SPB Git

spb/localvm-research Public License

Running LLMs larger than memory on a consumer Mac — falsification-driven research: margin-gated deferred refinement, out-of-core verification on Apple Silicon. TR-01 published.

Python 63.2% JavaScript 23.5% CSS 11.8% Shell 0.9% Makefile 0.5%

History of experiments/candidate_01/implementation/runtime.py · clear filter

  1. candidate_01 scale: StreamingVerifier (per-layer materialize->compute->re-lazify) + 32B benchmark
    …
    q4-32B resident (17.5 GB) + q8-32B (34.8 GB) layer-streamed from SSD per
sweep; peak residency = base + a few layers; SSD sees one sequential
checkpoint pass per sweep. Judge: independent 8B bf16.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
    Simon-Pierre Boucher committed 5 h ago (Aug 12, 2026) · 1 file changed +8 −2
  2. candidate_01 v2: incremental q8 verify-cache (O(window) sweeps) + bf16 quality judge
    …
    Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
    Simon-Pierre Boucher committed 5 h ago (Aug 12, 2026) · 1 file changed +18 −3
  3. candidate_01: fix truncating final-window rollback; verify EOS decisions; loop safety bound
    …
    Diagnostic: q8 prefill-vs-decode argmax divergence = 1.56% — the
fidelity ceiling for any teacher-forced verification on Metal.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
    Simon-Pierre Boucher committed 5 h ago (Aug 12, 2026) · 1 file changed +30 −14
  4. candidate_01: margin-gated deferred-refinement runtime + benchmark + hypothesis
    …
    q4 resident base generates with margin recording; low-margin tokens
deferred; windowed q8 sweeps verify (margin mode) or verify-all
(QSpec-style exact); rollback via KV trim. Baselines: pure q4/q8.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
    Simon-Pierre Boucher committed 5 h ago (Aug 12, 2026) · 1 file changed +129