[zaneni6]

Summary

Restructures CI around one workflow, Build and Test, that builds every FastFlowLM package and tests each one on the NPU runners as soon as its own build finishes.

Builds

Piece	What it does
.github/actions/win-build	Builds one Windows variant: XRT, HRX, or Source (engines built from source). Build only.
.github/actions/win-msi	Builds the MSI around an already-built flm.exe.
.github/actions/lin-build	Builds one Linux variant (XRT-deb, HRX-deb, XRT-portable, HRX-portable, Source-portable) in the build-environment container.
lin-build.yml / win-build.yml	Dispatch and workflow_call only: expand a variant (or all) into a matrix, package, upload. Artifact names unchanged.

• New HRX .deb for every distro: debian/rules reads FLM_USE_HRX / FLM_CMAKE_ARGS, Depends: libxrt-npu2 only for XRT, and src/CMakeLists.txt links non-portable HRX builds with -rpath-link to the system lib dir (the HRX SDK ships an unversioned libtinfo that broke the link).

Build and Test (build-and-test.yml)

Linux    XRT portable ─ test     HRX portable ─ test     Source portable ─ test
Windows  XRT zip ─ test          HRX zip ─ test          Source zip ─ test

• Each package is its own build call, so its tests start as soon as that build is done; XRT and HRX are tested side by side on any NPU runner of the platform.
• HRX skips api and serving, which test server logic the backend does not touch.
• A pull request runs only the suites its changes can affect (.github/script/ci/changes.py): an ASR-only change runs asr, a d3-only change decision, an engine change chat + prefix_cache; shared code runs everything. Docs-only changes build and test nothing.
• A new PR commit cancels the previous run; pushes to main, nightly and tags always finish.
• Nightly (full tier) tests every catalog model and every suite.
• .deb / MSI install tests are off unless the repository variable FLM_INSTALL_TESTS is true (the runners cannot install yet: no passwordless sudo on Linux, an FLM already installed on Windows).
• Stale flm processes left by earlier jobs are cleared before testing; a runner whose NPU is held by an unkillable one fails at once, named.
• New actions: detect-changes, install-package, remove-package, npu-test.

NPU test harness (.github/script/npu_test)

Stdlib-only Python, server mode, replaces sweep_*.py and asr_test.py:

Suite	Checks
chat	One model per engine family (19 on a PR instead of 49), plus every model of a family the change touches: known answer, streamed two turns, non-stream usage, vision colour check on generated images
serving	Queue full (503), concurrent requests routed to their own replies, client disconnect releasing the NPU, model switch, unknown model (/api/cancel skipped until fixed)
prefix_cache	Multi-slot KV cache: hit on return, LRU eviction with --prefix-slots 2, cached vs cold reply
api	src/test/openai_api_test + src/test/openai_usage (tools, logprobs, response_format, reasoning, sampling, continue_final_message, errors), never run in CI before
embedding	Shape, batching, similarity ordering, determinism
decision	d3 goldens, question types, errors, images, turbo/ultra prefix features (no longer reads the docs)
asr	Whisper WER/CER on FLEURS vs expected.json, formats, segments, language, prompt; chat transcription for gemma4-it e2b/e4b/12b and nemotron3.5-bolt

Release

release.yml runs Build and Test on the tag (no models-from-source) and publishes each backend only when all of its packages passed. HRX debs are now released with the HRX assets.

Removed

windows-build.yml, ubuntu-build.yml, debian-portable.yml, model-sweep.yml, asr-test.yml, resolve_build.sh, asr_gate.sh, sweep_*.py, asr_test.py, asr_summary.py. The CPU-only d3-contract / chat-continue tests move to cpu-tests.yml unchanged.

Known server bugs found by the new tests (not fixed here)

• /api/cancel registers itself under the request_id it names and cancels itself (src/server/server.cpp, process_task); serving/api-cancel is skipped until fixed.
• After a client disconnect the NPU lock was once never released and the process survived SIGKILL (strix runner); intermittent.
• An unknown model answers HTTP 200 with an error body instead of a 4xx (warns).
• A client disconnect leaks a connection slot (HttpSession::send_chunk_data skips active_connections_.fetch_sub).
• min_p is read as an integer (rest_handler.cpp:815).

Notes for reviewers

• Check names change (e.g. Test Linux XRT portable); update branch protection if any are required.
• ASR on PRs compares against expected.json instead of re-running main's build as a baseline.
• Runners: xsjnuckrck01 drops every job ("lost communication"); xconucgpt1-21 sets FLM_SERVE_PORT=52625 and something there holds that port (the harness now always picks a free port).
• d3:flash has no goldens file yet, so its goldens case warns.

Test plan

• actionlint clean; changes.py unit tests (17)
• Harness end to end against a mock flm
• Build and Test green on real NPU runners (run 40966757444, 41 min): Linux/Windows XRT and HRX portable, models-from-source, all builds incl. HRX .deb on 4 distros
• First nightly (full tier) after merge
• First tag through release.yml

🤖 Generated with Claude Code