File size: 3,856 Bytes
e1acc15 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 | # 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.
|