[Bug] Unsloth Studio Context Length Resets on Model Load/Startup

[TheAlex25]

Bug Report: Unsloth Studio Context Length Resets on Model Load/Startup

Status: New
Severity: Medium (Affects usability for long-context tasks)
Component: UI / Configuration Management

---

Summary

The "Run settings" in Unsloth Studio do not persist across sessions or model reloads. Specifically, the Context Length resets to the default value of 4096 tokens every time a model is loaded (manually or automatically), even if it was previously configured and saved at a higher value (e.g., 32768). Interestingly, "Sampling settings" persist correctly across sessions, suggesting the issue is isolated to the Run settings configuration logic.

Environment

• OS: Ubuntu 26.04
• Unsloth Studio Version: v0.1.471-beta
• Package Version: 2026.6.9
• llama.cpp Version: b9860-mix-c19e218

Steps to Reproduce

1. Launch Unsloth Studio.
2. Load a model.
3. Navigate to Run settings.
4. Change the Context Length from 4096 to 32768.
5. Save the configuration or restart the application/reload the model.
6. Observe the Context Length setting upon reload.

Expected Behavior

The Context Length should remain at the user-defined value (32768) when a model is loaded manually or via auto-load, matching the behavior of the "Sampling settings."

Actual Behavior

The Context Length reverts to 4096 tokens regardless of previous manual adjustments.

Note: The UI correctly identifies that higher context lengths exceed VRAM and suggests using system RAM for KV cache, but it fails to retain the setting itself.

---

Feature Requests & Suggestions

1. Persistence for Run Settings: Ensure that all "Run settings" (Context Length, KV Cache Dtype, etc.) are saved and restored alongside Sampling settings when a model is initialized.
2. Per-Model Configuration Profiles: Implement the ability to save specific configurations on a per-model basis. Currently, it appears settings are either global or failing to bind to specific model instances upon reload. This would allow users to have high context for large models and low context/high speed for smaller ones without manual readjustment every time.

Attachments

• (Refer to provided screenshots showing the discrepancy between the 4096 default state and the intended 32768 state with VRAM warning.)

[Sneakr]

Thanks for reporting. We are doing some rework on per model configs here #6647 which will be updated soon.

[TheAlex25]

Thanks for reporting. We are doing some rework on per model configs here #6647 which will be updated soon.

You're welcome. So, this issue would most likely be fixed in a future version of Unsloth? I do not use nightly builds, but will wait for the next version. Thank you.

[Sneakr]

You're welcome. So, this issue would most likely be fixed in a future version of Unsloth? I do not use nightly builds, but will wait for the next version. Thank you.

Correct! Thanks