rustane warm-page-cache pread fanout (ncdrone/rustane)
Parent: Mac local LLMs: MoE streaming and offload · Published reference · snapshot 2026-10-05
↓ Facts as markdownall context files
Anemll's docs credit rustane as the inspiration for cached-read fanout, describe the idea in prose and publish no rustane figures.
These notes link each claim to its source. A source may be a research report hosted on this site rather than the primary document. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.
Facts
- Anemll's docs credit rustane as the inspiration for cached-read fanout, describe the idea in prose and publish no rustane figures. [source]
- ncdrone/rustane describes itself as a Rust training and inference engine for the Apple Neural Engine and Metal GPU, validated for training from 48M to 5B parameters on an M4 Max 128 GB using reverse-engineered private ANE APIs. [source]
- The rustane README sections are Benchmarking, Checkpoint Inference (a generate CLI with Metal KV-cache decode) and HTTP Serving (OpenAI-style completion routes); none concerns expert streaming or page-cache reads. [source]
- The rustane results directory holds ANE and GPU TFLOPS, IOSurface staging, dual-load, f16 conversion, training probe and Rust-versus-Objective-C comparison notes, and no pread or page-cache note. [source]
- The rustane bench directory holds three Python scripts (dual_load.py, gpu_metal_matmul.py, gpu_tflops.py). [source]
- rustane's CREDITS file lists Anemll for ANE inference tricks and does not mention Flash-MoE or page-cache reads. [source]
- Anemll says its cache I/O split experiment "is not a code import from rustane" and is built on Flash-MoE's existing async routed-expert pread path. [source]
- Anemll's cachebench note says its scaling curve is "close to the pattern reported in rustane", naming only the shape, one worker near raw-storage speed then a sharp multi-worker climb. [source]
- The Anemll cachebench with split greater than 1 schedules the same total bytes as split 1 (for 128 experts and split 4, 512 tasks instead of 128), so the gain comes from more concurrent cached reads, not extra data. [source]
- ncdrone/rustane master shows its last commit on 2026-04-03 and 181 stars at the 2026-10-04 fetch. [source]
- Nothing in rustane's public default branch supports or refutes the warm-cache pread scaling curve; the Anemll cachebench is the only measured evidence for that curve. [source]
Children
- No children recorded.