Change FP8 Eval to default to activation dtype #3454
Merged
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
Bumps [coverage[toml]](https://github.com/nedbat/coveragepy) from 7.5.3 to 7.5.4. - [Release notes](https://github.com/nedbat/coveragepy/releases) - [Changelog](https://github.com/nedbat/coveragepy/blob/master/CHANGES.rst) - [Commits](nedbat/coveragepy@7.5.3...7.5.4) --- updated-dependencies: - dependency-name: coverage[toml] dependency-type: direct:development update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Updates the requirements on [psutil](https://github.com/giampaolo/psutil) to permit the latest version. - [Changelog](https://github.com/giampaolo/psutil/blob/master/HISTORY.rst) - [Commits](giampaolo/psutil@release-5.8.0...release-6.0.0) --- updated-dependencies: - dependency-name: psutil dependency-type: direct:production ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Co-authored-by: Mihir Patel <mihir.v.patel7@gmail.com>
* Add support for variable length dataloaders in dist training * Remove test file * Fix typo * Fixed batch referenced before assignment * Replace sentinel with None * Add unit test * Update unit test * Reduce tensor creation to one line Co-authored-by: Mihir Patel <mihir.v.patel7@gmail.com> * Remove requirement for gpu in test --------- Co-authored-by: Mihir Patel <mihir.v.patel7@gmail.com>
* fold ema fsdp state * debug * debug * more debug * keep debugging * debug * sanity check * debug * debug * use ema * debug * debug * debug * debug * debug * debug * more fix * filename test * revert test * fully parameterize * hsdp test * revert testing * typo * typo * hsdp * split off test * precommit * float to int * pyright * oom * print * rm tp * tp cfg * tp? * rm tp line * type annotation * revert * readd tp * type * world size * revert * revert monolithic cpkt + include sharded cpkt * enumerate * precommit * precommit * sharded * sync * only sync on first trainer * typo * hsdp * xfail * explicit sync * test * revert test * sync, docker issue * pre-commit * sync * pytest * xfail * rm world_size param * im so sorry pls forgive me king * the kings comments * Update tests/trainer/test_fsdp_checkpoint.py fix formatting Co-authored-by: Mihir Patel <mihir.v.patel7@gmail.com> * precommit --------- Co-authored-by: v-chen_data <v-chen_data@example.com> Co-authored-by: Daniel King <43149077+dakinggg@users.noreply.github.com> Co-authored-by: Mihir Patel <mihir.v.patel7@gmail.com>
* test * skip if torch version less than 2.3 * typo in ema * add remote * comments --------- Co-authored-by: v-chen_data <v-chen_data@example.com>
* remove save overwrite * fix tests * lint * remove bad test
…load (#3436) * lower the system metrics logging frequency * more frequent
Updates the requirements on [paramiko](https://github.com/paramiko/paramiko) to permit the latest version. - [Commits](paramiko/paramiko@2.11.0...3.4.0) --- updated-dependencies: - dependency-name: paramiko dependency-type: direct:development ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
* test * test * test * test * test * test * fix * sleep before skip * fix * pull request target * revert * revery pr_target branches * sleep 1 * 10 sec * uncomment * dist barrier * test * dist works! * update 0.0.9 * mihir comment Co-authored-by: Mihir Patel <mihir.v.patel7@gmail.com> --------- Co-authored-by: v-chen_data <v-chen_data@example.com> Co-authored-by: Mihir Patel <mihir.v.patel7@gmail.com>
added 2 commits
July 10, 2024 17:55
added 3 commits
July 10, 2024 18:06
mvpatel2000
reviewed
Jul 11, 2024
mvpatel2000
approved these changes
Jul 11, 2024
j316chuck
changed the title
j316chuck
changed the title
j316chuck
changed the title
j316chuck
changed the title
mvpatel2000
pushed a commit
to mvpatel2000/composer
that referenced
this pull request
Jul 21, 2024
mvpatel2000
pushed a commit
that referenced
this pull request
Jul 21, 2024
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
What does this PR do?
Turn off te fp8 autocast for eval. This logic works around the incompatability where composer eval batches can be of any arbitrary shape whereas TE requires that eval batches are divisible by 16.
What issue(s) does this change relate to?
https://databricks.atlassian.net/browse/GRT-3023
Unit Tests
WORLD_SIZE=1 EXTRA_ARGS='-k test_amp_fp8_eval_casts_to_bf16' make test-dist-gpu✅WORLD_SIZE=1 EXTRA_ARGS='-k test_amp_fp8_eval_casts_to_bf16' make test-dist-gpu✔️ (skipped)Regression Tests
fp8-llama3-8b-metamath-4ep-3AOZfj🔴 (fails on model eval)fp8-llama3-8b-metamath-4ep-tIKVb1✅ (works on model eval)Along for the ride