[lishuoshuo-amd]

• Description: what and why

Cuts release 1.1.2 (54 PRs merged since v1.1.1).

Version bump: every site that tracks the packaged version moves together:

Site	Change
pyproject.toml [project].version	1.1.1 → 1.1.2
docs/conf.py version_number	kept in sync with pyproject
README.md version badge	1.1.1 → 1.1.2
docs/compatibility.rst matrix row	Hyperloom column only
docs/install/install.md, examples/README.md, hyperloom-setup/SKILL.md	the three pip install hyperloom-inference-optimizer== pins
install_baremetal.sh HYPERLOOM_WHEEL_TAG	v1.1.1 → v1.1.2

Release notes: CHANGELOG.md was retired in #1589, so docs/release-notes.md is the only document to cut. ## Unreleased becomes ## Hyperloom 1.1.2 release, and an empty ## Unreleased stays for the next cycle. The new section has:

• an introduction and a Before upgrading from 1.1.1 guide. It uses separate tables for the removed command line, which a strict parser refuses, and the removed environment variables, which are silently ignored. It also covers installation prerequisites and report-field changes.
• 1.1.2 highlights: ROCm 10, ENABLEMENT as the sixth phase, retired supervision, ATOM end to end, the KernelForge assembly backend and salvage, and SBD V6 grading axis.
• Other changes: the inherited Unreleased entries are kept verbatim, except for the edits listed below. New entries are added for user-facing PRs that merged without one: Bring up vLLM on ROCm 10 and fix scheduler stalls found in the 6-leg run #1605, fix(agents): route every agent CLI choice through the shared rule, on official key names only #1541, fix(kb): address one kernel's records by one address, however its source and its environment spell it #1529, [FEAT] dissolve Explore Phase, unify the attempt ledger, judge dryness per lever #1587, Let the rewrite pipeline's caller own its task shape, its files and its stages #1596, Remove Hyperloom's MI355X achievable roofline fallback #1602, fix(grid): keep a measurement that served its whole protocol #1584, fix(profile): follow InferenceX's move of benchmark_serving.py under infx/ #1594, fix(kernel-patch): stop SGLang editable reinstall for in-tree JIT sources #1567, fix(agentx): let an agentic recipe satisfy the server-phase pid contract #1570, fix(kernel): preserve operator GEAK checkouts #1591. feat(forge): optimize compiler-emitted AMDGPU assembly #1426, KernelForge: add fuzzy KB warm-start fallback across framework versions and GPUs #1512, Feat/rocm10 e2e #1549, Added support for ATOM detailed profiling #1509, fix(bench): stop baseline benchmarking from a private InferenceX copy #1456, Removing patching for SGLang versions #1486, fix(sbd): give every timeline failure one shape and one place to fail #1578, fix(slurm): do not carry the host's GPU mask into the container #1582 and feat(sbd): complete the timeline events and pin their wire #1585 are covered in the highlights or the upgrade guide instead.

Edits to inherited entries:

• The "robustness monitor called a session over while it was still running" entry is dropped, and the --resume-from entry loses its robustness_monitor.sh sentence. refactor: retire robustness supervision and simplify benchmark watchdogs #1547 retires that monitor in this same release.
• The fix(slurm): reclaim GPUs from orphaned containers and run docker with --init #1572 orphan-container entry gains the fix(slurm): stop the GPU reclaim from killing a live neighbour job #1581 follow-up (literal /proc match, HL_ORPHAN_MIN_AGE_S).

ROCm 10 / image switch coverage: #1549, #1593 and #1605 left sites that still described the pre-ROCm 10 stack:

Site	Change
assets/slurm/models.tsv	the two vLLM rows move from vllm/vllm-openai-rocm:v0.29.0 to the rocm/vllm:rocm10.0.0_..._vllm_0.27.0 image the docker route uses
docs/compatibility.rst minimum requirements	ROCm 7.2.x → 7.2.x, 10.0
docs/compatibility.rst, docs/install/install.md	TheRock's pip-packaged ROCm is no longer called an untested layout; it is what the rocm10 images are built from (install_baremetal.sh ROCM_SDK_INDEX_URL)
docs/install/install.md bare-metal requirements	ROCm 10.0 listed as validated; the Python 3.12 requirement is scoped to the ROCm 7.2.x vLLM wheel (the source build takes >= 3.10, < 3.15)
hyperloom-setup/SKILL.md	the vllm (isolated) option covers both routes, and the glibc floor note is scoped to the wheel route; install_vllm_framework returns into install_vllm_from_source before assert_vllm_glibc_compatible
docs/reference/environment-variables.md, docs/install/slurm.md	new ROCPC_VENV (#1549) and HL_ORPHAN_MIN_AGE_S (#1581) rows
docs/compatibility.rst entrypoint note	no longer says the vLLM image's entrypoint is vllm serve; the --entrypoint tail override is kept for the real reason

The image configs, read from the Docker Hub registry:

Image	Entrypoint	Cmd
vllm/vllm-openai-rocm:v0.29.0 (old)	["vllm", "serve"]	none
rocm/vllm:rocm10.0.0_..._vllm_0.27.0	none	["/bin/bash"]
lmsysorg/sglang-rocm:v0.5.20-rocm10-mi30x-20260920	none	["/bin/bash"]
rocm/atom-dev:v0.1.7-rc0	none	["/bin/bash"]

The models.tsv swap also repairs the sbatch docker branch for the vLLM rows. It runs docker run ... "${IMAGE}" bash "$INCONTAINER" without --entrypoint, which the old image turned into vllm serve bash .... The pyxis branch never used the image entrypoint. .github/scripts/forge-ci-e2e-dispatch.sh keeps its own CI image default.

Compatibility surfaces in this release

Surface	Detail
Removed options	--conc-sweep-timeout-sec (#1547), --recipe-kb-strict-fingerprint (#1542). Both exit unrecognized arguments now.
Removed console script / packaged file	robustness-agent, tools/robustness_monitor.sh.example (#1547)
Removed env vars	HYPERLOOM_SUPERVISOR, HYPERLOOM_SUPERVISOR_ENFORCE, HYPERLOOM_SUPERVISOR_TICK_STALL_SEC, ROBUSTNESS_AGENT_ROOT, ROBUSTNESS_LLM_RCA_DISABLED (#1547); INFERENCE_OPTIMIZER_DISABLE_LOCAL_INFERENCEX, INFERENCE_OPTIMIZER_LOCAL_INFERENCEX_ROOT (#1456); MAGPIE_PATCH_STRICT (#1592)
Credential rename	LLM_GATEWAY_KEY as the only OpenAI-side credential → OPENAI_API_KEY (#1541)
Added CLI	recover-session --confirm-stopped TASK_ID --confirmation-reason TEXT (#1547), forge-loop --kernel-backend assembly (#1426), --applyback/--no-applyback (#1596)
Report fields	SBD V6 baseline_throughput/optimized_throughput → baseline_value/optimized_value, failure rows phase → stage; kernel_optimization_summary.json schema_version 2; sweep_failed stop reason
Unchanged	session record: LATEST_STATE_SCHEMA_VERSION is 6 in both v1.1.1 and this tree

--max-minutes-enablement-pct is also removed, but it was added and removed within this release, so it is not in the upgrade guide.

CI flake fix: test_optimize_loop_walkthrough.py::test_both_arms_dry_walks_the_rest_of_the_chain failed on py3.10 shard 1 ('KERNEL_AGENT' == 'CLOSE', teardown logged _on_phase_entered hook was cancelled). Since #1605 the KERNEL entry hook runs out of band and holds the phase until it returns. In this test the hook started geak_runner.py as a real subprocess, and KERNEL only reached a terminal geak_result once that process failed. Whether 11 ticks reached CLOSE therefore depended on the CI host. The test now makes _kernel_agent_tool_path unresolvable, so GEAK takes the existing runner_not_found branch in-process, and it asserts that branch. Delaying the geak_runner.py spawn by 30s fails the pre-fix test with the CI assertion and leaves the fixed test passing in 6.6s.

• Linked issue(s): close/fix refs

None.

• Tests: added/updated? commands run?

No code change. Verified by running:

• pytest -n 4 src/kernelforge/tests/test_rename_completeness.py src/hyperloom/inference_optimizer/tests/test_baremetal_doc_version_consistency.py src/hyperloom/inference_optimizer/tests/test_scaffold_smoke.py src/hyperloom/inference_optimizer/tests/test_pre_release_gate_orchestration.py src/hyperloom/inference_optimizer/tests/test_preflight_serving_framework.py src/hyperloom/inference_optimizer/tests/test_multi_node_scripts.py: 198 passed.

• Every removed option, console script and environment variable named in the upgrade guide was checked present at v1.1.1 and absent from this tree.

• Size/complexity triggers crossed: split, or why the shape is right

Six commits. The first two are eight one-line version bumps plus docs/release-notes.md, the same shape as #1519. The ROCm 10 coverage commit (4 files), the env-var doc commit (2 files) and the entrypoint note (1 file) are docs and a sample registry only; the last commit is a one-test fix.

• If this simplifies or refactors: which mechanism was removed, which contract was preserved, and the tests that show it:

N/A.

• Observable effect: what an operator will now see, and what the old behaviour cost them (if none, why the change is unobservable):

Installs pin 1.1.2, and the bare-metal installer downloads its hotfix asset from the v1.1.2 release. The release notes carry the 1.1.1 → 1.1.2 upgrade guide.

• Breaking changes: yes/no (details if yes)

The PR itself is not breaking. The release it cuts carries the removals listed above, and each has a "what to do" row in the upgrade guide.

• PR addresses single concern: yes/no (details if no):

Yes, the 1.1.2 release cut.

• Root cause is upstream (Magpie/TraceLens/GEAK/IntelliKit/AgentKernelArena), ticket filed:

N/A.

Before merging

The v1.1.2 GitHub release must carry the rocm-profiler-hotfix-libs.tar.gz asset, because HYPERLOOM_WHEEL_TAG now points the bare-metal installer at that tag. Same requirement as v1.1.1.

Test plan

• Release, doc and setup-skill tests: 198 passed.
• No remaining vllm-openai-rocm reference in docs, examples, skills or packaged assets.
• Removed CLI, console-script and environment-variable surface checked against v1.1.1.
• LATEST_STATE_SCHEMA_VERSION unchanged (6).
• Pre-release E2E gate: the pyproject.toml version bump against base selects the FULL run.
• Docs build (Sphinx).