fix(overlays/unstable): keep llama-cpp-multi green at llama-cpp 0.7 and 0.8

hermes commented

2026-11-01 18:59:14 +01:00

Contributor

0.7.0 made test-llama-context the generate-models fixture consumer. The exclusion list still named only the old state tests, so ctest re-added the excluded FIXTURES_SETUP to satisfy the included dependent, and its kolibri1 fixture failed checkPhase. Skip every consumer along with the setup.

0.8.0 ships native kolibri1 in llama-arch/model and a test fixture that writes its hparams; applying the full port there would insert duplicate enum, struct and case lines. Carry only the supplement: SIGMOID_LOGIT_ADD gating enum + graph semantics and the tokenizer-pre hunk. Derive both the patch and its ctest workaround from the channel version.

0.7.0 made test-llama-context the generate-models fixture consumer. The exclusion list still named only the old state tests, so ctest re-added the excluded FIXTURES_SETUP to satisfy the included dependent, and its kolibri1 fixture failed checkPhase. Skip every consumer along with the setup. 0.8.0 ships native kolibri1 in llama-arch/model and a test fixture that writes its hparams; applying the full port there would insert duplicate enum, struct and case lines. Carry only the supplement: SIGMOID_LOGIT_ADD gating enum + graph semantics and the tokenizer-pre hunk. Derive both the patch and its ctest workaround from the channel version.

hermes added 1 commit

2026-11-01 18:59:14 +01:00

fix(overlays/unstable): keep llama-cpp-multi green at llama-cpp 0.7 and 0.8

nix flake check / flake-check (pull_request) Failing after 4m15s

Details

fa2d578a1d

0.7.0 made test-llama-context the generate-models fixture consumer. The exclusion list still named only the old state tests, so ctest re-added the excluded FIXTURES_SETUP to satisfy the included dependent, and its kolibri1 fixture failed checkPhase. Skip every consumer along with the setup.

0.8.0 ships native kolibri1 in llama-arch/model and a test fixture that writes its hparams; applying the full port there would insert duplicate enum, struct and case lines. Carry only the supplement: SIGMOID_LOGIT_ADD gating enum + graph semantics and the tokenizer-pre hunk. Derive both the patch and its ctest workaround from the channel version.

hermes added 1 commit

2026-11-01 19:08:27 +01:00

chore(treefmt): the amd firewall port list

nix flake check / flake-check (pull_request) Successful in 4m24s

Details

bff7c34ae8

The base commit landed networking.firewall.interfaces."virbr-rialo" inline, so the repo-wide formatting gate fails every PR on that base. Reformat it so the overlay fix can pass flake-check.

hermes commented

2026-11-01 19:14:49 +01:00

Author

Contributor

Diagnose: the nightly failure is a stale disabledTests list, not a broken kolibri patch.
tests/CMakeLists.txt at llama.cpp 0.7.0 has test-generate-models as FIXTURES_SETUP generate-models, with test-llama-context as FIXTURES_REQUIRED. The old exclusion list named only the two 0.6 state consumers, so ctest re-added the excluded setup for the included dependent; it writes a kolibri1 GGUF without expert_feed_forward_length, which the port's loader rejects.

Fix: add test-llama-context to the exclusion set and derive patch/exclusions from the channel version. 0.8.0 ships native kolibri1, so only the SIGMOID_LOGIT_ADD gating and tokenizer-pre supplement is carried there; the full port would duplicate enum/struct/case lines.

Verification: at the channel tip rev 1c89083e451e891152365b799595cf7790a43325 the actual overlay llama-cpp-multi (CUDA+ROCm, drv d9ksvb2fqzzx8f53rbcysv6wnwd3fhdg) built green: 100% tests passed out of 46. At nixpkgs master 0.8.0 the supplement gives 100% tests passed out of 50 incl. test-generate-models/test-llama-context Passed. Local nix flake check . green.

The extra chore(treefmt) commit only stands up the repo-wide formatting gate: base 89c31315 landed the amd virbr-rialo port list inline.

Diagnose: the nightly failure is a stale `disabledTests` list, not a broken kolibri patch. `tests/CMakeLists.txt` at llama.cpp 0.7.0 has `test-generate-models` as `FIXTURES_SETUP generate-models`, with `test-llama-context` as `FIXTURES_REQUIRED`. The old exclusion list named only the two 0.6 state consumers, so ctest re-added the excluded setup for the included dependent; it writes a kolibri1 GGUF without `expert_feed_forward_length`, which the port's loader rejects. Fix: add `test-llama-context` to the exclusion set and derive patch/exclusions from the channel version. 0.8.0 ships native kolibri1, so only the `SIGMOID_LOGIT_ADD` gating and tokenizer-pre supplement is carried there; the full port would duplicate enum/struct/case lines. Verification: at the channel tip rev `1c89083e451e891152365b799595cf7790a43325` the actual overlay `llama-cpp-multi` (CUDA+ROCm, drv `d9ksvb2fqzzx8f53rbcysv6wnwd3fhdg`) built green: `100% tests passed out of 46`. At nixpkgs master 0.8.0 the supplement gives `100% tests passed out of 50` incl. `test-generate-models`/`test-llama-context` Passed. Local `nix flake check .` green. The extra `chore(treefmt)` commit only stands up the repo-wide formatting gate: base `89c31315` landed the amd virbr-rialo port list inline.

nix flake check / flake-check (pull_request) Successful in 4m24s

Details

This pull request has changes conflicting with the target branch.

overlays/unstable/default.nix

View command line instructions

Manual merge helper

Use this merge commit message when completing the merge manually.

Merge commit title

Checkout

From your project repository, check out a new branch and test the changes.

git fetch -u origin +refs/pull/75/head:hermes/cache-warm-llama-fixture

git switch hermes/cache-warm-llama-fixture

Rows
Columns

fix(overlays/unstable): keep llama-cpp-multi green at llama-cpp 0.7 and 0.8 #75

Manual merge helper

Checkout