Note: This article is adapted from the official Phala Network announcement.
If you’ve spent time fine-tuning large language models (LLM), you already know the two things that slow everything down. it’s definitely running out of VRAM and waiting on training runs that take longer than they should.
Unsloth by Phala cloud was built to fix both of those problems. It reduces memory usage significantly during fine-tuning and speeds up the process without requiring you to change your entire setup. For developers and research teams working with real hardware constraints, that combination is genuinely useful.
The challenge that doesn’t get talked about as much is what happens to your data during that process. Your training datasets, your adapter configurations, your tuning metadata, all of that has to live somewhere while the work is happening. In most setups, it’s sitting on infrastructure you don’t fully control, accessible to the host environment in ways that aren’t always visible to you. For institutions working with proprietary data or sensitive model configurations, that’s a real exposure point, not a theoretical one.
What Phala’s Template Actually Does
Phala released a deployment template that lets you run Unsloth directly on Phala Cloud inside a Confidential Virtual Machine. A CVM is a hardware-level isolated environment where your workload runs in a protected enclave. The host machine cannot read what’s happening inside it. Your datasets and configurations stay private by default, not because of a policy but because of how the hardware is structured. You get the performance benefits of Unsloth and the data protection of confidential compute in one deployment.
Why This Matters for Institutions and Developers
For individual developers, the practical benefit is straightforward. You can fine-tune models on sensitive data without routing it through infrastructure you don’t own or can’t verify. For institutions, the implications go further. Compliance requirements around data handling are tightening, particularly in regulated industries, and being able to demonstrate that training data never left a protected environment is increasingly something that needs to be documented, not assumed.
This is also a good example of how Phala is building in a way that’s different from most infrastructure projects. Rather than offering privacy as a feature you configure on top of an existing setup, they’re making it the default condition of the environment itself. The tooling connects to workflows developers are already using, which lowers the barrier to actually adopting it rather than just reading about it.
How to Get Started
The template is live at https://cloud.phala.com/templates/agno and the full code is available on Phala GitHub: https://github.com/Phala-Network/phala-cloud/tree/main/templates/prebuilt/agno and The upstream Unsloth is also open source at https://github.com/agno-agi/agno
If you’re already running fine-tuning workloads and want to understand what confidential compute actually adds to that process, this template is a straightforward way to see it in practice.
Top comments (0)