Build and release pipelines

Checked configurations

Workflow	Work performed	Evidence boundary
`ci.yml`	Python policy/installer tests, installed C99/C++17 SDK on Ubuntu 22.04/24.04, Windows 2022 and macOS 14, Linux management/ABI tests, and ASan+UBSan / TSan `modular_abi_test` and `gkmodule_test` gates	Windows/macOS SDK checks do not validate a Windows/macOS inference engine; the Linux non-CUDA build cannot perform CPU-only inference
`backend-integration.yml`	Build/test exact CPU and runner commits against the current installed SDK, fresh-process signed registration, full pinned CUDA backend compilation and manifest/operation checks	Signing uses a temporary test identity; hosted CUDA compilation has no device execution
`dev-build.yml`	Linux CUDA 13 inference executable, portable CPU baseline, linked CUDA runtime/BLAS, licenses, manifest and checksums	Development artifact; no real-GPU or full-model claim
`release.yml`	Tag/revision validation, all CI checks, distribution build, dedicated-GPU shipped-binary golden test, reviewer-protected publication and OIDC provenance	Publication waits for real GPU evidence and configured release protection

The CUDA build container is pinned by its amd64 digest. It builds SM75, SM86, SM89 and SM120 images, with SM120 PTX. These are compilation targets, not a claim that every GPU architecture has passed inference. Pascal validation on the development machine uses CUDA 12/SM61 and remains distinct from this CUDA13 distribution. HIP, SYCL, Windows inference and CPU-only inference are not release artifacts of these workflows.

Development artifacts are available in each successful Actions run, identified by the full Core SHA. They do not overwrite a moving development release/tag.

Distribution layout

tools/package_linux_cuda.py --build build --output build-dist/guild-linux-x86_64-cuda13 stages:

guild: launcher that resolves bundled libraries relative to its own location;
guild.bin: the actual CUDA executable;
lib/: the exact linked libcudart, libcublas and libcublasLt SONAMEs;
licenses/: Guild/Strata, independent CPU and CUDA backends, GGML and NVIDIA CUDA license files;
guild-release.json: structured build identity, dependency hashes and Linux system-library prerequisites.

The adjacent .tar.gz and .sha256 preserve executable permissions. Packaging checks the launcher, unchanged executable provenance and actual bundled-library resolution with a fresh empty package store. It requires GUILD_PORTABLE=ON for public artifacts. --allow-native-cpu is limited to explicitly host-specific local development artifacts. The NVIDIA driver and listed Linux system libraries are prerequisites. Model GGUFs, prepared packs and tokenizers are not included. Install/import a matching model manifest with those paths before use.

The CUDA baseline must include all three extracted groups (dense, attention and MoE), and its actual source revision must match the embedded lock. If an existing CMake cache retains a previous CUDA pin, reconfigure with -UGUILD_CUDA_BACKEND_REVISION before building and packaging.

Release setup and real-GPU gate

Before a release, configure the GitHub environment release with required reviewers. The validation job checks this configuration; it does not create an unprotected environment as a substitute. The release tag must identify a clean commit reachable from main, match VERSION.json, satisfy maturity/revision gates, and have no existing GitHub release.

Provide a dedicated runner with labels self-hosted, linux, x64, guild-cuda-validation, a compatible NVIDIA driver/GPU, sufficient RAM and prepared model fixtures. This runner is used only by the validated tag workflow, not by pull-request jobs. Set the release environment variable GUILD_RELEASE_GPU_FIXTURE to a local golden-fixture JSON path:

{
  "model": "ornith:35b",
  "model_manifest": "/srv/guild-fixtures/ornith-manifest.json",
  "prefill_batch": 256,
  "request": {
    "model": "ornith:35b",
    "messages": [{"role": "user", "content": "Count from one to ten."}],
    "max_tokens": 16
  },
  "expected_text": "EXACT COMPLETION FROM THE VALIDATED REFERENCE"
}

The example is a schema illustration, not an actual golden. Maintain the real fixture from a separately validated reference with exact GGUF/pack/tokenizer versions. The manifest is copied into an isolated store and supplies absolute GGUF, prepared-pack and tokenizer paths. The test starts the downloaded shipped executable, bound to 127.0.0.1 with an isolated package/model store, forces greedy non-streamed generation, requires eight completed decode steps, and compares exact output. It uploads response, server log, GPU identity and binary/fixture hashes. This smoke gate supplements the numerical model tests; it does not replace full logit/state/KV parity or testing on every supported GPU.

The publication job waits for this evidence and environment review. It verifies archive hashes, creates GitHub OIDC build-provenance attestations, and publishes new assets. Verify downloads using their checksums and gh attestation verify ARCHIVE --repo guild-inference/Guild. This authenticates Core release provenance; production Ed25519 signing of external .gkernel packages is a separate, unfinished release pipeline. No production package private key is in these workflows.