Repository navigation
Releases
Releases · modal-labs/modal-dojo
Modal Dojo 0.1.3
Modal Dojo 0.1.2
This release:
- Adds a converter from Miles native-LoRA checkpoints to PEFT adapters (#757)
- Adds a Reflection Beam 256k Miles LoRA recipe (#762)
- Adds a new login page for dashboards with basic auth (#741)
- Password-protects the dashboard by default (#740)
- Normalizes checkpoint conversion across Slime and Miles (#745, #764)
- Renames Slime conversion layout flags (#746)
- Improves handling of escape hatch arguments (#774, #775)
- Separates next-rollout waits and post-training evals in async Miles timing (#802)
- Fixes the base Miles image not resolving (#809)
- Fails fast on TorchMemorySaver allocator conflicts and invalid KL/over-sampling settings (#756)
- Retries checkpoint Volume commits within the commit timeout (#686)
- Fixes endpoint launching from Jupyter (#787)
- Other recipe tuneups, dependency upgrades, bug fixes, and stability improvements
New Contributors
- Eric Zhang (@ekzhang) made their first contribution in #791
- Fred Robinson (@frrad) made their first contribution in #787
- Jason Mancuso (@jvmncs) made their first contribution in #807
Full Changelog: v0.1.1...v0.1.2
Modal Dojo 0.1.1
This release:
- Makes endpoints and custom deployments authenticated using proxy tokens by default
- Updates the SGLang image to v0.5.22
- Adds recipes for DeepSeek V4.1 Flash LoRA and Muse Spark Open 64K
- Tunes the Qwen3.6-35B recipe's async layout
- Updates GLM-5.3-Flash 32k to native Miles LoRA
- Reuses converted checkpoints across parallelism layouts on Slime
- Adds explicit error handling for native LoRA resume scheduler mismatches
Modal Dojo 0.1.0
Initial release of the Modal Dojo. Install with:
uv add "modal-dojo==0.1.0"