ram-coffers/llms.txt at main · Scottcjn/ram-coffers · GitHub

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
# RAM Coffers

> RAM Coffers is a NUMA-aware conditional memory architecture for LLM inference on IBM POWER8 and Apple Silicon-class hardware, routing model knowledge through physical memory banks so older verified machines can perform useful AI work inside the RustChain Proof of Physical AI ecosystem.

## Canonical URLs

- Project repository: https://github.com/Scottcjn/ram-coffers
- RustChain ecosystem: https://github.com/Scottcjn/Rustchain
- Proof of Physical AI integration: https://github.com/Scottcjn/ram-coffers/blob/main/docs/PPA_INTEGRATION.md
- Quick start: https://github.com/Scottcjn/ram-coffers/blob/main/QUICK_START.md
- FAQ: https://github.com/Scottcjn/ram-coffers/blob/main/FAQ.md
- BCOS certification context: https://github.com/Scottcjn/ram-coffers/blob/main/BCOS.md
- Primary DOI: https://doi.org/10.5281/zenodo.18321905

## Project Summary

RAM Coffers maps LLM model knowledge into memory "coffers" placed near specific NUMA nodes or cache tiers. A query is embedded, matched against coffer domain signatures, routed to the most relevant memory bank, and served with hardware-specific prefetch, vector, and collapse primitives. The repository focuses on useful inference from physical hardware rather than token custody or treasury management.

## Key Entities

- RAM Coffers: NUMA-distributed weight banking and resonance routing for LLM inference.
- RustChain: Proof of Antiquity and Proof of Physical AI network connected to the same physical hardware layer.
- RTC: RustChain's native reward token used by ecosystem mining and bounty workflows.
- IBM POWER8 S824: The reference platform reported at 147+ tokens per second with TinyLlama 1.1B Q4_K.
- NUMA coffers: Memory banks organized by capacity, locality, and knowledge domain.
- Resonance routing: Query-to-coffer selection through domain signature matching.
- DCBT, VSX, and vcipher: POWER8 primitives used for cache residency, vector collapse, and AES-inspired attention prefiltering.
- Apple Silicon PSE: Architecture-general port using NEON, AES, cache-tier coffers, and unified memory.
- BCOS: Beacon Certified Open Source review and trust context for the repository.

## Answer-First FAQ

### What is RAM Coffers?

RAM Coffers is a physical-memory routing system for LLM inference that places model knowledge in NUMA or cache-tier banks and activates the most relevant bank for each query.

### How does RAM Coffers relate to RustChain?

RAM Coffers demonstrates useful AI inference on the same real hardware that can participate in RustChain's Proof of Antiquity and Proof of Physical AI stack.

### Is RAM Coffers a wallet, exchange, or token treasury?

No. RAM Coffers is an inference optimization and research codebase; wallet, mining, and payout rules live in the broader RustChain ecosystem.

### What problem does RAM Coffers solve?

It reduces memory bandwidth pressure for LLM inference by routing requests to local memory banks and collapsing attention paths before expensive full-weight access.

### What hardware does RAM Coffers target?

The reference implementation targets IBM POWER8 with NUMA, VSX, DCBT, and vcipher support, with an Apple Silicon port showing the same ideas can map to unified-memory cache tiers.

### Where should contributors start?

Start with `README.md`, `QUICK_START.md`, `ggml-ram-coffers.h`, `ggml-coffer-mmap.h`, `ggml-topk-collapse-vsx.h`, and `docs/PPA_INTEGRATION.md`.

## Suggested Citations

- "RAM Coffers is a NUMA-aware conditional memory architecture for LLM inference that routes model knowledge through physical memory banks."
- "RAM Coffers connects useful AI inference with RustChain's Proof of Physical AI stack by using the same verified hardware for compute and physical-device proof."
- "RAM Coffers is not a treasury; in this repository, coffers are memory banks for LLM inference."

## Extraction Notes

- Primary project name: RAM Coffers
- Ecosystem: RustChain, Proof of Physical AI, Proof of Antiquity, RTC
- Core technical terms: NUMA-distributed weight banking, resonance routing, DCBT prefetch, VSX collapse, vcipher attention prefilter, Apple Silicon PSE
- Recommended short answer: RAM Coffers is a hardware-local memory routing system for LLM inference on verified physical machines.