PR-N4: remove SDK conftest stub + finalize no-doubles cleanup#56
Merged
FluffyAIcode merged 2 commits intoJun 2, 2026
Merged
Conversation
Final installment of the no-test-doubles cleanup. Closes the
sequence PR-N1 \u2192 N2 \u2192 N3 \u2192 N4. After PR-N4 lands, NO
test doubles implementing the verifier / engine / tokenizer
protocols remain in the Linux test tree.
What was deleted
----------------
tests/sdk/python/conftest.py
-203 lines. Contained _start_runtime / _stop_runtime helpers
that spun up an in-process gRPC server with a FakeVerifier
(later replaced by _MinimalVerifierStub in PR-N1's preview
cleanup) on a background thread. The runtime_address +
runtime_address_no_inspector fixtures are gone with it.
tests/sdk/python/test_client.py
-157 lines, 13 tests. Exercised Client + Session lifecycle
against the FakeVerifier-backed runtime.
tests/sdk/python/test_session.py
-502 lines, 33 tests. Exercised Session.append + .generate +
.info + .close end-to-end against the FakeVerifier-backed
runtime.
What was added
--------------
tests/integration/test_sdk_real.py +137 lines, 11 tests
SDK Client + Session integration tests against a real
Qwen3-0.6B-backed gRPC runtime:
- Client: create_session round-trip, eos_token_ids round-
trip, idempotent close, address property
- Session: append + generate yield tokens + metadata,
info reflects history, close returns final length,
close is locally idempotent
- End-to-end error mapping: SessionNotFoundError on
unknown id, InvalidArgumentError on max_tokens=0,
SessionClosedError on append-after-close
tests/integration/conftest.py +180 lines
- pytest_collection_modifyitems hook (auto-marks everything
under tests/integration/ with @pytest.mark.integration)
- real_speculative_engine fixture (session-scoped, Qwen3-0.6B)
- real_grpc_runtime_address fixture (session-scoped,
in-process gRPC server backed by real Qwen3-0.6B verifier
on a background thread; yields the host:port the SDK can
connect to)
tests/integration/__init__.py +0 lines (placeholder)
scripts/review_pr_n4_on_mac.sh +93 lines
Mac M4 reviewer aid running the full accumulated integration
suite (PR-E1 INV-3 + PR-N1 coordinator/generator + PR-N2
scheduler + PR-N3 http_shim/engine/tokenizer/streaming +
PR-N4 SDK).
What stays on Linux
-------------------
tests/sdk/python/test_errors.py (unchanged, 9 tests)
Pure _wrap_grpc_error mapping with synthesized
grpc.RpcError objects. Verifier-independent;
transport-only error-class translation. Stays on Linux.
CI workflow change
------------------
.github/workflows/ci.yaml: dropped kakeya.client and kakeya.session
from the --include= filter. Linux gate now covers ONLY:
inference_engine/server/{auth, config, errors, grpc_app,
metrics, schemas, proto_gen}
inference_engine/memory/*
inference_engine/scheduler/{config, session, pooled_verifier}
inference_engine/pipeline/*
inference_engine/session/store
sdks/python/kakeya/{__init__, errors}
training/repr_align/*
That's the verifier-independent boundary, frozen post PR-N4.
Final state of the no-doubles cleanup
-------------------------------------
PR-N1 (#53): retired FakeVerifier hierarchy
(tests/inference_engine/session/test_coordinator.py,
test_generator.py, test_grpc_app.py FakeVerifier-using
sections).
PR-N2 (#54): retired DeterministicEngine + DeterministicTokenizer
(tests/inference_engine/scheduler/conftest.py +
test_scheduler.py).
PR-N3 (#55): retired the HTTP shim cluster (server/conftest.py
+ 6 test files + their subtypes).
PR-N4 (this): retired the SDK conftest stub.
The integration suite at tests/integration/ now contains:
test_inv3_session_determinism_gate.py (PR-E1)
test_coordinator_real.py (PR-N1)
test_generator_real.py (PR-N1)
test_scheduler_real.py (PR-N2)
test_http_shim_real.py (PR-N3)
test_engine_real.py (PR-N3)
test_tokenizer_real.py (PR-N3)
test_streaming_real.py (PR-N3)
test_sdk_real.py (PR-N4)
Linux verification
------------------
PYTHONPATH=.:sdks/python coverage run -m pytest <Linux gate paths>:
649 passed (was 695 on main; -46 net = removed 46 SDK runtime
tests, kept 9 SDK error-mapping tests).
100% coverage on 999 stmts (was 1660 on main; -661 net stmts is
all verifier-dependent modules now integration-only).
Mac M4 evidence (REQUIRED for merge)
------------------------------------
Per ADR 0008 \u00a79: this PR's runtime correctness lives in the
integration suite. Reviewer runs:
bash scripts/review_pr_n4_on_mac.sh
git add results/platform-tests/pr-n4-mac-*
git commit -m 'Mac M4 review evidence for PR-N4'
git push
Stack
-----
PR-N4 is branched off main, independent of PR-N1 (#53) /
PR-N2 (#54) / PR-N3 (#55) at the file level. Conftests in
tests/integration/ added by N1/N2/N3/N4 are file-disjoint from
each other (each adds one fixture) but the file IS shared, so
post-merge the four contributors' fixture defs need to be
reconciled. The recommended merge order:
1. PR-N1 (verifier doubles) — adds conftest with marker hook
2. PR-N2 (engine/tokenizer doubles) — adds real_speculative_engine
3. PR-N3 (HTTP shim doubles) — uses real_speculative_engine
4. PR-N4 (this, SDK doubles) — adds real_grpc_runtime_address
If a different order lands first, the integration conftest
needs a small merge to combine fixtures.
Co-authored-by: FluffyAIcode <FluffyAIcode@users.noreply.github.com>
…4-sdk-conftest-stub-cleanup-8e7f Co-authored-by: Cursor <cursoragent@cursor.com> # Conflicts: # .github/workflows/ci.yaml # tests/integration/conftest.py # tests/sdk/python/conftest.py
4 tasks
FluffyAIcode
added a commit
that referenced
this pull request
Jun 2, 2026
…tion workflow
Closes the loop on automated GA gating. After PR-N1..N4 retired all
verifier-protocol test doubles from the Linux gate, the integration
suite (tests/integration/) became the binding correctness gate for
runtime modules \u2014 inference_engine.session.coordinator,
inference_engine.session.generator,
inference_engine.scheduler.scheduler,
inference_engine.server.{app,engine,tokenizer,streaming}, and
kakeya.{client,session}. Until this PR, that suite ran manually
via scripts/review_pr_*_on_mac.sh; PR-E2 wires it into CI on every
PR labelled needs-mac-m4.
Three artifacts ship:
.github/workflows/integration.yaml +136 lines
Self-hosted runner workflow targeting [self-hosted, macOS,
ARM64, kakeya-mac-m4]. Triggers on PR events when the
needs-mac-m4 label is present, plus on workflow_dispatch
for manual re-runs. Steps:
1. Checkout (full history).
2. Verify host shape (chip, memory, python version).
3. Verify Qwen/Qwen3-0.6B is in HF cache (HF_HUB_OFFLINE=1
at test time \u2014 no downloads in CI; cache miss fails
fast with a clear pre-warm command).
4. pip install -e . + pytest dependencies (warm pip cache
keeps this <30 s).
5. pytest -m integration tests/integration/ \u2014 expected
runtime 60-120 s on M4 with warm cache. 90-min timeout
is a safety margin, not the operating point.
6. Upload JUnit XML artifact.
7. On failure, inline the test names + first-line error
messages into the Action log so triage doesn't require
downloading the artifact.
Concurrency: cancel-in-progress per PR, so a new push
supersedes the previous run.
.github/workflows/auto-label-mac.yaml +89 lines
pull_request_target workflow that auto-applies (or removes)
the needs-mac-m4 label based on which paths the PR touches.
Trigger paths:
inference_engine/ \u2014 runtime, scheduler, session, server
sdks/ \u2014 Python + TypeScript SDK
proto/ \u2014 wire contract
tests/integration/ \u2014 the integration suite itself
kv_cache_proposer/ \u2014 verifier + decoder
Doc-only or CI-only PRs are NOT labelled \u2014 they skip the
integration gate entirely, saving runner time. The label is
automatically dropped if a subsequent push removes all
verifier-dependent edits.
docs/ops/mac-m4-runner-setup.md +137 lines
Operator runbook for the self-hosted runner: hardware
requirements (24 GB minimum, ~50 GB free disk), runner
registration with the kakeya-mac-m4 label, HF cache
pre-warm command (Qwen3-0.6B), Python toolchain setup,
runtime expectations, cache hygiene cron, runner upgrade
procedure, and failure triage steps.
CI workflow split rationale
---------------------------
The pre-existing .github/workflows/ci.yaml stays as the Linux gate
(verifier-independent, runs on github-hosted ubuntu-latest, fires
on every PR). PR-E2 adds integration.yaml as a SEPARATE workflow
because:
1. Self-hosted runners are slow / few; doc-only PRs shouldn't
touch them.
2. The integration gate is intentionally OPT-IN by label; ci.yaml
is non-optional.
3. Failure semantics differ: Linux gate failure blocks merge
unconditionally; Mac M4 gate failure surfaces a structured
report but the merge decision is a human one until v0.3.0
final ships.
Together the two workflows form the post-cleanup gating model:
- Linux gate (ci.yaml):
verifier-independent code; 100% coverage; every PR.
- Mac M4 gate (integration.yaml):
verifier-dependent code; binding GA gate; PRs touching
runtime / SDK / proto / integration tests.
Stack
-----
PR-E2 is branched off main, independent of the cleanup PRs (#49,
#50, #51, #52, #53, #54, #55, #56). The workflow doesn't fail at
launch even before PR-E1 lands; it just won't find any tests
under tests/integration/ until that PR is merged. Recommended
merge order: cleanup PRs first (so the workflow has tests to
run), then PR-E2.
Per ADR 0008 \u00a79
----------------
PR-E2 ships ONLY workflow YAML + a runbook \u2014 no Python source
changes. No Mac M4 evidence required for this PR (the workflow
itself becomes the Mac M4 evidence machinery for ALL future PRs).
Co-authored-by: FluffyAIcode <FluffyAIcode@users.noreply.github.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Why
Final installment of the no-test-doubles cleanup. Closes the sequence PR-N1 → N2 → N3 → N4. After PR-N4 lands, NO test doubles implementing the verifier / engine / tokenizer protocols remain in the Linux test tree.
What was deleted
tests/sdk/python/conftest.py_start_runtime/_stop_runtimehelpers +runtime_addressfixture (FakeVerifier-backed)tests/sdk/python/test_client.pytests/sdk/python/test_session.pyTotal: ~860 lines, 46 tests deleted.
What was added
tests/integration/test_sdk_real.pytests/integration/conftest.pypytest_collection_modifyitemshook +real_speculative_engine(Qwen3-0.6B SpeculativeEngine, session-scoped) + NEWreal_grpc_runtime_address(in-process gRPC server backed by real verifier on background thread, yieldshost:port)tests/integration/__init__.pyscripts/review_pr_n4_on_mac.shWhat stays on Linux
tests/sdk/python/test_errors.py(9 tests) — pure_wrap_grpc_errormapping with synthesizedgrpc.RpcErrorobjects. Verifier-independent; transport-only error-class translation.CI workflow change
.github/workflows/ci.yaml— droppedkakeya.client,kakeya.sessionfrom--include=filter. Linux gate now covers ONLY:This is the verifier-independent boundary, frozen post PR-N4.
Final state of the no-doubles cleanup
test_coordinator_real.py,test_generator_real.pytest_scheduler_real.pytest_http_shim_real.py,test_engine_real.py,test_tokenizer_real.py,test_streaming_real.pytest_sdk_real.pyAfter all four merge,
tests/integration/contains:Linux verification
Mac M4 evidence (REQUIRED for merge)
bash scripts/review_pr_n4_on_mac.shruns the full accumulated integration suite (PR-E1 INV-3 + PR-N1 coordinator/generator + PR-N2 scheduler + PR-N3 http_shim/engine/tokenizer/streaming + PR-N4 SDK) against real Qwen3-0.6B and producespr-n4-mac-integration-tests-<unix>.jsonevidence.Stack & merge order note
PR-N4 is branched off
main, independent at the file level from PR-N1 (#53), PR-N2 (#54), PR-N3 (#55). However all four PRs add totests/integration/conftest.py(each contributes a fixture), so post-merge the conftest needs a small reconciliation. Recommended merge order:real_speculative_engine)real_speculative_engine)real_grpc_runtime_address)PR-N4's
tests/integration/conftest.pyis a superset of N1/N2's; if N4 lands first, the others need to skip their conftest creation and reuse the merged version.