Chore - Bump llama.cpp 1.0

.agents/skills/bump-engines/SKILL.md

@@ -15,10 +15,11 @@ Each engine's `engines/<engine>/AGENTS.md` owns its pin, release id convention a
 
 4. Sweep for the old release id, version and digest. Keep measurement-provenance comments in model files. A bump that changes no flag never needs backend test edits; if one does, the test asserts a committed declaration, which is a coupling bug to fix. When upstream forces an adapter change (step 2), its tests change with it.
 
-5. Qualify on hardware. Flag audits miss default-behaviour changes, which only the gated tests and real deployments catch, so this step is required:
-   * Run `just backend::test-nvidia` on an amd64 NVIDIA host and on GB10.
-   * Deploy each changed engine's profiles on every platform it supports, each on a host whose memory fits its per-device minimum. A row passes when it reaches Ready, answers two completions, shows draft acceptance above 0 where speculative, and launches with the argv the adapter produces.
-   * Report sizing estimates that read low by more than 5% or 500 MiB, measured with the check of the `model-sizing` skill, and a largest fit that doesn't start or lands far below the engine's real start limit; don't refit sizing in a bump unless upstream changed how the engine holds memory, which needs the user's approval.
+5. Qualify on every platform the engine runs on (root `AGENTS.md` "Platform scope"), macOS included. Flag audits miss default-behaviour changes, which only the hardware tests and real deployments catch, so this step is required:
+   * Run the hardware tests in the engine's `AGENTS.md` "Validate" on a host of each platform.
+   * Deploy each changed engine's profiles on one platform, each on a host whose memory fits its per-device minimum. On the other platforms, deploy only the profiles most exposed to the audited upstream changes (drafts, the engine's memory paths) and any that read low; full sweeps cost hours of hardware and many GB of downloads. A row passes when it reaches Ready, answers two completions, shows draft acceptance above 0 where speculative, and launches with the argv the adapter produces.
+   * Report sizing estimates that read low by more than 5% or 500 MiB, measured with the check of the `model-sizing` skill, and a largest fit that doesn't start or lands far below the engine's real start limit; don't refit sizing in a bump unless upstream changed how the engine holds memory, which needs the user's approval. A model that only overruns its margin gets, with approval, a small offset fitted to the measurements already taken, not a new sizing campaign.
    * Run remote hosts one at a time, with one remote command per row: an ssh agent that prompts for approval rejects concurrent signing.
+   * Check a host's free disk before each new model download. The model cache is shared with other workspaces, so report a nearly full disk to the user instead of pruning it.
 
 6. Fix any `AGENTS.md` step that proved wrong.

backend/core/src/model/engine/llama_cpp/binding.rs

@@ -56,7 +56,7 @@ fn classify(rejections: &[(u16, &str)], status: u16, body: &[u8]) -> Option<Inva
 /// message format, where `%d` stands for a token count.
 ///
 /// A prompt larger than the micro-batch is a server error, and one larger than a slot's context
-/// is a context-size error (b11723 `tools/server/server-context.cpp`).
+/// is a context-size error (b12359 `tools/server/server-context.cpp`).
 const OVERSIZED_PROMPT_REJECTIONS: [(u16, &str); 2] = [
     (
         500,
@@ -69,7 +69,7 @@ const OVERSIZED_PROMPT_REJECTIONS: [(u16, &str); 2] = [
 ];
 
 /// llama.cpp's refusal of a chat prompt larger than a slot's context: the HTTP status it answers
-/// with and its message format (b11723 `tools/server/server-context.cpp`).
+/// with and its message format (b12359 `tools/server/server-context.cpp`).
 const OVERSIZED_CHAT_REJECTIONS: [(u16, &str); 1] = [(
     400,
     "request (%d tokens) exceeds the available context size (%d tokens), try increasing it",

backend/core/src/model/resolved.rs

@@ -1255,7 +1255,7 @@ mod tests {
                 inputs.hardware_class_id.as_str(),
             ),
             (
-                "llama-cpp-2026-10-19",
+                "llama-cpp-2026-11-13",
                 "metal",
                 dotcore_domain::test_support::nix::STORE_PATH,
                 "apple",

backend/core/src/test_support/nix.rs

@@ -25,7 +25,7 @@ pub(crate) fn apple_metal_snapshot(model_id: &str) -> ModelDeploymentSnapshot {
         panic!("Apple Metal fixture requires a local snapshot");
     };
     let engine = local.engine_artifact.as_mut().expect("snapshot engine");
-    engine.engine_release_id = "llama-cpp-2026-10-19".to_owned();
+    engine.engine_release_id = "llama-cpp-2026-11-13".to_owned();
     let mut artifact = EngineArtifact {
         selectors: Vec::new(),
         ..engine_artifact("metal", &locked_flake(), "darwin", "arm64", "apple")

backend/core/tests/gated/apple_hardware_integration.rs

@@ -129,11 +129,12 @@ fn hw_apple_gpu_is_observed_with_a_fixed_device_id_across_polls() {
         NodeAcceleratorObservation::from_hardware_snapshot("node-a", &snapshots[0], false),
         NodeAcceleratorObservation::available_at(
             "node-a",
-            NodeAcceleratorFacts::apple(),
+            NodeAcceleratorFacts::apple()
+                .with_device_memory(gpu.device_id.clone(), None, None)
+                .with_device_model(gpu.model.clone())
+                .with_host_memory(host_total, host_available),
             observed_at
         )
-        .with_device_memory(gpu.device_id.clone(), None, None)
-        .with_host_memory(host_total, host_available)
     );
 }
 

backend/fixtures/llama-cpp-nvidia-site/catalog/models/stories15m-split-qualification.yaml

@@ -21,7 +21,7 @@ artifacts:
 configurations:
   - configurationId: "nvidia-cuda"
     engineArtifacts:
-      - engineReleaseId: "llama-cpp-linux-2026-10-19"
+      - engineReleaseId: "llama-cpp-linux-2026-11-13"
         engineArtifactId: "cuda13"
     artifactBindings:
       - bindingId: "split-q8"

backend/site/src/node/target_inventory.rs

@@ -250,6 +250,17 @@ impl NodeAcceleratorFacts {
         self
     }
 
+    /// Names the product model the vendor reports for the device added last.
+    #[cfg(any(test, feature = "test-support"))]
+    pub fn with_device_model(mut self, model: impl Into<String>) -> Self {
+        self.memory
+            .devices
+            .last_mut()
+            .expect("a device to name")
+            .model = model.into();
+        self
+    }
+
     /// Replaces the host memory facts.
     #[cfg(any(test, feature = "test-support"))]
     pub fn with_host_memory(mut self, total_bytes: u64, available_bytes: u64) -> Self {

catalog/engines/llama-cpp-macos.yaml

@@ -2,9 +2,9 @@
 # flake as `engines.llama-cpp.native`. The packaged master catalog rewrites `self` to the locked Dotcore
 # revision with precomputed store paths, and fails unless the version equals the engine package
 # version (checked by `checks.dotcore-catalog`).
-engineReleaseId: "llama-cpp-2026-10-19"
+engineReleaseId: "llama-cpp-2026-11-13"
 engineKind: "ENGINE_KIND_LLAMA_CPP"
-version: "b11723"
+version: "b12359" # v1.0.0
 parameterSchema:
   schemaId: "llama-cpp-v2"
   parameters:

catalog/engines/llama-cpp.yaml

@@ -1,8 +1,8 @@
 # Each artifact names a Docker image derivation package; the package pins the exact image.
 # NVIDIA host integration testing remains an explicit responsibility.
-engineReleaseId: "llama-cpp-linux-2026-10-19"
+engineReleaseId: "llama-cpp-linux-2026-11-13"
 engineKind: "ENGINE_KIND_LLAMA_CPP"
-version: "b11723"
+version: "b12359" # v1.0.0
 # Same schema id as the macOS release, so one model configuration can approve artifacts of both.
 parameterSchema:
   schemaId: "llama-cpp-v2"

catalog/models/clef-flash.yaml

@@ -30,9 +30,9 @@ configurations:
     # The model card's encoder default; the weights accept up to the model's maximum context.
     maximumContextTokens: 16384
     engineArtifacts:
-      - engineReleaseId: "llama-cpp-linux-2026-10-19"
+      - engineReleaseId: "llama-cpp-linux-2026-11-13"
         engineArtifactId: "cuda13"
-      - engineReleaseId: "llama-cpp-2026-10-19"
+      - engineReleaseId: "llama-cpp-2026-11-13"
         engineArtifactId: "metal"
     artifactBindings:
       - bindingId: "q4-k-m"

catalog/models/gemma-4-12b-it-qat.yaml

@@ -40,8 +40,8 @@ configurations:
   # targets and the Metal binary on macOS native targets; both releases share a parameter schema.
   - configurationId: "llama-cpp"
     engineArtifacts:
-      - { engineReleaseId: "llama-cpp-linux-2026-10-19", engineArtifactId: "cuda13" }
-      - { engineReleaseId: "llama-cpp-2026-10-19", engineArtifactId: "metal" }
+      - { engineReleaseId: "llama-cpp-linux-2026-11-13", engineArtifactId: "cuda13" }
+      - { engineReleaseId: "llama-cpp-2026-11-13", engineArtifactId: "metal" }
     artifactBindings:
       - bindingId: "primary"
         modelArtifactId: "primary"

catalog/models/gemma-4-26b-a4b-it.yaml

@@ -106,9 +106,9 @@ configurations:
   # targets and the Metal binary on macOS native targets; both releases share a parameter schema.
   - configurationId: "llama-cpp"
     engineArtifacts:
-      - engineReleaseId: "llama-cpp-linux-2026-10-19"
+      - engineReleaseId: "llama-cpp-linux-2026-11-13"
         engineArtifactId: "cuda13"
-      - engineReleaseId: "llama-cpp-2026-10-19"
+      - engineReleaseId: "llama-cpp-2026-11-13"
         engineArtifactId: "metal"
     artifactBindings:
       - bindingId: "q3-xl"

catalog/models/gemma-4-31b-it.yaml

@@ -110,9 +110,9 @@ configurations:
   # targets and the Metal binary on macOS native targets; both releases share a parameter schema.
   - configurationId: "llama-cpp"
     engineArtifacts:
-      - engineReleaseId: "llama-cpp-linux-2026-10-19"
+      - engineReleaseId: "llama-cpp-linux-2026-11-13"
         engineArtifactId: "cuda13"
-      - engineReleaseId: "llama-cpp-2026-10-19"
+      - engineReleaseId: "llama-cpp-2026-11-13"
         engineArtifactId: "metal"
     artifactBindings:
       - bindingId: "q3-xl"

catalog/models/gemma-4-e2b.yaml

@@ -48,9 +48,9 @@ configurations:
   # targets and the Metal binary on macOS native targets; both releases share a parameter schema.
   - configurationId: "llama-cpp"
     engineArtifacts:
-      - engineReleaseId: "llama-cpp-linux-2026-10-19"
+      - engineReleaseId: "llama-cpp-linux-2026-11-13"
         engineArtifactId: "cuda13"
-      - engineReleaseId: "llama-cpp-2026-10-19"
+      - engineReleaseId: "llama-cpp-2026-11-13"
         engineArtifactId: "metal"
     artifactBindings:
       - bindingId: "q4-k-m"

catalog/models/gemma-4-e4b-it-qat.yaml

@@ -39,8 +39,8 @@ configurations:
   # targets and the Metal binary on macOS native targets; both releases share a parameter schema.
   - configurationId: "llama-cpp"
     engineArtifacts:
-      - { engineReleaseId: "llama-cpp-linux-2026-10-19", engineArtifactId: "cuda13" }
-      - { engineReleaseId: "llama-cpp-2026-10-19", engineArtifactId: "metal" }
+      - { engineReleaseId: "llama-cpp-linux-2026-11-13", engineArtifactId: "cuda13" }
+      - { engineReleaseId: "llama-cpp-2026-11-13", engineArtifactId: "metal" }
     artifactBindings:
       - bindingId: "primary"
         modelArtifactId: "primary"

catalog/models/muse-glimmer-30b.yaml

@@ -50,9 +50,9 @@ configurations:
   # targets and the Metal binary on macOS native targets; both releases share a parameter schema.
   - configurationId: "llama-cpp"
     engineArtifacts:
-      - engineReleaseId: "llama-cpp-linux-2026-10-19"
+      - engineReleaseId: "llama-cpp-linux-2026-11-13"
         engineArtifactId: "cuda13"
-      - engineReleaseId: "llama-cpp-2026-10-19"
+      - engineReleaseId: "llama-cpp-2026-11-13"
         engineArtifactId: "metal"
     artifactBindings:
       - bindingId: "q3-xl"

catalog/models/nanbeige-4-3-33b.yaml

@@ -57,7 +57,7 @@ artifacts:
 configurations:
   - configurationId: "llama-cpp"
     engineArtifacts:
-      - engineReleaseId: "llama-cpp-linux-2026-10-19"
+      - engineReleaseId: "llama-cpp-linux-2026-11-13"
         engineArtifactId: "cuda13"
     artifactBindings:
       - bindingId: "q3-iq3-s"

catalog/models/nemotron-3-5-lightning-30b-a3b.yaml

@@ -45,9 +45,9 @@ configurations:
   # targets and the Metal binary on macOS native targets; both releases share a parameter schema.
   - configurationId: "llama-cpp"
     engineArtifacts:
-      - engineReleaseId: "llama-cpp-linux-2026-10-19"
+      - engineReleaseId: "llama-cpp-linux-2026-11-13"
         engineArtifactId: "cuda13"
-      - engineReleaseId: "llama-cpp-2026-10-19"
+      - engineReleaseId: "llama-cpp-2026-11-13"
         engineArtifactId: "metal"
     artifactBindings:
       - bindingId: "q3-xl"

catalog/models/ornith-1-5-35b-a3b.yaml

@@ -421,9 +421,9 @@ configurations:
   # targets and the Metal binary on macOS native targets; both releases share a parameter schema.
   - configurationId: "llama-cpp"
     engineArtifacts:
-      - engineReleaseId: "llama-cpp-linux-2026-10-19"
+      - engineReleaseId: "llama-cpp-linux-2026-11-13"
         engineArtifactId: "cuda13"
-      - engineReleaseId: "llama-cpp-2026-10-19"
+      - engineReleaseId: "llama-cpp-2026-11-13"
         engineArtifactId: "metal"
     artifactBindings:
       - bindingId: "q4-k-m"
@@ -479,7 +479,9 @@ configurations:
         # The prompt cache shares the unified pool with the model.
         parameters:
           promptCacheOnAccelerator: 1
-          baseOffsetBytes: 0
+          # Measured on a GB10 with llama.cpp b12359 (2026-11-13): the derived base terms read up to
+          # 392 MiB low at the profile defaults (Q4_K_M; Q6_K 267 MiB).
+          baseOffsetBytes: 469762048
         bindingSizing:
           # Derived from qwen3-6 q4-xl (GB10): main-weight difference -1,158,046,240 B.
           - bindingId: "q4-k-m"
@@ -513,8 +515,9 @@ configurations:
         # The prompt cache stays in host memory, outside the GPU's own memory.
         parameters:
           promptCacheOnAccelerator: 0
-          # The qwen3-6 dedicated-memory offset (400 MiB, from its RTX 5090 smoke runs).
-          baseOffsetBytes: 419430400
+          # Measured on an RTX 5090 with llama.cpp b12359 (2026-11-13): the derived base terms read up
+          # to 766 MiB low at the profile defaults (Q4_K_M; Q5_K_M 704 MiB, Q6_K 640 MiB).
+          baseOffsetBytes: 889192448
         # The binding parameters are the nvidia-unified values verbatim, as in qwen3-6.
         bindingSizing:
           - bindingId: "q4-k-m"

catalog/models/qwen3-5-4b.yaml

@@ -49,9 +49,9 @@ configurations:
   # targets and the Metal binary on macOS native targets; both releases share a parameter schema.
   - configurationId: "llama-cpp"
     engineArtifacts:
-      - engineReleaseId: "llama-cpp-linux-2026-10-19"
+      - engineReleaseId: "llama-cpp-linux-2026-11-13"
         engineArtifactId: "cuda13"
-      - engineReleaseId: "llama-cpp-2026-10-19"
+      - engineReleaseId: "llama-cpp-2026-11-13"
         engineArtifactId: "metal"
     artifactBindings:
       - bindingId: "q3-xl"

catalog/models/qwen3-5-9b.yaml

@@ -47,9 +47,9 @@ configurations:
   # targets and the Metal binary on macOS native targets; both releases share a parameter schema.
   - configurationId: "llama-cpp"
     engineArtifacts:
-      - engineReleaseId: "llama-cpp-linux-2026-10-19"
+      - engineReleaseId: "llama-cpp-linux-2026-11-13"
         engineArtifactId: "cuda13"
-      - engineReleaseId: "llama-cpp-2026-10-19"
+      - engineReleaseId: "llama-cpp-2026-11-13"
         engineArtifactId: "metal"
     artifactBindings:
       - bindingId: "q3-xl"

catalog/models/qwen3-6-35b-a3b.yaml

@@ -113,9 +113,9 @@ configurations:
   # One configuration for every platform: CUDA on Linux NVIDIA, Metal on macOS.
   - configurationId: "llama-cpp"
     engineArtifacts:
-      - engineReleaseId: "llama-cpp-linux-2026-10-19"
+      - engineReleaseId: "llama-cpp-linux-2026-11-13"
         engineArtifactId: "cuda13"
-      - engineReleaseId: "llama-cpp-2026-10-19"
+      - engineReleaseId: "llama-cpp-2026-11-13"
         engineArtifactId: "metal"
     artifactBindings:
       - bindingId: "q3-xl"

catalog/models/qwen3-8-27b.yaml

@@ -89,9 +89,9 @@ configurations:
   # targets and the Metal binary on macOS native targets; both releases share a parameter schema.
   - configurationId: "llama-cpp"
     engineArtifacts:
-      - engineReleaseId: "llama-cpp-linux-2026-10-19"
+      - engineReleaseId: "llama-cpp-linux-2026-11-13"
         engineArtifactId: "cuda13"
-      - engineReleaseId: "llama-cpp-2026-10-19"
+      - engineReleaseId: "llama-cpp-2026-11-13"
         engineArtifactId: "metal"
     artifactBindings:
       - bindingId: "q3-xl"

engines/llama-cpp/AGENTS.md

@@ -2,38 +2,28 @@
 
 ## How it is built
 
-* Linux: [`docker-image.lock.json`](docker-image.lock.json) pins the upstream image by digest; the
-  node pulls it on first use or loads its stored image. The release declares the driver floor.
-* macOS: `engines.llama-cpp.native` ([`package.nix`](package.nix)), `aarch64-darwin` only. The
-  release tarball is installed unmodified (rewriting the Mach-O files would break their
-  signatures): tools and dylibs under `libexec/llama-cpp/`, linked into `bin/`; the main program is
-  `llama-server`.
+* Linux: [`docker-image.lock.json`](docker-image.lock.json) pins upstream's CUDA image by digest.
+  The release declares the driver floor.
+* macOS: [`package.nix`](package.nix) installs upstream's macOS arm64 release tarball unmodified;
+  rewriting its Mach-O files would break their signatures.
 * The package builds with the repository flake's nixpkgs, so a root nixpkgs bump can change its
   store path; existing deployments keep the store path they pinned.
 * Nothing here is part of root `just build`, `test` or `check`.
-* Shared parameters: [`engines/AGENTS.md`](../AGENTS.md) says where their flags and rules live.
 
 ## Update
 
-1. Pick the newest tag that has both the Linux image and the macOS asset (the two usually move
-   together). Linux image: `just image-lock ghcr.io/ggml-org/llama.cpp:server-cuda13-<tag>`. Images
-   appear after the release tarballs, so the newest tag can lack one; the recipe then fails.
-2. macOS package: check the release has a `llama-<tag>-bin-macos-arm64.tar.gz` asset (some, such as
-   `b10156`, have none). Set `version` in [`package.nix`](package.nix) to the tag, and `src.hash`
-   to the asset's SHA-256 in SRI form (`sha256-...`): the release asset's `digest`, or
-   `nix store prefetch-file` on its URL; convert hex with `nix hash convert --hash-algo sha256 <hex>`.
-3. Declare new releases with the tag as `version`: `llama-cpp-linux-<date>` in `llama-cpp.yaml` and
-   `llama-cpp-<date>` in `llama-cpp-macos.yaml`, after the build's publish date, and move every
-   model approving the old releases to them. Releases are immutable, and the macOS package `version`
-   must equal its release's `version` (`checks.dotcore-catalog` fails otherwise); keep an older
-   release by pinning it to the `github:dotcoreai/dotcore/<rev>` revision that built it.
+* Target upstream's newest stable release (`v<semver>`), not a nightly. It shares its commit with a
+  nightly build tag (`b<build>`): both artifacts are published under that build, and it is the
+  release `version` (note the semver in a comment).
+* Lock the release's CUDA image by its index digest, never by tag. Images can lag the tarballs, and
+  not every nightly gets one.
+* The macOS package `version` must equal its release's `version`; the catalog check fails
+  otherwise.
+* Releases are immutable: declare new ones with the date convention of `llama-cpp.yaml` and
+  `llama-cpp-macos.yaml`, and move every model approving the old ones.
 
 ## Validate
 
-1. `just smoke` (on a Mac): prints the package store path and `llama-server --version`; the build
-   number must match the tag. No model or GPU needed.
-2. Linux image, on an NVIDIA host: `just image-load`, then from the root
-   `just dev::workload-launch <model> <profile> --device nvidia.com/gpu=0` with a small model
-   approving the new release; the workload answers a request.
-3. `just backend::test-nvidia` on an NVIDIA host (from the root) deploys the engine from the
-   workspace's development source through its loader, as a node does.
+* `just smoke` on a Mac checks the package and the build number it reports.
+* The backend's hardware test recipes on each platform in the root "Platform scope":
+  `test-nvidia` (NVIDIA hosts), `test-gb10` (GB10) and `test-apple` (Mac).