Skip to content

CLAUDE.md

Latest commit

 

History

History
173 lines (135 loc) · 5.2 KB

CLAUDE.md

File metadata and controls

173 lines (135 loc) · 5.2 KB

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Project Overview

quantum-llama.cpp is a modified fork of llama.cpp that integrates Quantum Random Number Generators (QRNGs) into token generation. The core idea: "the output is co-authored by the universe itself."

Key difference from upstream: Uses true quantum randomness instead of pseudo-random number generation for sampling.

Build Commands

# Standard build (MUST use -DLLAMA_CURL=OFF for quantum features)
cmake -B build -DLLAMA_CURL=OFF
cmake --build build --config Release -j

# With CUDA
cmake -B build -DLLAMA_CURL=OFF -DGGML_CUDA=ON
cmake --build build --config Release -j

# Debug build
cmake -B build -DCMAKE_BUILD_TYPE=Debug -DLLAMA_CURL=OFF
cmake --build build

Built binaries are placed in build/bin/.

Testing

# Run test suite
ctest --test-dir build --output-on-failure -j

# Server tests (requires Python venv)
cd tools/server/tests
source ../../../.venv/bin/activate
./tests.sh

Code Formatting

Always format before committing:

git clang-format

Key style rules:

  • 4-space indentation, 120-column limit
  • Pointer/reference: void * ptr, int & a
  • snake_case for functions, variables, types
  • Enum values: ENUM_NAME_VALUE (uppercase with prefix)

Architecture

Core Directories

  • src/ - Main llama library (llama-*.cpp/h modules)
  • include/llama.h - Public C API
  • ggml/ - Vendored tensor library
  • tools/ - Executables (llama-cli, llama-server, etc.)
  • common/ - Shared utilities
  • libpsirngclient/ - Git submodule for gRPC QRNG client

Quantum Integration (in src/)

  • psirngclient-manager.cpp/h - Singleton managing QRNG connections
  • anu-qrng-client.cpp/h - HTTP client for ANU QRNG API (hex16 mode-based sampling)
  • llama-sampling.cpp - Sampling pipeline (integration point for quantum RNG)

Quantum RNG Flow

Token Logits
    ↓
Calculate Entropy (normalized 0-1)
    ↓
[entropy < 0.50?] ─YES─→ GREEDY (no QRNG) → Done
    │
    NO
    ↓
Apply EDT Temperature: T = T₀ × 0.8^(θ/entropy)
    ↓
ANU API call (hex16, length=1024, size=10)
    ↓
Find mode of ~20K uint8 values
    ↓
Use mode/256 for inverse CDF sampling
    ↓
Done

Key principle: Each token selection makes a fresh API call. No buffering - this preserves temporal correlation between consciousness and token selection.

Adaptive Entropy-Based Sampling

  • entropy < 0.50 → Greedy sampling (no API call, saves bandwidth)
  • entropy ≥ 0.50 → EDT temperature + QRNG sampling
  • Typically saves 50-80% of API calls for predictable text

EDT (Entropy-based Dynamic Temperature)

  • Formula: T = T₀ × 0.8^(θ/entropy)
  • Defaults: T₀=2.0, θ=1.0
  • Higher entropy → higher temperature (more creative exploration)
  • Lower entropy → lower temperature (more focused selection)
  • At max entropy (1.0): T ≈ 1.6

Running with Quantum RNG

ANU QRNG Setup (Required)

  1. Get your FREE API key at: https://quantumnumbers.anu.edu.au/

  2. Set the environment variable:

# Linux/Mac
export ANU_API_KEY="your-api-key-here"

# Windows CMD
set ANU_API_KEY=your-api-key-here

# PowerShell
$env:ANU_API_KEY="your-api-key-here"
  1. Run:
./build/bin/llama-cli -m model.gguf -p "prompt" -n 128 -no-cnv

Quantum CLI Arguments

Argument Description Default
--quantum-verbose Show entropy/temperature for each token off
--quantum-statistics Print sampling statistics at end off
--quantum-entropy-threshold N Entropy cutoff for greedy vs QRNG 0.40
--quantum-edt-t0 N EDT upper bound temperature 2.0
--quantum-edt-theta N EDT entropy sensitivity 1.0
--no-quantum-adaptive-sampling Always use QRNG (no greedy) -
--no-quantum-edt Use fixed temperature instead of EDT -

psirng Service (Alternative, requires external setup)

export PSIRNG_HOST=192.0.2.10
export PSIRNG_GRPC_PORT=50051
export PSIRNG_CERT_PATH=/path/to/cert.pem
./build/bin/llama-cli -m model.gguf -p "prompt"

Development Guidelines

  • NEVER buffer quantum random data - Each token selection MUST use fresh quantum data from a new API call. Buffering destroys the temporal correlation between consciousness and token selection.
  • Never use -DLLAMA_CURL=ON - incompatible with quantum features
  • Clone with --recurse-submodules to get libpsirngclient
  • Avoid adding third-party dependencies
  • Use basic C++ patterns, avoid fancy STL constructs
  • Vertical alignment for readability
  • Tensor storage is row-major (dim 0=columns, 1=rows, 2=matrices)
  • Matrix multiplication: C = ggml_mul_mat(ctx, A, B) means C^T = AB^T

Naming Conventions

  • Pattern: <class>_<method> where method is <action>_<noun>
  • Examples: llama_model_init(), llama_sampler_chain_remove()
  • Optimize for longest common prefix: number_small, number_big (not small_number)

Performance Validation

# Benchmark
./build/bin/llama-bench -m model.gguf

# Evaluate perplexity
./build/bin/llama-perplexity -m model.gguf -f dataset.txt

# Test backend ops
./build/bin/test-backend-ops