layer-scan 0.2.2

Automated LLM layer duplication configuration scanner with heatmap visualization

layer-scan

Automated LLM layer duplication config scanner — find the optimal (i,j) for any model + task.

Given an open-source LLM and an evaluation probe, layer-scan finds the layer duplication config that maximizes model capability without modifying weights.

Features

• Full (i,j) scanning — automated search across all valid configs
• Logit distribution scoring — deterministic, no text generation needed
• Multi-probe analysis — Pareto-optimal configs across tasks
• Sparse-then-dense — two-phase scanning for large models
• mergekit export — one-click YAML for mergekit passthrough
• Cross-tool annotation — overlay neuro-scan labels
• Interactive heatmaps — Plotly with hover details

Install

pipx install layer-scan
# or
pip install layer-scan

# ExLlamaV2 for 70B+ models:
pip install layer-scan[exllamav2]

Quick Start

# Scan with math probe
layer-scan scan --model Qwen/Qwen2-7B --probe math

# Scan + export mergekit config
layer-scan scan --model Qwen/Qwen2-7B --probe math --export-mergekit config.yaml

# Multi-probe Pareto analysis
layer-scan multi-probe --model Qwen/Qwen2-7B --probes "math,eq,json"

# Annotate with neuro-scan labels
layer-scan annotate --results results.json --neuro-report report.json

# ExLlamaV2 for large models
layer-scan scan --model /models/qwen2-72b-exl2 --backend exllamav2 --gpu-split "22000,22000"

# Then merge
mergekit-yaml config.yaml ./merged-model --copy-tokenizer

Commands

Command	Description
scan	Scan (i,j) configs with a single probe
multi-probe	Cross-probe scan, Pareto-optimal configs
annotate	Overlay neuro-scan labels on heatmap
lookup	Fetch pre-computed results from HF Hub
probes	List available probes

Key Options

Option	Default	Description
--model, -m	required	Model path or HF ID
--probe, -p	math	Probe: math, eq, json, custom
--backend, -b	transformers	transformers or exllamav2
--min-block	7	Min duplicated block size
--top-k, -k	5	Top configs to report
--sparse-first	off	Sparse scan then refine
--export-mergekit	—	Export as mergekit YAML

Probes

Probe	Samples	Tests
math	16	Arithmetic, geometry, calculus
eq	12	Emotions, social cues
json	10	JSON extraction, schema
custom	variable	Load from JSON file

Backends

Backend	Best for	Multi-GPU
transformers	Small-medium models	—
exllamav2	70B+ quantized	--gpu-split

References

• Repeat Yourself: Layer Duplication
• SOLAR 10.7B: Depth Up-Scaling
• MergeKit

License

MIT