Skip to content

[mergeability] Cluster cluster-43366-4: merged 1 PRs#2

Closed
evalstate wants to merge 10 commits intomainfrom
merge-cluster-cluster-43366-4-20260421165257
Closed

[mergeability] Cluster cluster-43366-4: merged 1 PRs#2
evalstate wants to merge 10 commits intomainfrom
merge-cluster-cluster-43366-4-20260421165257

Conversation

@evalstate
Copy link
Copy Markdown
Owner

Cluster: cluster-43366-4
Base: origin/main

Merged:

Skipped:

Failed:

  • None.

Notes:

Next steps:

evalstate pushed a commit that referenced this pull request Apr 23, 2026
* Support the new privacy model (second try)

This time I'm building the conversion scripts and everything from main branch, rather than building on top of existing GPT-OSS support (#2).

Tested by converting multiple checkpoints and comparing logits and predictions on several texts.

Signed-off-by: Mihai Maruseac <mihaimaruseac@openai.com>

* Keep existing gpt_oss converter and duplicate rather than dry

Signed-off-by: Mihai Maruseac <mihaimaruseac@openai.com>

* Implement reviewer feedback

This is partial implementation as we also simplified the model since then. For example, there's no more need for scale tensors.

Still have to do documentation and some integration tests, but wanted a next round of review.

Signed-off-by: Mihai Maruseac <mihaimaruseac@openai.com>

* Add an integration test

Signed-off-by: Mihai Maruseac <mihaimaruseac@openai.com>

* Add a stub of documentation

Signed-off-by: Mihai Maruseac <mihaimaruseac@openai.com>

* more modular and interface friendly

* fixup

* fix

* fixup

* Make export still match logits and predictions with upstream model

Signed-off-by: Mihai Maruseac <mihaimaruseac@openai.com>

* alignments

* Migrate to newest model checkpoint

Signed-off-by: Mihai Maruseac <mihaimaruseac@openai.com>

* Remove some of the code that is no longer needed

Signed-off-by: Mihai Maruseac <mihaimaruseac@openai.com>

* Remove some of the code that is no longer needed

Signed-off-by: Mihai Maruseac <mihaimaruseac@openai.com>

* sync with latest changes

* keep sinks in fp32 - adjust FA and Flex to cast into half (less precise)

needs moe to be fixed

* try fix moe accumulation

* push to sync

* default to eager moe

* lets readd this after merge just for the signature

* last nits

* Apply suggestions from code review

Co-authored-by: Anton Vlasjuk <73884904+vasqu@users.noreply.github.com>

* Migrate to the proper name

Signed-off-by: Mihai Maruseac <mihaimaruseac@openai.com>

* Sort

Signed-off-by: Mihai Maruseac <mihaimaruseac@openai.com>

* Sort

Signed-off-by: Mihai Maruseac <mihaimaruseac@openai.com>

* Finalize documentation

Signed-off-by: Mihai Maruseac <mihaimaruseac@openai.com>

* small fixes

* style

---------

Signed-off-by: Mihai Maruseac <mihaimaruseac@openai.com>
Co-authored-by: Mihai Maruseac <mihaimaruseac@openai.com>
@evalstate evalstate closed this Apr 23, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants