Skip to content
This repository was archived by the owner on Aug 7, 2024. It is now read-only.

Added initial compile tests for flat-parameter FSDP #215

Closed
wants to merge 6 commits into from

Conversation

awgu
Copy link

@awgu awgu commented Feb 14, 2024

Stack from ghstack (oldest at bottom):

The tests added in this PR test (1) compile with frame_count to check that we are compiling the expected number of functions and (2) numerics against manual data parallelism. We exclude checking activation checkpointing and FSDP's mixed precision in this PR.

There are existing tests in test_fsdp.py and test_fsdp_compile.py. I am not sure how we should consolidate. The tests in test_fsdp_compile.py seem somewhat obsolete since they do not test numerics.

awgu pushed a commit that referenced this pull request Feb 14, 2024
ghstack-source-id: 39228ff
Pull Request resolved: #215
@facebook-github-bot facebook-github-bot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Feb 14, 2024
@awgu awgu marked this pull request as ready for review February 14, 2024 22:54
Andrew Gu added 2 commits February 15, 2024 11:16
The tests added in this PR test (1) compile with `frame_count` to check that we are compiling the expected number of functions and (2) numerics against manual data parallelism (including using FSDP's mixed precision). These add ~42 seconds to the test time.

There are existing tests in `test_fsdp.py` and `test_fsdp_compile.py`. I am not sure how we should consolidate. The tests in `test_fsdp_compile.py` seem somewhat obsolete since they do not test numerics.

[ghstack-poisoned]
The tests added in this PR test (1) compile with `frame_count` to check that we are compiling the expected number of functions and (2) numerics against manual data parallelism (including using FSDP's mixed precision). These add ~42 seconds to the test time.

There are existing tests in `test_fsdp.py` and `test_fsdp_compile.py`. I am not sure how we should consolidate. The tests in `test_fsdp_compile.py` seem somewhat obsolete since they do not test numerics.

[ghstack-poisoned]
Andrew Gu added 2 commits February 16, 2024 07:36
The tests added in this PR test (1) compile with `frame_count` to check that we are compiling the expected number of functions and (2) numerics against manual data parallelism (including using FSDP's mixed precision). These add ~42 seconds to the test time.

There are existing tests in `test_fsdp.py` and `test_fsdp_compile.py`. I am not sure how we should consolidate. The tests in `test_fsdp_compile.py` seem somewhat obsolete since they do not test numerics.

[ghstack-poisoned]
The tests added in this PR test (1) compile with `frame_count` to check that we are compiling the expected number of functions and (2) numerics against manual data parallelism (including using FSDP's mixed precision). These add ~42 seconds to the test time.

There are existing tests in `test_fsdp.py` and `test_fsdp_compile.py`. I am not sure how we should consolidate. The tests in `test_fsdp_compile.py` seem somewhat obsolete since they do not test numerics.

[ghstack-poisoned]
@awgu awgu changed the title Added compile tests for flat-parameter FSDP Added initial compile tests for flat-parameter FSDP Feb 16, 2024
The tests added in this PR test (1) compile with `frame_count` to check that we are compiling the expected number of functions and (2) numerics against manual data parallelism (including using FSDP's mixed precision). These add ~42 seconds to the test time.

There are existing tests in `test_fsdp.py` and `test_fsdp_compile.py`. I am not sure how we should consolidate. The tests in `test_fsdp_compile.py` seem somewhat obsolete since they do not test numerics.

[ghstack-poisoned]
@awgu awgu closed this Feb 27, 2024
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Labels
CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants