Float8Linear` * PRs 2..n: one by one, add features implemented in `Float8DynamicLinear` to `Float8Linear`, as necessary * last PR: delete `Float8DynamicLinear` Test Plan: ``` ./test/test_everything.sh ``` Reviewers: Subscribers: Tasks: Tags: [ghstack-poisoned]
vkuzo
added a commit
that referenced
this pull request
Jun 28, 2024
Summary: At a high level, we need to make dynamic vs delayed scaling configurable separately for activations, weights and gradients. The way I am approaching this is as follows: * PR 1 (this PR): add basic support for dynamic scaling, configurable by tensor, to `Float8Linear` * PRs 2..n: one by one, add features implemented in `Float8DynamicLinear` to `Float8Linear`, as necessary * last PR: delete `Float8DynamicLinear` Test Plan: ``` ./test/test_everything.sh ``` Reviewers: Subscribers: Tasks: Tags: ghstack-source-id: 7020d5e88e29710c77baaf8c95ff1d859c91bebd Pull Request resolved: #290
Summary: At a high level, we need to make dynamic vs delayed scaling configurable separately for activations, weights and gradients. The way I am approaching this is as follows: * PR 1 (this PR): add basic support for dynamic scaling, configurable by tensor, to `Float8Linear` * PRs 2..n: one by one, add features implemented in `Float8DynamicLinear` to `Float8Linear`, as necessary * last PR: delete `Float8DynamicLinear` Test Plan: ``` ./test/test_everything.sh ``` Reviewers: Subscribers: Tasks: Tags: [ghstack-poisoned]
vkuzo
added a commit
that referenced
this pull request
Jun 28, 2024
Summary: At a high level, we need to make dynamic vs delayed scaling configurable separately for activations, weights and gradients. The way I am approaching this is as follows: * PR 1 (this PR): add basic support for dynamic scaling, configurable by tensor, to `Float8Linear` * PRs 2..n: one by one, add features implemented in `Float8DynamicLinear` to `Float8Linear`, as necessary * last PR: delete `Float8DynamicLinear` Test Plan: ``` ./test/test_everything.sh ``` Reviewers: Subscribers: Tasks: Tags: ghstack-source-id: 0aeb524df8b1b693f0538a85c898d24f3f347081 Pull Request resolved: #290
Summary: At a high level, we need to make dynamic vs delayed scaling configurable separately for activations, weights and gradients. The way I am approaching this is as follows: * PR 1 (this PR): add basic support for dynamic scaling, configurable by tensor, to `Float8Linear` * PRs 2..n: one by one, add features implemented in `Float8DynamicLinear` to `Float8Linear`, as necessary * last PR: delete `Float8DynamicLinear` Test Plan: ``` ./test/test_everything.sh ``` Reviewers: Subscribers: Tasks: Tags: [ghstack-poisoned]
vkuzo
added a commit
that referenced
this pull request
Jun 28, 2024
Summary: At a high level, we need to make dynamic vs delayed scaling configurable separately for activations, weights and gradients. The way I am approaching this is as follows: * PR 1 (this PR): add basic support for dynamic scaling, configurable by tensor, to `Float8Linear` * PRs 2..n: one by one, add features implemented in `Float8DynamicLinear` to `Float8Linear`, as necessary * last PR: delete `Float8DynamicLinear` Test Plan: ``` ./test/test_everything.sh ``` Reviewers: Subscribers: Tasks: Tags: ghstack-source-id: bb1bfe5234f6b654bc8e52d48a158d8b1b3bd616 Pull Request resolved: #290
drisspg
reviewed
Jun 28, 2024
drisspg
reviewed
Jun 28, 2024
| """Returns whether the given linear_type requires sync before forward.""" | ||
| return linear_type in REQUIRES_SYNC | ||
| return linear_type is LinearType.DELAYED and any( |
There was a problem hiding this comment.
since Float8DynamicLinear does not support TensorScalingType, I think and is right?
drisspg
reviewed
Jun 28, 2024
| emulate=emulate, | ||
| ) | ||
| if linear_type is LinearType.DYNAMIC: | ||
| return Float8DynamicLinear.from_float( |
There was a problem hiding this comment.
should we assert that non of the scaling types are True?
There was a problem hiding this comment.
technically yes, I was just lazy and unmotivated since this stack is trying to delete Float8DynamicLinear
drisspg
reviewed
Jun 28, 2024
| def linear_requires_sync(linear_type: LinearType): | ||
| def linear_requires_sync( | ||
| linear_type: LinearType, | ||
| scaling_type_x: TensorScalingType = TensorScalingType.DELAYED, |
There was a problem hiding this comment.
Nit: we should probably remove the defaults right?
There was a problem hiding this comment.
hmm, long term probably yes. I'm taking the approach of tackling the "what's the default" recipe separately and not changing default behavior for now.
Summary: At a high level, we need to make dynamic vs delayed scaling configurable separately for activations, weights and gradients. The way I am approaching this is as follows: * PR 1 (this PR): add basic support for dynamic scaling, configurable by tensor, to `Float8Linear` * PRs 2..n: one by one, add features implemented in `Float8DynamicLinear` to `Float8Linear`, as necessary * last PR: delete `Float8DynamicLinear` Test Plan: ``` ./test/test_everything.sh ``` Reviewers: Subscribers: Tasks: Tags: [ghstack-poisoned]
This was referenced Jul 1, 2024
|
@vkuzo has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
This pull request has been merged in 3cb42e1. |
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Stack from ghstack (oldest at bottom):
Summary:
At a high level, we need to make dynamic vs delayed scaling configurable
separately for activations, weights and gradients. The way I am
approaching this is as follows:
Float8LinearFloat8DynamicLineartoFloat8Linear, as necessaryFloat8DynamicLinearTest Plan:
Reviewers:
Subscribers:
Tasks:
Tags:
Differential Revision: D59305792