Skip to content

aprender 0.66.0 - the honest-GPU model release

Choose a tag to compare

View all tags
@noahgift noahgift released this 10 Sep 05:39
v0.66.0
53dd489
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Pre-built CUDA binaries (Linux)

Rebuilt 2026-09-10 14:38Z on the fleet's disposable GPU runners (gx10 aarch64, yoga x86_64) inside rust:1.93.0-bullseye: measured glibc floor GLIBC_2.30 on both assets (objdump -T), so they start on Ubuntu 22.04 (glibc 2.35) and newer; sha256 verified, apr --version → 0.66.0, --features cuda proven in the bytes (libcuda loader present). Earlier native builds needed GLIBC_2.43 and did not start on 22.04.

apr built with --features cuda, built on the fleet's own GPU boxes (aarch64 on gx10, x86_64 on yoga) and smoke-tested on both. The NVIDIA driver (libcuda.so.1) is loaded at runtime; the glibc floor of each asset is printed in its build job's summary.

target asset
x86_64-unknown-linux-gnu apr-v0.66.0-x86_64-unknown-linux-gnu-cuda.tar.gz (+ .sha256)
aarch64-unknown-linux-gnu apr-v0.66.0-aarch64-unknown-linux-gnu-cuda.tar.gz (+ .sha256) 
curl -LO https://github.com/paiml/aprender/releases/download/v0.66.0/apr-v0.66.0-x86_64-unknown-linux-gnu-cuda.tar.gz
curl -LO https://github.com/paiml/aprender/releases/download/v0.66.0/apr-v0.66.0-x86_64-unknown-linux-gnu-cuda.tar.gz.sha256
sha256sum -c apr-v0.66.0-x86_64-unknown-linux-gnu-cuda.tar.gz.sha256
tar xzf apr-v0.66.0-x86_64-unknown-linux-gnu-cuda.tar.gz && ./apr-v0.66.0-x86_64-unknown-linux-gnu-cuda/apr --version

From source instead: cargo install aprender --version 0.66.0 --features cuda (the default cargo install aprender is CPU only).

The honest-GPU model release. Milestone 0.66.0 was narrowed by the operator
on 2026-09-08 to the two P0 defects a user hits first, plus the CI change that
makes the GPU crates' default-feature tests part of the required check. The
PP-066 obligation rows (backend registry, release assets, installer, C0 gates)
moved to 0.68.0 on 2026-09-09; no speed number ships in 0.66 — instruments
and speed ship later (PP-066 claims ratchet).

Fixed

  • apr chat answered with the toy demo model for a sharded SafeTensors
    index     (#3022, #3024; PP-066 row F-1, PMAT-1080, #3050). Path::extension()
    on model.safetensors.index.json — the exact file apr pull writes and then
    recommends — is json, which matched no arm and fell through to the demo
    model while the banner still printed the real path. One decision
    (resolve_chat_format: suffix, then magic bytes, then a refusal with exit 6)
    replaces the two that never compared notes; Demo is no longer an outcome for
    a path that exists. Six-row case table, both polarities; before/after records
    in evidence/format-honesty/.
  • GPU inference computed a different function than CPU for Qwen2.5-1.5B
    (hidden 1536, 12 heads, 2 KV heads)     (#2971, #3017; PP-066 rows L0-1a/L0-1b,
    #3026, #3032). Every model in evidence/models/supported.yaml now computes the
    same function on GPU as on CPU over at least 64 positions or the GPU refuses
    it (C14, scripts/check_model_parity.sh --manifest; SKIP_PARITY_GATE is a
    printed override that never passes). apr parity --per-op names the first
    diverging op; for the 1.5B it was the post-FFN residual in layer 26 and the CPU
    Q8_K reference was the inaccurate side. Measured green on lambda (sm_89) and
    gx10 (sm_121); records under evidence/parity/l0-1/.
  • apr-cli's integration surface went 14 red to 0 on a clean main (#3051,
    #3053) — three root causes, each fixed as a guard; the generated 28 MB
    test.apr is untracked and the race it hid is fixed (#3059).
  • main was red under pmat 3.39.0 (#3028, #3030): twelve legacy nested
    subtask records the new validator refuses as duplicate ids are gone, and id
    uniqueness is checked in-repo, independent of the analyser pin.
  • Silicon Nightly tested a package that has never existed (#2793,
    paiml/infra#361).

Added

  • aprender-gpu and aprender-cuda-edge run in workspace-test (#3063;
    T0 of the NVIDIA CUDA Rust spec, #3062). Both are default = [] with no
    build.rs; their default-feature tests had never been in a required check.
    The cuda-gated modules (driver, kernels, memory, ptx) still run only under
    --features cuda (#3067).
  • Build-system enforcement (BSE-001 M2, BSE-17) (#3037, #3039, #3044): a
    fail-closed composite make gate pinned to origin/main, the guard_tree
    dispatcher and guard-tree job, predict_merge, sorted-insert for the
    roadmap, asserted tool pins; ratchet verdicts are a function of (comparand,
    merge) only; two-tier tests — quick on the PR, full behind it.
  • PP-066 guards — G-10: the shipped-path ratchet runs under one pinned
    analyser with a stamped baseline (#3011); G-11: row PRs never write the DAG,
    roadmap, spec block or README counts — DAG status is derived (#3020, #3012);
    G-4/G-6/C0-7: the obligation DAG as data with invariants in CI, the
    roadmap-additive guard, the receipt terminal marker (#2987, #2981).
  • PR review receipts judged from the base (#2985, C0-5): one base-owned
    quorum workflow on pull_request_target and merge_group.

Documentation

  • PP-066 release spec v1.5 → v1.6 with the S0 discovery ledger (23 premises
    measured), the 0.66 parity report and the 0.65.2 post-publish host receipts
    (parity NO-GO, measured) (#2872, #2875, #2868, #2871, #3000, #2858).

Release receipt: docs/audits/impl-PMAT-1096-receipt.md · quorum record: docs/audits/quorum-PMAT-1096-2971.json

🤖 Generated with Claude Code

https://claude.ai/code/session_01BjhtNUSensCYpQb3mCYLod

Assets 14
Loading