Version mismatch on latest libs and packs

[livid tree]

is there a version guide somewhere whcih shows matching version of trl, transformers etc with unlsoth? i use uv but even if there are no conflicts, latest unlsoth does not match with latest trl, transformers etc keeps on giving erroe about Attribute not exists for certain Trainer everytime even thought I am using only SFT.

[untold furnace]

There is currently no official, up-to-date compatibility matrix or version guide published by Unsloth that lists matching versions of Unsloth, TRL, Transformers, and other dependencies. This is a common pain point, and users have requested such a guide due to frequent breakages when mixing latest versions of these libraries. The Unsloth team and community often recommend using specific, known-working versions based on recent issues and discussions, but these are scattered and not centrally documented. For example, users have found that only certain combinations (e.g., trl==0.19.1 or 0.20.0, transformers==4.53.3, unsloth==2025.8.1) work reliably, and upgrading any of these can break compatibility, especially with SFTTrainer and GRPOTrainer. See the discussion in this issue and this discussion.

If you need a working setup, it's best to follow the specific version recommendations found in recent GitHub issues or to use the versions listed in the official Unsloth example notebooks. If you encounter errors like "Attribute not exists for certain Trainer," it's likely due to a version mismatch between Unsloth and its dependencies. The maintainers have acknowledged the need for stricter and clearer version control, but as of now, you must piece together compatible versions from community reports and issue threads. Would you like a step-by-step example of a known working version set for SFT fine-tuning?

Sources:

• Dependency Conflicts Between unsloth, unsloth_zoo, torch, triton, and transformers in Python 3.10 Environment
• TypeError: chunk(): argument 'input' (position 1) must be Tensor, not list during GRPO training

💡 **Hint:** Mention @RunLLM in the thread for followups.

[knotty trench]

not that i know of

but if you are running locally and need a stable working environment, i'd recommend you use our unsloth docker container. It's pre-installed , pre-tested and comes with working notebooks

https://hub.docker.com/r/unsloth/unsloth
and https://hub.docker.com/r/unsloth/unsloth-blackwell only if you have a blackwell architecture card

they aren't yet gpt-oss compatible though . This will happen later once all bugs are cleared . the driver is to always have a stable container environment and for users not to have to worry about dependencies but spend time on ML/AI

also if you don't mind , no need to tag people everytime 🙏 . We see your questions and someone will always attend to them

aside from containers, we should potentially pin dependencies on pypi releases. That might help up to a certain extent

[livid tree]

okay got it, thanks,
also, will remember not to mention anyone to avoid unnecessary notifications, my bad 😅