Files
gps-denied-onboard/_docs/03_implementation/batch_33_cycle1_report.md
T
Oleksandr Bezdieniezhnykh 0dfe7c5301 [AZ-321] C10 EngineCompiler: hardware-tied TRT compile + cache reuse
Land the C10 per-model engine compile + cache-reuse orchestrator.
`EngineCompiler.compile_engines_for_corpus(request)` walks the
corpus, computes the canonical engine filename via AZ-281
`EngineFilenameSchema.build`, and either reuses the cached binary
(cache hit, AZ-280 `Sha256Sidecar.verify` returns True) or delegates
to the AZ-297 `compile_engine` on the injected runtime (cache miss;
the runtime owns the write path). Returns one `EngineCompileResult`
per backbone carrying the canonical `EngineCacheEntry`, outcome
(BUILT / REUSED), and `compile_duration_s` (None on reuse).
Hardware-tied reuse (D-C10-6 / D-C10-7) falls out of the filename
schema — a host change rebuilds at the new path and leaves the old
files untouched (AC-4).

Design corrections vs. the task spec body:
- The spec proposed a c10-local `EngineCacheEntry` carrying outcome
  and duration; that name is already taken by the AZ-297 canonical
  DTO. The wrapper is renamed `EngineCompileResult`; the canonical
  shape wins.
- The spec called `InferenceRuntime.host_info()`, which is not in
  the AZ-297 Protocol. `HostCapabilities` is threaded through
  `EngineCompileRequest` instead so the composition root owns host
  probing and the compiler stays decoupled.
- The c10 layer cannot import `components.c7_inference` (arch rule
  `test_az270_compose_root.test_ac6`). `engine_compiler.py` defines
  `CompileEngineCallable` — a structural Protocol cut of
  `InferenceRuntime` exposing only `compile_engine` — and catches
  broad `Exception` (re-raising preserves the original type;
  `error_class` is recorded in the ERROR log payload).

Production
- engine_compiler.py: `CompileOutcome` enum, `BackboneSpec`,
  `EngineCompileRequest`, `EngineCompileResult`,
  `EngineCompileSummary` DTOs; `CompileEngineCallable` Protocol;
  `EngineCompiler` with the single public method.
- config.py: `BackboneConfig` + `C10ProvisioningConfig`
  (`workspace_mb` default 4 GiB to match C7 NFT-LIM-01); validate
  positive shape dims and duplicate model_name detection in
  `__post_init__`.
- runtime_root/c10_factory.py: `build_engine_compiler(config)` wires
  the existing `build_inference_runtime` factory through;
  `build_backbone_specs(config)` materialises the `BackboneSpec`
  tuple from the config block.
- components/c10_provisioning/__init__.py: re-exports the AZ-321
  surface and registers the new config block.

Tests
- test_engine_compiler.py: covers AC-1..AC-10 + missing-sidecar
  sibling case for AC-5. Tier-1 via fake runtime that writes through
  the REAL `Sha256Sidecar.write_atomic_and_sidecar`. Tier-2
  placeholders for the cache-hit p99 NFR (200 MB engine sweep) and
  kill-during-compile atomic-write NFR.

Docs
- module-layout.md: c10_provisioning Per-Component Mapping lists the
  new internal modules (engine_compiler.py, config.py), the
  composition-root c10_factory.py, the AZ-321 public re-export
  surface, and the registered config block.
- batch_33_cycle1_report.md + reviews/batch_33_review.md:
  PASS_WITH_WARNINGS (4 Low findings accepted).

Tests run: c10_provisioning 13 passing + 2 Tier-2 skips; combined
unit suite (excluding pending components) 543 passing, 21
env-skipped.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-05-13 00:09:53 +03:00

13 KiB
Raw Blame History

Batch 33 / Cycle 1 — Implementation Report

Date: 2026-05-12 Tasks: AZ-321 (C10 EngineCompiler — per-model TRT compile + hardware-tied cache reuse + AZ-281 filename schema + AZ-280 sidecar gate) Story points landed: 5 Status: complete (AZ-321 → In Testing)

Scope summary

Single-task batch landing the C10 per-model engine compile + cache- reuse orchestrator. EngineCompiler.compile_engines_for_corpus(req) walks the corpus, computes the canonical engine filename via AZ-281 EngineFilenameSchema.build(...), and either reuses the cached binary (cache hit; AZ-280 Sha256Sidecar.verify returns True) or delegates to the AZ-297 compile_engine method on the injected runtime (cache miss; the runtime owns the write path and the sidecar emission). The orchestrator returns one EngineCompileResult per backbone carrying the canonical EngineCacheEntry, the CompileOutcome.{BUILT,REUSED} label, and the compile_duration_s (None on reuse). Hardware-tied cache reuse (D-C10-6 / D-C10-7) falls out naturally from the filename schema: an engine compiled on (sm=87, jp=6.2, trt=10.3, fp16) lives at a different path than one compiled on (sm=89, jp=6.3, trt=10.5, fp16), so a hardware change produces cache misses for the new device and leaves the old files untouched (AC-4).

Two design corrections vs. the task spec body:

  • EngineCacheEntry shape — the task spec proposed a c10-local EngineCacheEntry with outcome and compile_duration_s fields. That clashes with the canonical AZ-297 _types.inference.EngineCacheEntry already re-exported from components.c10_provisioning. The canonical shape wins; the AZ-321 wrapper is renamed EngineCompileResult and carries {entry, outcome, compile_duration_s} cleanly.
  • InferenceRuntime.host_info() — the task spec calls a hypothetical host_info() method on the runtime to retrieve (sm, jp, trt). The AZ-297 Protocol does NOT expose host info. Rather than expand the frozen Protocol mid-cycle, we accept a HostCapabilities field on EngineCompileRequest so the composition root threads the host from its own probe (Tier-2 device introspection or Tier-1 test fixture). The compiler stays decoupled from any runtime-side introspection surface.

The C10 layer is also forbidden by test_az270_compose_root.test_ac6 from importing from components.c7_inference directly — that rule applies across all components/*/*.py files regardless of what the prose-level module-layout.md declares the "Imports from" list to be. The lint test wins. To respect it, engine_compiler.py defines CompileEngineCallable — a structural Protocol cut of InferenceRuntime exposing only the single compile_engine method the compiler actually uses — and catches the broader Exception class (the AZ-297 C7 error family stays the runtime's contract; the compiler dispatches on type(exc).__name__ in its ERROR log payload and re-raises so the original exception type propagates to the caller intact).

Files added / modified

New (production)

  • src/gps_denied_onboard/components/c10_provisioning/engine_compiler.pyCompileOutcome enum (BUILT / REUSED), BackboneSpec DTO, EngineCompileRequest DTO, EngineCompileResult DTO, EngineCompileSummary DTO, CompileEngineCallable structural Protocol, and the EngineCompiler class with the single public compile_engines_for_corpus method. Helpers: _build_config_for_backbone (synthesises one OptimizationProfile with min == opt == max == expected_input_shape from the backbone spec; richer dynamic-shape ranges are out of scope for AZ-321), _summarise (aggregate counts for the c10.engine.compile.summary log).
  • src/gps_denied_onboard/components/c10_provisioning/config.pyBackboneConfig DTO (model_name, onnx_path, expected_input_shape, input_name with "input" default) + C10ProvisioningConfig (backbones tuple, workspace_mb default 4096 to match C7 NFT-LIM-01). Both validate in __post_init__ (non-empty strings, positive shape dims, duplicate model_name detection).
  • src/gps_denied_onboard/runtime_root/c10_factory.pybuild_engine_compiler(config) wires the existing build_inference_runtime factory through to a new EngineCompiler instance with a c10-scoped structured logger; build_backbone_specs (config) materialises the BackboneSpec tuple from config.components['c10_provisioning'].backbones.

Modified (production)

  • src/gps_denied_onboard/components/c10_provisioning/__init__.py — re-exports the AZ-321 public surface (EngineCompiler, BackboneSpec, EngineCompileRequest, EngineCompileResult, CompileOutcome, EngineCompileSummary, CompileEngineCallable, BackboneConfig, C10ProvisioningConfig) and registers the new config block via register_component_block("c10_provisioning", C10ProvisioningConfig). CacheProvisioner / Manifest / EngineCacheEntry re-exports unchanged.

New (tests)

  • tests/unit/c10_provisioning/test_engine_compiler.pyNEW Tier-1 suite covering every AC + the 2 Tier-2 NFR placeholders:
    • AC-1 cold cache + 3 backbones → all BUILT; 3 .engine + 3 .sha256 files on disk; 3 c10.engine.cache.miss WARN logs; 1 c10.engine.compile.summary INFO log with engines_built=3.
    • AC-2 warm cache + identical request → all REUSED; compile_duration_s is None for every result; ZERO calls to the fake runtime; 3 c10.engine.cache.hit INFO logs.
    • AC-3 mixed (1 hit + 2 miss) — DINOv2 reused, LightGlue + ALIKED built; 2 calls to the fake runtime.
    • AC-4 hardware change (sm 87→89, jp 6.2→6.3, trt 10.3→10.5): every backbone rebuilt at the new filename; old files at the old filename untouched on disk.
    • AC-5 tampered sidecar (overwrite LightGlue's .sha256 with 0×64): LightGlue rebuilt; DINOv2 + ALIKED still reused; 1 c10.engine.sidecar.mismatch WARN log with model_name= lightglue and reason=digest_mismatch. Plus a sibling case where the sidecar file is deleted entirely (Sha256Sidecar.verify raises) — same WARN-then-rebuild outcome.
    • AC-6 EngineBuildError mid-corpus (backbone 2 of 3 fails): error propagates; backbone 1 (pre-cached, reused) untouched on disk; backbone 2's would-be engine NOT on disk (atomic-write guarantee from the fake mirrors AZ-298's real behaviour); backbone 3 never attempted (single call recorded for backbone 2).
    • AC-7 CalibrationCacheError propagates with the c10.engine.compile.error ERROR log carrying model_name, calibration_path, error_class=CalibrationCacheError.
    • AC-8 filename is exactly dinov2_vpr__sm87_jp6.2_trt10.3_fp16.engine (per AZ-281 canonical schema with the __ separator between model and sm); sidecar at *.engine.sha256 with 64-hex digest; EngineFilenameSchema.parse round-trip + Sha256Sidecar.verify both pass.
    • AC-9 compile_duration_s is a positive float for every BUILT result, None for every REUSED result.
    • AC-10 empty backbones tuple → empty result; ZERO runtime calls; ZERO files written; 1 summary log with all-zero counts.
    • NFR-perf-cache-hit Tier-2 placeholder skip (200 MB engine sweep belongs in the AZ-321 microbench harness on Jetson).
    • NFR-reliability-atomic-write Tier-2 placeholder skip (kill- during-compile scenario lives in the microbench harness; the atomicity contract itself is owned by AZ-280's tests).

The Tier-1 tests use a _FakeRuntime that satisfies CompileEngineCallable and writes deterministic engine bytes via the REAL Sha256Sidecar.write_atomic_and_sidecar — so the cache-hit / cache-miss / tampered-sidecar paths run against the same helper the production wiring uses. Only the C7-runtime-specific compile internals (TRT engine bytes, calibration cache, GPU memory) are mocked.

Modified (docs)

  • _docs/02_document/module-layout.md — c10_provisioning Per- Component Mapping now lists the new internal modules (engine_compiler.py, config.py) and the composition-root c10_factory.py; the Public API re-export list is extended with the AZ-321 surface; the Config block line is added (registered on import). default_provisioner.py row marked pending until the AZ-325 task lands.

Acceptance criteria coverage

AC Test Status
AC-1 Cold cache → all built test_ac1_cold_cache_compiles_every_backbone passing
AC-2 Warm cache → all reused, zero compile calls test_ac2_warm_cache_reuses_every_backbone passing
AC-3 Mixed cache test_ac3_mixed_cache_hits_and_misses passing
AC-4 Hardware change invalidates filename test_ac4_hardware_change_invalidates_cache passing
AC-5 Tampered sidecar + missing sidecar paths test_ac5_tampered_sidecar_invalidates_that_engine + test_missing_sidecar_treated_as_cache_miss passing
AC-6 EngineBuildError propagates, partial state consistent test_ac6_engine_build_error_propagates_and_third_backbone_untouched passing
AC-7 CalibrationCacheError propagates with diagnostic log test_ac7_calibration_cache_error_propagates passing
AC-8 Filename + sidecar layout matches AZ-281 schema test_ac8_filename_and_sidecar_layout passing
AC-9 compile_duration_s recorded for built only test_ac9_compile_duration_recorded_for_built_only passing
AC-10 Empty backbones → empty result, no side effects test_ac10_empty_backbones_returns_empty passing
NFR-perf-cache-hit p99 ≤ 1.5 s for 200 MB engine test_nfr_perf_cache_hit_p99_under_1500ms_for_200mb_engine (Tier-2) Tier-2 skipped
NFR-reliability atomic-write no half-engine after kill test_nfr_reliability_atomic_write_no_half_engine_after_kill (Tier-2) Tier-2 skipped

AC Test Coverage: 10 of 10 covered (+ 2 NFRs)

Code Review Verdict: PASS_WITH_WARNINGS (4 Low accepted; see Findings)

Auto-Fix Attempts: 0

Stuck Agents: None

Findings (self-review)

# Severity Category Location Note Resolution
1 Low Architecture engine_compiler.py::_compile_one Catches the broad Exception (not the specific AZ-297 RuntimeError family) because the c10 layer cannot import components.c7_inference (architecture rule test_az270_compose_root.test_ac6). The C7 contract scopes its runtime exceptions to its own family; ANY exception bubbling out of compile_engine is treated as a compile failure here. Re-raise preserves the original type. Inline comment documents the rule. Open (Low) — accepted; architecture rule wins.
2 Low Maintainability engine_compiler.py::CompileEngineCallable Duplicates the compile_engine method shape from the C7 InferenceRuntime Protocol. Mirrors the LightGlue dual-Protocol pattern already in _types/manifests.py (consumer-side structural cut vs. producer-side opaque marker). Open (Low) — accepted; matches established pattern.
3 Low Architecture engine_compiler.py ↔ C7InferenceConfig EngineCompileRequest.cache_root MUST equal the directory the C7 runtime writes to (C7InferenceConfig.engine_cache_dir). The composition root (build_engine_compiler + the C10 corpus driver T5 in AZ-325) is responsible for keeping the two in sync; the compiler itself trusts the request. A divergence would cause cache hits to always miss. Open (Low) — flagged for AZ-325 to enforce.
4 Low Scope engine_compiler.py::_build_config_for_backbone Synthesises exactly one OptimizationProfile with min == opt == max == expected_input_shape. Backbones requiring dynamic input ranges would need a richer BackboneSpec carrying explicit OptimizationProfile tuples. None of the AZ-321 corpus backbones (DINOv2-VPR, LightGlue, ALIKED) need dynamic shapes today, but the limitation is real. Open (Low) — accepted; future extension.

Tracker

  • AZ-321 transitioned to In Progress at session start; will move to In Testing post-commit per protocols.md.

Test suite

  • tests/unit/c10_provisioning/ — 13 passing, 2 Tier-2 skips (cache-hit p99 NFR + atomic-write kill scenario).
  • Combined unit suite excluding pending components (c1, c2, c2.5, c3, c3.5, c4, c5, c8, c11, c12) and the c6 collection blocker on this host (missing psycopg_pool is a known dev-machine env issue, pre-existing) — 543 passing, 21 environment-skipped, 1 warning (pre-existing pynvml FutureWarning unrelated to AZ-321).

Next batch

Cycle 1 advances per the greenfield queue — autodev re-detects the next AZ ticket in the Step 7 batch loop. AZ-321 unblocks AZ-322 (C10 Descriptor Batcher), AZ-337 (C2 UltraVPR), AZ-345 / AZ-346 / AZ-347 (C3 matchers), and AZ-349 (C3.5 refiner) at the topological level; the next ready batch is computed by compute-next-batch.

A cumulative review (batches 3133) will fire at the next sub-skill phase boundary per Step 14.5's K=3 trigger.