| # PP4 measured selector qualification |
|
|
| `fused-qualification.json` is the final hash-bound synthetic solo receipt. |
| It matches the committed `sampled_fused.py` and `fused_probe.py`. All16 cases |
| pass (B1–B4, depth4/5, all-positive and mixedT0/T1), with360 exact conditional |
| position comparisons and1,600 graph replays. Tokens, support IDs and every |
| dyadic q value match the original materialized Markov+PP3 producer reference. |
| Graph/eager outputs match exactly under the same keys. Ties, signed zero, |
| NaN/inf/all-masked refusal and safe continuation after an invalid draw pass. |
| Both physical-peer specializations compile without executing peer traffic; |
| this is explicitly not physical-pair qualification. |
|
|
| MEASURED on x86host's assigned bc62df2a at unchanged225W, CPython3.13.12, |
| Torch2.12.1+cu130. Peak RSS979,392KiB; peak CUDA allocation264,765,440B. |
| Runtime5.02s in the warmed compiler cache; qualification follows the final typing-only qualifier edit. Synthetic all-positive selector |
| medians (including both logical shards and Markov, excluding backbone/pair): |
|
|
| |B|depth4 ms|depth5 ms| |
| |---|---:|---:| |
| |1|1.021952|1.346560| |
| |2|1.727488|2.307072| |
| |3|2.014208|2.676736| |
| |4|2.686976|3.578880| |
|
|
| These uncached-heavy synthetic costs are neither whole-worker timings nor |
| incremental serving costs. Whole-worker timing uses real N4 states separately. |
| No kernel timings are transferred to GB10 or a physical pair. |
|
|
| The frozen pre-change CPU selection passed339 tests with41 existing |
| GPU/checkpoint/optional-lib skips. The same changed selection passed343 with |
| the same41 skips before two additional priming regressions; the final focused |
| dispatch selection passes6/6. The initial unfiltered worker selection also |
| attempted39 explicitly GPU-only fused tests with CUDA hidden, all failing with |
| no-device errors; no assertion or skip was changed. The selected CPU regression |
| omits that file explicitly. Logs remain in the campaign directory. |
|
|
| Strict mypy covers the new fused runtime, GPU qualifier and CPU dispatch tests |
| with the existing narrow Triton JIT stubs; legacy tp2 is not claimed newly |
| strict in its entirety. New runtime boundaries are annotated. Ruff, shell |
| syntax and diff checks pass. |
|
|
| Reproduce only on the designated solo: stage code into PP4 x86host scratch, |
| set `PP4_SOURCE_ROOT`, start an owned detached named tmux with private |
| `TMUX_TMPDIR` before server creation, then run `fused_probe_locked.sh --out` |
| with an output beneath PP4 scratch. The launcher takes the existing solo lock |
| with `flock -w3600`, checks occupancy/225W and uses the required interpreter, |
| affinity/nice/thread limits and private caches. It changes no power or unit. |
| Preserve aggregate receipts before deleting scratch. CPU tests use TMUX unset, |
| private TMPDIR/TMUX_TMPDIR, CUDA hidden, bytecode disabled, and repository root |
| plus `offload/dspark` on PYTHONPATH. |
| |
| |
| ## Round2 CPU validation |
| |
| Tests in offload/launch/windows/tests/test_pp4_{window,remote,prepare,pair}.py |
| cover receiving-host second-hop stop/run, timed-out recovery/retry, delayed |
| pair-open, exact-HOLD deferred close/retention, strict current-production env |
| mounts, both actual K1 None branches and exact base reconstruction, failed |
| counter growth/reset/missing/late increments, physical receipt case/teardown |
| coverage, pair-first failure restoration, two restored ABBA segments and |
| measured whole-block budget refusal. No new hardware skip qualifies a path. |
| |
| New pair_probe/pair_qualify scripts are syntax/type inspected only on Controller. |
| Their MEASURED physical labels are emitted only when real checked subprocesses |
| finish on the exact locked pair; no such receipt was generated by this lane. |
| Synthetic CPU receipts are confined to test fixtures and never admission. |
| The coordinator must obtain independent review and actual hardware/GB10 |
| qualification; a passing CPU suite is not execution clearance. |
| |
