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%
-
expH: adaptive per-cell budget (~3s wall) + iostat ground-truth logger
…
Short cells (~10ms wall) produced impossible throughput readings; iostat log validates that measured bytes came from the disk controller. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
-
expH fix: write test file with F_NOCACHE — first run measured page cache, not SSD
…
F_NOCACHE on a read fd does not bypass already-resident pages; writing the test file normally made every page resident, so 'cold' cells hit unified memory (42-115 GB/s, impossible for NVMe). Also recreate the file before the GPU-load phase since the warm phase repopulates cache. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
-
expH: SSD feasibility benchmark implementation + hypothesis
…
Measures internal NVMe read throughput: block sizes 4KB-4MB, random vs sequential, F_NOCACHE vs warm page cache, 1/4/8 threads, and key cells under concurrent MLX Metal matmul load. Embeds hardware manifest. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
-
Bootstrap: charter-compliant skeleton, header enforcement, hardware manifest, experiment scaffolds
…
- Repository layout per CLAUDE.md §3 - tools/check_headers.py enforcing §0.1 (58 files pass) - tools/new_experiment.py scaffolding with seven-field hypothesis block (§10) - benchmarks/hardware_manifest.py: macOS-native fingerprint (M5 Max / 48GB verified) - Eight micro-experiment scaffolds (§9, exp A–H) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>