[GH-ISSUE #3498] Teflon (a new part of Mesa on Linux) NPU delegate support #2154

Open
opened 2026-04-12 12:22:57 -05:00 by GiteaMirror · 1 comment
Owner

Originally created by @leaf-node on GitHub (Apr 5, 2024).
Original GitHub issue: https://github.com/ollama/ollama/issues/3498

What are you trying to do?

Teflon is a new front end library for NPU acceleration part of the latest versions of Mesa on Linux. It's in early stages, but more NPU drivers may be added in the future.

I'd like to use ollama with an NPU via an open source stack when a device that I own is supported by Mesa.

How should we solve this?

It hope to see Teflon support in ollama. According to the Teflon documentation, it's already supported in TensorFlow Lite as a "delegate".

What is the impact of not solving this?

When an upcoming Arm based processor with strong NPU compute capabilities appears on the market, Teflon support for that device may eventually be added to Mesa. If ollama doesn't support Teflon, ollama may miss out on the opportunity to be an early adopter of this open source general purpose API.

Anything else?

Thanks! : )

Originally created by @leaf-node on GitHub (Apr 5, 2024). Original GitHub issue: https://github.com/ollama/ollama/issues/3498 ### What are you trying to do? [Teflon](https://docs.mesa3d.org/teflon.html) is a [new](https://www.phoronix.com/news/Gallium3D-Teflon-Merged) front end library for NPU acceleration part of the latest versions of [Mesa](https://www.mesa3d.org/) on Linux. It's in early stages, but more NPU drivers may be added in the future. I'd like to use ollama with an NPU via an open source stack when a device that I own is supported by Mesa. ### How should we solve this? It hope to see Teflon support in ollama. According to the Teflon documentation, it's already supported in TensorFlow Lite as a "delegate". ### What is the impact of not solving this? When an upcoming Arm based processor with strong NPU compute capabilities appears on the market, Teflon support for that device may eventually be added to Mesa. If ollama doesn't support Teflon, ollama may miss out on the opportunity to be an early adopter of this open source general purpose API. ### Anything else? Thanks! : )
Author
Owner

@tomeuv commented on GitHub (Nov 21, 2024):

Teflon author here. I would need some help to understand through which mechanism could ollama delegate the execution of some or all layers to Mesa drivers. Teflon only exposes the TFLite delegate API, which I don't know if ollama would want to use. But other frontends similar to Teflon could be added to Mesa, to give frameworks access to the HW-specific NPU drivers.

<!-- gh-comment-id:2490744632 --> @tomeuv commented on GitHub (Nov 21, 2024): Teflon author here. I would need some help to understand through which mechanism could ollama delegate the execution of some or all layers to Mesa drivers. Teflon only exposes the TFLite delegate API, which I don't know if ollama would want to use. But other frontends similar to Teflon could be added to Mesa, to give frameworks access to the HW-specific NPU drivers.
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#2154