Add Muse Glimmer chat template - #6739
Merged
Merged
Conversation
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
albertvillanova
approved these changes
Aug 17, 2026
qgallouedec
force-pushed
the
add-muse-glimmer-chat-template
branch
from
August 18, 2026 03:22
8d5e1c3 to
a32797c
Compare
This file contains hidden or 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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Stacked on
add-muse-glimmer.muse_glimmer.jinja(as shipped by meta-models/Muse-Glimmer-30B) andmuse_glimmer_training.jinja, registered inget_training_chat_template.{% generation %}/{% endgeneration %}, same treatment asgptoss_training.jinja, which shares the<|start|>assistantcue and one-block-per-channel layout. Rendering is byte-identical to the original.add_response_schemaentry: the checkpoint ships its ownresponse_template, which TRL already leaves alone.test_chat_template_utils.py.test_parse_response_tool_call_with_contentgets a Muse Glimmer branch next to the existing GPT-OSS one — a turn with tool calls has noto=userchannel, so text alongside a tool call belongs in theto=selfchannel.pytest tests/test_chat_template_utils.py -k muse_glimmer: 21 passed, 1 skipped.Note
Low Risk
Additive templates, registry wiring, and tests only; no changes to training loops or auth/data paths.
Overview
Adds Muse Glimmer as a supported chat-template family so TRL can recognize the model and swap in a training template for SFT/GRPO.
Vendored
muse_glimmer.jinja(ATEM:to=selfreasoning, per-tool channels with<atem:function_calls>,to=usercontent, multimodal placeholders) andmuse_glimmer_training.jinja, which wraps the full assistant turn—all ATEM channels—in{% generation %}/{% endgeneration %}(same pattern as GPT-OSS).get_training_chat_templatemaps the reference template to the training variant; docs list Muse Glimmer among supported families.Tests cover tool calling, training-template behavior, and response parsing (including
reasoning_contentalongside tool calls, since tool turns have noto=userchannel). Noadd_response_schemahook—the checkpoint already shipsresponse_template.Reviewed by Cursor Bugbot for commit dd63c49. Bugbot is set up for automated code reviews on this repo. Configure here.