Merge pull request #879 from docker/use-llamacpp-images

Use official llama.cpp images for Linux

Merge pull request #880 from docker/fix/vllm-metal-chat-template

fix: pass chat template to vllm-metal backend

refactor: normalize ModelPack config to Docker format in API responses (

#875)

feat: add CNCF ModelPack type definitions (#859)

* feat: add CNCF ModelPack type definitions, conversion utilities, and manifest ArtifactType support

- Add CNCF ModelPack media type constants and Model type (modelpack/types.go)
- Add DockerConfigToModelPack conversion utility (modelpack/convert.go)
- Add CNCFModel partial type for CNCF format support (partial/cncf_model.go)
- Add ArtifactType field to OCI Manifest struct (oci/manifest.go)
- Update partial/mutate layers for format-aware manifest options
- Replace deprecated WithConfigMediaType with WithManifestOptions
- Add model-spec v0.0.7 dependency

* feat: add --format flag to package command and CNCF build support in builder

- Add BuildFormat type and WithFormat option to builder for Docker/CNCF selection
- Implement CNCF artifact creation with media type remapping and config conversion
- Add --format CLI flag to 'docker model package' command with validation
- Improve context size error handling (return error for invalid values)
- Add comprehensive builder tests for CNCF format (artifact validation, media types)
- Add integration tests for CNCF format packaging
- Update CLI reference documentation

* feat: add CNCF model unpacking, detection, and mmproj file handling

- Add DiffID extraction support for CNCF remote layers (oci/remote)
- Add CNCF layer handling in bundle unpack with media type classification
- Detect mmproj files by filepath annotation in CNCF generic weight layers
- Add findModelFileExcluding helper for mmproj-aware model file discovery
- Add comprehensive tests for DiffID extraction, CNCF layer unpacking,
  mmproj detection, and parse with CNCF mmproj GGUF files

* fix: harden remappedLayer descriptor and add format auto-detection

Address review feedback from PR #860:

1. remappedLayer.GetDescriptor: pre-compute digest and size at construction
   time via newRemappedLayer(), surfacing errors eagerly instead of silently
   producing invalid OCI descriptors.

2. FromModel format inheritance: when --format is not explicitly set, auto-
   detect the source model's format (Docker vs CNCF) from its config. This
   prevents producing inconsistent artifacts when repackaging a CNCF model
   without an explicit --format flag.

3. CLI conditional --format: only pass WithFormat to the builder when the
   user explicitly sets --format, allowing the builder's auto-detection to
   work for the repackaging case.

* fix: preserve determinism in DockerConfigToModelPack and fix normalizeParamSize comment

- Remove time.Now() fallback in DockerConfigToModelPack: propagate
  desc.Created directly (nil when unset) so callers control timestamps
  and converted configs remain deterministic across runs.
- Fix normalizeParamSize comment to accurately describe its behavior
  (lowercases, does not truncate decimal precision).

* refactor: remove normalizeParamSize, pass Parameters as-is to CNCF config

The function only lowercased the string, which is unnecessary — the CNCF
model-spec does not mandate lowercase paramSize values. Pass cfg.Parameters
directly to keep the same value as the Docker format.

* fix: use newRemappedLayer in FromDirectory for consistent digest/size caching

The CNCF path in FromDirectory was constructing remappedLayer directly,
bypassing the digest/size pre-computation added in newRemappedLayer.
This left cachedDigest/cachedSize zero-valued for layers that aren't
descriptor providers, producing invalid OCI descriptors.

---------

Co-authored-by: Eric Curtin <eric.curtin@docker.com>

chore(deps): bump go.opentelemetry.io/otel/sdk from 1.41.0 to 1.43.0 (#…

…851)

Bumps [go.opentelemetry.io/otel/sdk](https://github.com/open-telemetry/opentelemetry-go) from 1.41.0 to 1.43.0.
- [Release notes](https://github.com/open-telemetry/opentelemetry-go/releases)
- [Changelog](https://github.com/open-telemetry/opentelemetry-go/blob/main/CHANGELOG.md)
- [Commits](open-telemetry/opentelemetry-go@v1.41.0...v1.43.0)

---
updated-dependencies:
- dependency-name: go.opentelemetry.io/otel/sdk
  dependency-version: 1.43.0
  dependency-type: indirect
...

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

fix: update context size precedence to prioritize backend configurati…

…on over model configuration (#847)

* fix: update context size precedence to prioritize backend configuration over model configuration

* fix: ensure context size from model configuration is positive before using it

* fix: validate context size range and update backend configuration handling

* fix: implement GetMaxTokens function to prioritize backend context size over model configuration in mlx also added goleak detector in mlx package

feat: enhance progress bar status handling and alignment (#845)

Merge pull request #813 from docker/unpack-cncf

Add CNCF ModelPack support and improve model format detection

chore: bump docker-model and llamacpp versions in build.json (#843)

* chore: bump docker-model and llamacpp versions in build.json

* chore: update llamacpp version fetching logic in release workflow

* chore: streamline release workflow by adding model-runner checkout and loading GO version conditionally

chore: update llamacpp version fetching logic in release workflow