HaloLoom v0.1.1 — source-current Strix Halo workbench

HaloLoom v0.1.1

A container-based toolkit for running and improving LLM inference on AMD Strix Halo.

HaloLoom brings Hyperloom, vLLM, SGLang, AMD Quark, and GPU profiling tools together in a pinned ROCm 10 environment for the Radeon 8060S (gfx1151). It is designed for users who want to serve models, quantize them, or explore agent-assisted optimization without assembling the underlying Python and GPU toolchains on the host.

Quickstart · User guide · Build guide · Test results

What's included

• vLLM and SGLang images for OpenAI-compatible model serving and Hyperloom optimization.
• AMD Quark for agent-assisted model quantization.
• KernelForge and GEAK for generating, compiling, and testing kernel candidates.
• Magpie, InferenceX, TraceLens, and IntelliKit Metrix for benchmarks, accuracy checks, and GPU profiling.
• Optional AITER tools for selected kernel-development tasks.
• Install, verification, serving, quantization, source-fetch, and rebuild helpers.
• Integration with an existing Claude Code, Codex, or Hermes CLI. Agent executables are not bundled in the images.

What's new in v0.1.1

• Native Radeon 8060S detection and the corresponding vLLM/SGLang benchmark runners in the optimizer.
• Codex/ChatGPT subscription login support across the OpenAI-side optimizer roles, alongside the existing API-key and Claude subscription paths.
• Working vLLM native GPU profiling with the required library settings included in the image and Compose configuration.
• Updated SGLang 0.5.19 development sources and writable runtime/JIT cache settings for non-root operation.
• A public, checksum-verified rebuild workflow using pinned base images and a downloadable build-input archive.

Exact component commits and image references are listed in the release manifest.

Get started

The host needs Linux with a working AMD GPU driver, Docker with Compose v2, Git, Python 3, access to the video/render groups, and an installed supported agent CLI. Model capacity depends on available RAM and the requested context and concurrency.

git clone https://github.com/HawgAuto/HaloLoom.git
cd HaloLoom
unset HSA_OVERRIDE_GFX_VERSION
./scripts/install.sh
./scripts/haloloom vllm verify

The installer pulls the vLLM, SGLang, and Quark images. Add --include-aiter for the optional tools image or --agent codex to choose a CLI explicitly. It does not install Docker, the kernel driver, an agent CLI, or host Python packages.

Follow the quickstart to start a small model and send an API request. The full optimizer also needs supported orchestration credentials; a Hermes CLI alone does not provide them. See authentication.

Downloads and rebuilding

Most users should install the published container images rather than assemble the release files manually. The attached assets are:

• hyperloom_inference_optimizer-1.0.0-py3-none-any.whl — the pinned Hyperloom package used by the images; not a standalone installation of the full stack.
• haloloom-v0.1.1-build-inputs.tar.gz — the source and overlay inputs used by the public rebuild workflow.
• SHA256SUMS — checksums for the wheel and archive.

To rebuild, use ./scripts/build_images.sh as described in the build guide. Framework builds use the pinned public bases; AITER is reused at its recorded digest.

Tested behavior and current limits

Release checks covered an anonymous clone and image installation, 123 passing tests, a public rebuild, and a real API response from the released vLLM image. Native vLLM profiling and SGLang/Quark model requests were also exercised. Detailed configurations, results, and original evidence records are in Test results and limitations.

Before use:

• Keep native gfx1151 detection; do not set HSA_OVERRIDE_GFX_VERSION.
• Use a fixed vLLM KV-cache budget if shared-memory changes trip automatic sizing. The published small-model API check was text-only, not broad multimodal coverage.
• Keep SGLang's supplied Triton attention configuration with CUDA graphs disabled. Graph discovery, capture/replay, and EAGLE operation are outside current test coverage.
• AITER attention and sampling are unavailable. Optional low-bit kernels have separate, model/framework-specific coverage.
• The default server port is published on host interfaces without API authentication. Restrict access, trust your model sources, and read the security policy.

Quantization and optimization results must be evaluated for your workload; no general quality or speed improvement is promised. HaloLoom does not automatically change existing inference services or apply optimizer results.

HaloLoom v0.1.0

HaloLoom v0.1.0

Turnkey ROCm 10 / AMD Radeon 8060S (gfx1151) distribution of the faithful Hyperloom port.

Included

• reviewed Hyperloom wheel hyperloom_inference_optimizer-1.0.0-py3-none-any.whl (sha256:9cc9884911db814c8d72d7e1219fe580f94f98849674f3b468f16caf3a8e09de)
• full vLLM and SGLang serving workbench images
• faithful Quark agent image
• optional AITER tools/JIT image
• installed pinned Magpie, IntelliKit Metrix, TraceLens, GEAK, InferenceX and built-in KernelForge
• Compose, preflight, source-sync, serving and quantization launchers
• automatic read-only plug-in for an existing host Claude Code, Codex or Hermes CLI; no agent CLI is bundled

Exact public image references and registry digests are in manifests/components.json.

Low-bit package

The six-route HIP module and framework adapters are independently owned and released at:

https://github.com/HawgAuto/Strix-Halo-Lowbit-Kernel-Pack/releases/tag/v0.1.0

All six routes pass direct physical qualification. W4A8 additionally passes final-image vLLM and SGLang request/dispatch qualification. No performance win is claimed.

Boundaries

• Native gfx1151 is required; do not set HSA_OVERRIDE_GFX_VERSION.
• SGLang uses Triton attention with CUDA graphs disabled.
• AITER core, gemm-common and RMSNorm canaries pass; AITER greedy sampling and AITER attention are unavailable.
• The prior Quark INT8 PTQ candidate was quality-rejected; no quantization quality win is claimed.
• Clean low-bit-free framework wheels are provenance-bound but are not v0.1 images because they have not passed separate physical serving.
• Nothing is promoted into an existing production service automatically.