Hyperloom v1.0.0b2

Hyperloom v1.0.0b2

Highlights

Official upstream vLLM ROCm image

Every vLLM image reference moves from rocm/hyperloom:vllm-v0.27.1-rocm7.2.3 to the official upstream vllm/vllm-openai-rocm:v0.27.1, because AMD deprecated rocm/vllm and rocm/vllm-dev. The tag is a 1:1 replacement, but its entrypoint is vllm serve, so a long-running Hyperloom container has to override it (for example --entrypoint tail). SGLang images are unchanged.

Magpie benchmark upgraded to v0.2.0

The default benchmark dependency moves from v0.1.0 to v0.2.0. Both the installer and the runtime preflight stay pinned to the immutable v0.2.0 release commit, so installs remain reproducible.

One current remote Recipe KB Store contract

Remote mode reads a single identity-addressed inference Recipe carrying replay config, the ordered patch timeline, and nested kernel columns, then publishes one final CLOSE session with verified artifacts under the same throughput champion. Degraded configuration donors now require exact precision, and a permanently missing owner patch is dead-lettered instead of blocking publication of the remaining Recipe sections. Local Recipe storage and non-Recipe GBrain integrations are unchanged.

KERNEL_OPT_BACKEND_ORDER is the single kernel-backend switch

The GEAK gate (_geak_enabled) no longer falls back to the persisted shared_state.kernel_optimizer field, so the backend choice is identical on a resume. Only an exact forge opts out of the default GEAK phase. The field itself is unchanged and still feeds the session breakdown. The KERNEL_OPT_BACKENDS environment variable is removed; no production code read it. (Breaking)

Action metadata consolidated into code

actions/_meta/*.yaml and orchestrator/actions/registry.py are replaced by ACTION_CATALOGUE in inference_optimizer/protocol/action_surfaces.py. Editing a yaml no longer changes anything because there is no yaml. The preferred_backend, preferred_model, max_turns, and params_schema fields are dropped because no runtime code ever read them; the operational verdict_class is kept. (Breaking)

Three unexecutable kernel actions removed

vendor_kernel_config, operator_tuning, and deep_kernel_analysis never had an executor or a KERNEL_REQUEST_HANDLERS kind, so every request for them was answered with unknown_kernel_kind; they were authored for the kernel_agent LLM role that #1095 retired. Sessions recorded under the old build that carry these names in state.json / coordinator.db are no longer resumable, and no migration is provided. (Breaking)

Write-only artifacts no longer produced

agent_transcript.jsonl, orchestration_turns.jsonl, mn_input_params_*.json, and the work_dir copy of semantic_audit.json had no reader, and the first three persisted secrets or raw LLM transcripts past a redactor that inspected values but not keys. (Breaking)

Magpie leak salvage is now opt-in

Salvage no longer defaults to /workspace/ and runs only when $INFERENCE_OPTIMIZER_RESCUE_PATHS is set. The generic {framework}_{gpu_type}.sh scripts respect $RESULT_DIR and never needed salvage, but a script pinned through params.benchmark_script that hardcodes /workspace/ now fails the task with no_report unless the environment variable is set explicitly. (Breaking)

kernel_optimization.py drops --test-command and --test-harness-path

The unittest-harness contract they fed had no reachable caller. An external invoker still passing either flag now fails in argparse rather than being silently ignored. (Breaking)

What's included

Images and benchmarks

• Bump the recommended vLLM stack to v0.27.1 (ROCm 7.2.3) - #1195
• Use the official vllm/vllm-openai-rocm image for vLLM - #1220
• Upgrade Magpie to v0.2.0 - #1218

Knowledge base / Recipe

• Harden Recipe finalization and warm replay - #1199
• Make the launch shape and reference recipe outlive the process that parsed them - #1197

Consolidation and removals

• Retire the unreachable kernel-harness stack; fix benchmark workspace selection - #1163
• Finish the workspace fix, close the catalogue drift, stop writing dead guidance - #1183

Release

• Cut 1.0.0b2 - #1219

Requirements

(unchanged from v1.0.0b1) ROCm platform with a supported AMD Instinct GPU (MI355X / MI325X / MI300X; MI308X detection); Docker with AMD ROCm support, Kubernetes with AMD GPU nodes, or the supported bare-metal setup; Cursor IDE with Remote SSH; an Anthropic API key or compatible Anthropic gateway, or an OpenAI-compatible gateway; tmux, jq, Node.js >= 18, and Claude CLI.

The beta wheels are published on PyPI, so the documented install no longer names a GitHub release asset URL. The version pin is required: a pre-release is skipped by the resolver unless it is asked for by name.

python3 -m pip install "hyperloom-inference_optimizer==1.0.0b2" --target .

Beta notice

v1.0.0b2 is a beta release. Beta signals that the surface is frozen: configuration surfaces, orchestration policies, and installation workflows are not expected to change further before v1.0.0.