ADu2021/fairy2i-complex-llm-quantization
skills/skillxiv-v0.0.2-claude-opus-4.6/fairy2i-complex-llm-quantization/SKILL.md
Enable extreme 1-2bit quantization of pre-trained LLMs by leveraging complex-valued arithmetic. Convert real-valued linear layers to complex domain losslessly, quantize to fourth roots of unity {±1, ±i}, and apply recursive residual error quantization for near full-precision performance.
$ install --global
skillsauthnpx skillsauth add ADu2021/skillXiv fairy2i-complex-llm-quantizationInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
Security Scan Results
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
3
Scanners Passed
9
Scanners in report
Clean
TrivyContainer and dependency vulnerability scanner
95%
Clean
SemgrepStatic code analysis for vulnerabilities
95%
Clean
mcp-scan (Snyk)Model Context Protocol security validation
95%
Skipped
Snyk (dep)Open source security scanning
50%
Skipped
Socket.devSupply chain security analysis
50%
Skipped
VirusTotalMulti-engine malware detection
50%
Skipped
CrowdStrikeAdvanced threat intelligence
50%
Skipped
OSV-ScannerOpen Source Vulnerability database check
50%
Skipped
OWASP Dep-Check
50%
Last scanned: Apr 24, 2026, 9:01 PM1.8s1 file scanned
SKILL.md
- name:
- fairy2i-complex-llm-quantization
- title:
- Fairy2i: Extreme Quantization of LLMs via Complex-Valued Arithmetic and Phase-Aware Projection
- version:
- 0.0.2
- engine:
- skillxiv-v0.0.2-claude-opus-4.6
- license:
- MIT
- url:
- https://arxiv.org/abs/2512.02901
- keywords:
- [quantization, complex-arithmetic, LLM-compression, 1-2bit, phase-aware]
- description:
- Enable extreme 1-2bit quantization of pre-trained LLMs by leveraging complex-valued arithmetic. Convert real-valued linear layers to complex domain losslessly, quantize to fourth roots of unity {±1, ±i}, and apply recursive residual error quantization for near full-precision performance.
Skill Summary
Fairy2i introduces extreme quantization through complex-valued arithmetic, enabling conversion of pre-trained real-valued LLMs to 1-2bit precision without retraining. The method proves that any real-valued linear layer can be expressed as a complex-valued operation, then quantizes weights to the fourth roots of unity {±1, ±i} using phase-based projection. Recursive residual error quantization further reduces approximation errors, achieving near full-precision performance on LLaMA-2 7B at effective 2-bit precision.
When To Use
- Compressing large pre-trained LLMs for deployment on memory-constrained hardware
- Projects requiring extreme (1-2bit) quantization from existing full-precision checkpoints
- Scenarios where retraining is computationally infeasible but compression is necessary
- Research exploring complex-valued neural networks and their compression properties
When NOT To Use
- Scenarios requiring maximum inference speed on standard hardware (complex arithmetic adds overhead)
- Projects where 1-2bit quantization causes unacceptable performance degradation for your domain
- Domains with very limited domain data making quantization-aware fine-tuning impractical
- Hardware platforms not optimized for complex-valued operations
Core Technique
Three key technical innovations enable extreme compression:
1. Widely-Linear Transformation Prove that any real-valued linear layer can be mathematically expressed as an equivalent complex-valued operation combining "a complex-linear part and a conjugate-linear part." This lossless reparameterization converts pre-trained real models into the complex domain without altering behavior before quantization.
2. Phase-Aware Complex Quantization Rather than using real-valued binary or ternary sets, quantize complex weights to the "fourth roots of unity ({±1,±i})" using phase-based projection. This utilizes the full 2-bit encoding space more efficiently than real alternatives.
3. Recursive Residual Error Quantization Iteratively quantize the remaining error after each stage using the same codebook. At deployment, the final weight is a sum of multiple ultra-low-bit terms, reducing approximation errors with minimal overhead.
Implementation Notes
Start with pre-trained real-valued checkpoints. Apply widely-linear transformation to convert layers to complex domain. Implement phase-aware quantization to fourth roots of unity. Apply recursive residual error quantization for multi-term weight representation. Validate performance on downstream tasks—method typically preserves functionality within 1-3% accuracy loss.
References
- Original paper: Fairy2i (Dec 2025)
- Complex-valued neural network literature
- Quantization-aware training for LLMs
Related Skills
ADu2021/flow-map-trajectory-tilting
testing
VerifiedTrustedCommunity
Uses flow maps as look-ahead operators to enable principled reward-guided diffusion by predicting trajectory endpoints at any denoising step. Deploy when applying rewards or preferences to diffusion trajectories with meaningful gradients throughout generation.
2SKILL.mdUpdated Apr 17, 2026
ADu2021/flexible-data-mixture-of-experts
testing
VerifiedTrustedCommunity
Train language models where each expert learns independently on closed datasets, enabling flexible inference with selective data inclusion or exclusion. 41% performance improvement while allowing users to opt out of specific data sources without retraining.
2SKILL.mdUpdated Apr 17, 2026
ADu2021/flexibility-trap-diffusion-reasoning
data-ai
VerifiedTrustedCommunity
Understand how token generation flexibility in diffusion LMs paradoxically constrains reasoning, as models exploit ordering flexibility to avoid uncertain tokens, and apply simplified approaches that preserve parallel decoding benefits. Use when optimizing diffusion-based language models for reasoning tasks.
2SKILL.mdUpdated Apr 17, 2026
ADu2021/flex-continuous-agent-evolution
devops
VerifiedTrustedCommunity
Enable LLM agents to improve continuously during deployment by constructing structured experience libraries through self-reflection on successes and failures—achieving 23% improvement on reasoning without gradient-based parameter updates or external training.
2SKILL.mdUpdated Apr 17, 2026