Validation
Validation layers
| Layer | What it establishes |
|---|---|
| CPU contracts | Routing, configuration, reports, discovery and mocked ownership |
| Native CPU contracts | File/storage behavior and Linux admission observations |
| Small CUDA contracts | Payload correctness, registration reuse, event ownership, close refusal, interruption and idle trim |
| Model validation | Correct output and public reports for a fixed model/workload |
| Installed wheel | Discovery, linkage and separately tested native behavior of the actual packaged bytes |
Compilation and source-checkout tests do not establish installed-wheel acceptance. A successful model does not certify all routes, quantizers or platforms.
Installed-package check without CUDA
From the checkout, run scripts/probe_installed_native_cpu.py with the installed
environment's Python and -I -B. It checks the native library hash and automatic
discovery, then simulates the child-process bootstrap without loading the DLL/.so
or running CUDA. This is a maintainer release check, separate from GPU tests and
complete model runs.
Selected installed-wheel acceptance
The current published pair is 0.1.0rc3; see its
final artifact hashes and registry checks.
The final distribution passed strict metadata validation, clean Windows and
Ubuntu 22.04 container CPU installation checks, and a clean Windows PyPI
installation/discovery check without CUDA. Model validation below identifies
the original wheel containers and native payloads separately; registry
publication does not broaden their acceptance scope.
The 0.1.0rc2 allocator/DDO candidate pair has selected installed allocator
smoke-test acceptance on Windows, native Linux and WSL. The unchanged Linux
wheel was used on both Linux targets.
| Target | Native artifact SHA256 | Scope |
|---|---|---|
| Windows AMD64, CUDA 13.2 | 4354a4dce319559c3cbb3533554798d688e9b779f2ca08991f30e28309e69c0b |
Installed discovery and idle-pool trim |
| Linux x86_64 / WSL2, CUDA 13.2 | ea8f8cc61b2303c4921f1527eae53250833613c83db48b5d2111f1d79309ab59 |
Installed discovery, dependency resolution and idle-pool trim |
The reproducible contract is
ddo-vmm-allocator/scripts/vmm_idle_trim_probe.py --library <installed-library>.
It preserves a live payload while reclaiming idle backing, permits allocation
after trim and checks final ownership/failure counters. It is not a model or
source-cache performance test.
Hardware and environments records the selected test host. Package delivery describes candidate generation and identity. These results do not establish dependency-complete model integration, arbitrary driver support or manylinux compliance.
Local rc3 Linux wheel pair
The locally assembled 0.1.0rc3 DDO/allocator pair has selected installed-wheel
LTX image functional acceptance on native Linux and WSL2. Both targets used the
same Docker-produced wheels without rebuilding the native library:
| Artifact | SHA256 |
|---|---|
| DDO wheel | 7b5554f2b0c9cb6a2863f02d253b7aedd1dbe08ceff74ba5bcceefe4dc9634ec |
| Linux allocator wheel | 1d82ef5011030aa4052d7b13ad74ed65751652da58a939d534ecf5bddc6b0240 |
| Packaged native ELF | 71183e987bb8b1d6a252c71ab938589ee89ac13f1e48339cb184731569d55535 |
Each target completed its fixed BF16, 1280 x 704, eight-step, seed-43 image workload including decode, exited 0 and matched its retained rc2 image exactly. Expected component routes and capacities were checked; available pre-close native failure/quarantine counters were zero. Final native counters and latent finiteness were not exposed by the unchanged runner. Returned cleanup and process exit do not prove final driver-memory release.
The WSL workload used bounded registration; native Linux used full-source registration. Do not infer equal capacities or compare their speeds as matched workloads. Native Linux also emitted a preserved group-offload lazy-prefetch warning for unused vision-tower layers; the text-prompt workload does not cover image-conditioned execution. See the platform-specific model conditions.
Separately, the installed Linux library passed the small explicit execution-slot reclaim/recreation gate on WSL: two 4 MiB buffers, 8 MiB physical/24 MiB virtual capacity, one 4 MiB release/recreation, stable address and exact payloads, with leased/pending refusals and checked pool close. This is explicit-reclaim acceptance; the model runs do not establish automatic pressure-trigger coverage. Automatic DXGI reclaim remains Windows-only and disabled by default.
These are development artifacts, not a published rc3 or manylinux certification. Model timings are observations, not performance acceptance. Native Linux, WSL, Windows, full video and arbitrary models remain separate validation scopes.
Source-level coverage
Updated execution-backing candidate
The subsequent local rc3 Windows allocator wheel has selected installed Qwen
image acceptance with opt-in execution-backing reclaim. Its native SHA256 is
eb0e3d8241817853a724cca11c294c3d98540b20e0285e36d0c0640cb29c995b;
wheel SHA256 is
45e361c6ea2706bd810ce408306e889e2b8465be4aafba8d8e56f6b284e86644.
The fixed BF16, 1024-square, 20-step workload matched its retained image exactly,
completed source-ticket collection and recorded 29 cumulative transformer
pressure queries with zero failures or unavailable observations. No reclaim
trigger occurred; this model result establishes pressure-path reachability,
not release under pressure. A separate small native gate verified idle-only
release/recreation, refusal of leased/pending buffers and exact payloads.
A new Docker-produced Linux wheel pair also passed selected installed LTX image checks separately on WSL and native Linux, using the same bytes:
| Artifact | SHA256 |
|---|---|
| DDO wheel | b80068a72b8bbb31a0ff1a6700b3958258bc05fc9bb5ef3c22f7357deec0e482 |
| Linux allocator wheel | 125865e1cb9de65e9249ddd12cafbe4086eb28b53d86e83d7898768e112324a2 |
| Native ELF | 3ac4cacb966dacc8ee7e96bf6502d526173e9d8497654272d2bd60f4c9044289 |
Both targets completed the eight-step BF16 1280 x 704 workload, matched their retained reference pixels and reported zero available pre-close native failure and quarantine counters. Original platform-specific staging/registration limits were preserved. Final native counters and latent finiteness remain unavailable; process exit and returned cleanup do not prove final driver-memory release. The native Linux text-prompt run retained the warning about unexecuted vision-tower layers; image-conditioned execution remains unvalidated.
The reuse-pressure hook is Windows/DXGI-only and compiled out on Linux/WSL. These are functional development-wheel results, not matched performance, automatic Linux pressure-policy acceptance, manylinux certification or publication.
Public tests cover policy/precedence, loading/source association, reports, dense convolution ownership, source-cache/event lifetimes and bounded cleanup. Saved TorchAO FP8 has separate payload/root-state/interruption contracts; BF16 success does not imply FP8 success. Refer to FP8 support for the supported recipe.
Native Linux admission must preserve overcommit mode, physical/process bounds and finite cgroup limits. Unknown observations refuse growth. WSL additionally requires bounded host-registration validation; an observed capacity is not a universal preset.
Full model comparisons and their exact reproduction conditions are maintained in DDO examples, not duplicated here.
Residual CPU source ownership
The explicit detach_auxiliary_cpu_sources contract has CPU coverage for byte
preservation, aliases, source provenance, pre-copy budget refusal, transactional
rollback and protection of retained native CPU payloads. A mapped-storage
regression checks that tiny registered residual aliases release their large
original storage owner after detachment, without treating storage extent as a
measurement of recovered RAM or private commit.
Selected Windows two-step model validation also checked complete finite, bitwise-identical video/audio latents and safe closure after bounded detachment. Its commit observations and limits are recorded in the model measurement details. This does not establish full inference/decode, other platforms/models, wheel acceptance, automatic enablement or a speed improvement.
Acceptance rules
Use manual validation to record routes, versions, native identity, output and measurement scope. Missing fields are unavailable, not zero. Host callbacks, CUDA events and whole-pipeline wall time remain distinct.
Successful inference does not prove recovery from a poisoned CUDA context, all-platform support, long-run stability or final release of driver memory. Retain failed-run diagnostics separately from accepted performance results.