Land the fallback InferenceRuntime strategy that satisfies C7-IT-05:
when the TRT-direct path (AZ-298) cannot deserialise a cached engine
or when the operator explicitly selects ORT, the system stays in the
air at degraded latency rather than dropping the request. Conforms to
the AZ-297 Protocol; current_runtime_label() == "onnx_trt_ep".
Production
- onnx_trt_ep_runtime.py: compile_engine is a no-op returning an
EngineCacheEntry pointing at the source .onnx; deserialize_engine
is gate-first for .engine entries and gate-skip for .onnx, builds
an ORT InferenceSession with the provider list
[TensorrtExecutionProvider, CUDAExecutionProvider,
CPUExecutionProvider], stages cached engines into the ORT TRT EP
cache directory via symlink-or-copy, warms up with one session.run
after construction, and honours config.inference.ort_disallow_cpu_
fallback by raising EngineDeserializeError when the active provider
resolves to CPU; infer emits a one-shot c7.fallback_to_onnx_trt_ep
WARN log plus gcs_alert callback on first call when is_fallback=
True; release_engine is idempotent. _build_provider_args is the
single point that pins TRT EP option-key names (Risk-3) and caps
trt_max_workspace_size at gpu_memory_budget_bytes // 4 (AC-8).
- config.py: adds ort_trt_cache_dir (validated non-empty) and
ort_disallow_cpu_fallback to C7InferenceConfig.
- fdr_client/records.py: adds c7.fallback_to_onnx_trt_ep and
c7.cpu_fallback FDR record kinds.
Tests
- test_onnx_trt_ep_runtime.py: covers AC-1..AC-8 + Risk-2 CPU-fallback
alert + Risk-3 option-key pin + NFR-reliability error rewrap; Tier-1
via fake ORT session; Tier-2 placeholders skip on macOS dev for
numerical FP16 comparison and session-creation perf NFR.
- test_protocol_conformance.py: drops onnx_trt_ep from the missing-
module parametrize now that the module ships.
- test_az272_fdr_record_schema.py: extends per-kind fixture builder
to cover the two new C7 FDR kinds in the roundtrip / schema-version
AC tests.
Docs
- module-layout.md: replaces the pending onnx_trt_runtime row with
the shipped onnx_trt_ep_runtime row + capabilities list.
- batch_32_cycle1_report.md + reviews/batch_32_review.md: full batch
+ self-review (PASS_WITH_WARNINGS, 4 Low findings accepted).
Tests run: c7_inference 139 passing + 17 Tier-2 skips; combined unit
suite (excluding pending components) 529 passing, 19 env-skipped.
Co-authored-by: Cursor <cursoragent@cursor.com>
Implement the production-default InferenceRuntime strategy on JetPack
6.2 + TensorRT 10.3 (per D-C7-9). The runtime owns the full TRT
lifecycle: compile_engine via the Polygraphy + trtexec + IBuilderConfig
hybrid (FP16 / INT8 / Mixed precision), deserialize_engine with
EngineGate-first ordering and a pre-allocation GPU memory budget gate,
infer via H2D -> enqueueV3 -> D2H -> stream sync on the owned CUDA
stream, idempotent release_engine, and an injected
ThermalStatePublisher delegation for thermal_state.
INT8 calibration cache trust (D-C10-6, AC-2/3/4) is enforced by a
.calib_cache.sha256 file-integrity sidecar (AZ-280) plus a new
.calib_cache.dataset_sha256 sidecar that records the dataset content
hash at compile time; reuse only when both agree, rebuild silently on
dataset hash mismatch, raise CalibrationCacheError on corrupt sidecar
(never silently overwritten).
GPU memory budget (NFT-LIM-01, default 4 GiB) is checked BEFORE any
TRT call beyond the gate (AC-6); a pre-allocation refusal raises
OutOfMemoryError and leaves the resident state unchanged.
TensorRT 10.3 / Polygraphy / PyCUDA are lazy-imported inside the
methods that need them so the module loads cleanly on Tier-0 hosts.
A standalone CLI entry (python -m
gps_denied_onboard.components.c7_inference.tensorrt_runtime compile
<onnx> <build_config.json>) is wired for C10 CacheProvisioner
(AZ-321) to invoke pre-flight without holding a runtime instance.
C7InferenceConfig gains gpu_memory_budget_bytes (default 4 GiB) and
trtexec_timeout_s (default 600 s, Risk 4 mitigation), both validated
in __post_init__.
Tests: 26 active + 6 Tier-2-gated skips; AC-1 / AC-3 / AC-4 / AC-5
/ AC-6 / AC-7 / AC-10 + NFR-reliability fully covered on Tier-1
via fake CUDA / TRT modules; AC-2 / AC-8 / AC-9 / NFR-perf-deserialize
placeholders skip with prerequisite reason and live in the AZ-298
Tier-2 microbench harness. Code review verdict
PASS_WITH_WARNINGS (1 Medium hot-path hoist fix auto-applied).
Co-authored-by: Cursor <cursoragent@cursor.com>
Implements AZ-297 InferenceRuntime's thermal_state() side: a singleton
background-thread publisher that polls jtop (preferred) or pynvml
(fallback) at config.thermal_poll_hz, stores an atomic ThermalState
snapshot, and emits c7.thermal_transition FDR records on every throttle
flip with a WARN log on entry and an INFO log on exit. Default-safe on
TelemetryUnavailableError per Invariant I-6 with a 1-Hz rate-limited
WARN.
Sources return a raw ThermalReading; the publisher stamps measured_at_ns
via its injected Clock so _JtopSource / _PynvmlSource stay clean of
direct time.* calls (Invariant 2). _poll_once is the deterministic test
seam — start() spawns the production thread.
- c7.thermal_transition registered in fdr_client.records KNOWN_PAYLOAD_KEYS
- [telemetry] optional dep group (jetson-stats, pynvml) added to pyproject
- 14 unit tests (AC-1..AC-6, AC-8, NFR-default-safe, structural)
green; AC-7 / AC-1 microbench / NFR-perf-poll Tier-2 deferred
- full unit suite: 1140 passed, 11 expected Tier-2/CUDA skips
Co-authored-by: Cursor <cursoragent@cursor.com>
AZ-301 takeoff-side validator every InferenceRuntime strategy calls
before deserialize_engine. Five-step deterministic refusal pipeline,
in order:
1. filename schema parse -> EngineSchemaMismatchError(reason=...)
2. schema tuple match -> EngineSchemaMismatchError(expected,got)
3. sidecar present -> EngineSidecarMissingError
4. sidecar trust -> EngineHashMismatchError(stage=sidecar)
5. manifest match -> EngineHashMismatchError(stage=manifest)
Refusal order is part of the public contract (AC-7 verifies a
fixture that is BOTH schema-mismatched AND missing-sidecar refuses
at step 1).
Production code (new):
- components/c7_inference/engine_gate.py -- EngineGate, HostTuple,
read_host_tuple (Jetson: pynvml + /etc/nv_tegra_release +
tensorrt.__version__; raises RuntimeError on Tier-1)
- components/c7_inference/manifest.py -- DeploymentManifest,
ManifestReader, ManifestReaderProtocol. Risk-2 enforced at the
type level: __getitem__ raises EngineHashMismatchError on
missing key, NEVER KeyError, so the gate cannot silently pass
- components/c7_inference/__init__.py -- re-exports the new
public surface
Tests (new): tests/unit/c7_inference/test_engine_gate.py covers
AC-1..AC-7 + NFR-reliability-no-write + manifest reader + refusal
log emission. 14 tests unconditional + AC-8 Tier-2 skip (needs
real NVML + L4T release file + tensorrt binding).
Three task-spec -> as-built deltas documented in
_docs/02_tasks/done/AZ-301_c7_engine_gate.md Implementation Notes:
1. HostTuple lives in engine_gate.py (the only consumer);
re-exported from package __init__.py.
2. read_host_tuple takes precision as a keyword argument — three
of four fields come from the host, precision is engine-build
metadata supplied by the caller.
3. AC-8 is Tier-2-only; AC-1..AC-7 + NFR-reliability + extras
run on every CI host.
Risk-2 (manifest reader silently treats missing entry as pass):
DeploymentManifest.__getitem__ raises EngineHashMismatchError with
"missing manifest entry for {path}" — covered by
test_manifest_missing_entry_raises_hash_mismatch.
NFR-perf-validate (p99 <= 50 ms): tier-2 only — a real 500 MB
engine streaming sha256 cannot be benchmarked on Tier-1 fixtures.
AZ-302 (ThermalStatePublisher) + AZ-304 (C6 Postgres schema)
deferred to batches 26 / 27 to keep the 1-task batch cadence and
isolate their respective env / testcontainer surface areas.
Suite: 1134 passed / 11 skipped. No regressions outside the new
files.
Co-authored-by: Cursor <cursoragent@cursor.com>
Freezes the c7_inference Public API per
_docs/02_document/contracts/c7_inference/inference_runtime_protocol.md
v1.0.0:
- InferenceRuntime Protocol (6 methods: compile_engine,
deserialize_engine, infer, release_engine, thermal_state,
current_runtime_label) in components/c7_inference/interface.py.
- DTOs (PrecisionMode enum, OptimizationProfile, BuildConfig,
EngineCacheEntry, EngineHandle opaque marker) in _types/inference.py
— placed at the L1 types layer so C10 can re-export EngineCacheEntry
without crossing the components.* boundary (AZ-270 AC-6).
- ThermalState DTO expanded in _types/thermal.py from the AZ-355
forward-declared stub to the AZ-297 contract shape (cpu/gpu temp,
thermal_throttle_active, measured_clock_mhz, measured_at_ns,
is_telemetry_available). Invariant I-6: when telemetry is
unavailable, throttle is False.
- Error family rooted at c7_inference.errors.RuntimeError (9 subtypes:
EngineBuildError, EngineDeserializeError, EngineHashMismatchError,
EngineSchemaMismatchError, EngineSidecarMissingError,
CalibrationCacheError, InferenceError, OutOfMemoryError,
TelemetryUnavailableError). RuntimeNotAvailableError stays in
runtime_root/errors.py — composition-time, outside the family.
- C7InferenceConfig per-component config block (runtime label,
thermal_poll_hz, engine_cache_dir) with constructor-time validation
rejecting unknown runtime labels.
- Composition-root factory build_inference_runtime in
runtime_root/inference_factory.py with three BUILD_* gates
(BUILD_TENSORRT_RUNTIME, BUILD_ONNX_TRT_EP_RUNTIME,
BUILD_PYTORCH_FP16_RUNTIME). Concrete strategy modules are imported
lazily via __import__ AFTER the flag check, so a Tier-0 build with
the flag OFF MUST NOT load the strategy module (AC-5 / I-5;
verifiable via sys.modules).
- 37 conformance tests cover all 8 ACs + NFR-perf-factory
(p99 build under 200 ms × 1000 calls) + NFR-reliability-error-family.
AC-8 introspects the contract file's Shape table and asserts method
parity against the runtime Protocol; also asserts all 9 error
subtypes are documented.
Retired the AZ-263 scaffolding EngineCacheEntry from _types/manifests.py
(replaced by the AZ-297 canonical shape in _types/inference.py); updated
the LightGlue-flavoured EngineHandle Protocol docstring in
_types/manifests.py to rationalize its intentional dual existence
with the C7 opaque EngineHandle (same name, different consumer-side
cut, mirroring the C4/C5 ISam2GraphHandle pattern).
Stale ThermalState.throttle docstring references in c4_pose/config.py,
c4_pose/interface.py, and _types/pose.py updated to
thermal_throttle_active.
Full unit-test sweep: 843 passed, 2 pre-existing environment skips
(cmake, actionlint).
Co-authored-by: Cursor <cursoragent@cursor.com>
Oleksandr Bezdieniezhnykh