Hi team,
We would like to release the gemma4 snap as stable. This is an inference snap that installs a silicon-optimized Gemma 4 engine for local inference. To function, this snap requires access similar to the gemma3 snap. The software stack is identical to the other inference snaps.
- name: gemma4
- description: An inference snap, installing a silicon-optimized gemma4 engine
- snapcraft: gemma4-snap/snap/snapcraft.yaml at main · canonical/gemma4-snap · GitHub
- upstream: GitHub - canonical/gemma4-snap: Local inference with Gemma4 · GitHub
- upstream-relation: -
- interfaces:
- hardware-observe:
- request-type: auto-connection
- reasoning: To detect the host hardware (CPU, accelerator, etc) and install the right engine for AI inference
- home: + read all
- request-type: installation, manual connection with
read: allattribute - reasoning: The engine that runs inside as a root daemon requires access to other users’ home for loading local models based on user configuration. We expect the user to manually connect
homeif they need to sideload a model.
- request-type: installation, manual connection with
- process-control:
- request-type: auto-connection
- reasoning: ROCm requires control of processor affinity to ensure that ROCm helper threads run on the correct cores. Anything that loads the ROCm libraries fails without it.
- hardware-observe:
The gemma3 snap has already been reviewed for the same set of interfaces:
- Auto-connection of hardware-observe and home for gemma3 - #3 by jslarraz
- Autoconnect request: process-control for gemma3-snap
Thanks for your support!