Review whether the ollama-cuda pin can be removed safely #101

New issue

Open

opened 2026-07-27 12:39:02 +01:00 by nimmo · 0 comments

nimmo commented

2026-07-27 12:39:02 +01:00

Owner

Context

The repository currently pins ollama-cuda via the dedicated nixpkgs-ollama-cuda flake input because the main nixpkgs update to e318e6adc497b5e8a5ac433c6c6ed21ecdab9029 broke CUDA-backed Ollama builds.

Observed failure during Lyra/Electra-related builds:

Could not find nvcc executable in path specified by environment variable CUDAToolkit_ROOT=...
CUDA Toolkit not found

To keep NVIDIA-backed Ollama buildable, only pkgs.ollama-cuda is overlaid from the pinned nixpkgs revision 421eebfd0ec7bccd4abe826ce62d7e6e83129493.

Review task

Test pkgs.ollama-cuda from the main nixpkgs input on Lyra.
Confirm whether Electra's disabled dgpu Ollama path would also evaluate/build correctly with upstream pkgs.ollama-cuda.
If upstream is fixed, remove the nixpkgs-ollama-cuda pin and overlay.
Update lib/auto-update-policy.nix, flake.nix, flake.lock, and README to remove references to the pin.

Exit criteria

Lyra builds successfully using upstream pkgs.ollama-cuda from the main nixpkgs input.
Electra's dgpu Ollama configuration is safe to re-enable later without relying on the pinned package.
The dedicated nixpkgs-ollama-cuda input and overlay are removed cleanly.
Auto-update policy and docs no longer mention the pin.

## Context The repository currently pins `ollama-cuda` via the dedicated `nixpkgs-ollama-cuda` flake input because the main `nixpkgs` update to `e318e6adc497b5e8a5ac433c6c6ed21ecdab9029` broke CUDA-backed Ollama builds. Observed failure during Lyra/Electra-related builds: - `Could not find nvcc executable in path specified by environment variable CUDAToolkit_ROOT=...` - `CUDA Toolkit not found` To keep NVIDIA-backed Ollama buildable, only `pkgs.ollama-cuda` is overlaid from the pinned nixpkgs revision `421eebfd0ec7bccd4abe826ce62d7e6e83129493`. ## Review task - Test `pkgs.ollama-cuda` from the main `nixpkgs` input on Lyra. - Confirm whether Electra's disabled dgpu Ollama path would also evaluate/build correctly with upstream `pkgs.ollama-cuda`. - If upstream is fixed, remove the `nixpkgs-ollama-cuda` pin and overlay. - Update `lib/auto-update-policy.nix`, `flake.nix`, `flake.lock`, and README to remove references to the pin. ## Exit criteria - Lyra builds successfully using upstream `pkgs.ollama-cuda` from the main `nixpkgs` input. - Electra's dgpu Ollama configuration is safe to re-enable later without relying on the pinned package. - The dedicated `nixpkgs-ollama-cuda` input and overlay are removed cleanly. - Auto-update policy and docs no longer mention the pin.

nimmo referenced this issue from a commit

2026-07-27 12:44:17 +01:00

Pin ollama-cuda and disable Electra dgpu Ollama

nimmo referenced this issue

2026-07-27 12:46:13 +01:00

Pin ollama-cuda and disable Electra dgpu Ollama #102