Metadata-Version: 2.4
Name: coolingcube-q
Version: 0.1.1
Summary: Quantization feasibility analyser — is your quant NFF for coding, math, reasoning or chat?
Author-email: Devon Langford <CoolingCubeInfo@proton.me>
Project-URL: Homepage, https://coolingcube.cc
Keywords: quantization,llm,exllama,kld,feasibility,nff
Classifier: Programming Language :: Python :: 3
Classifier: License :: OSI Approved :: MIT License
Requires-Python: >=3.8
Description-Content-Type: text/markdown

# coolingcube-q

Quantization feasibility analyser. Takes your KLD and top-1 agreement measurements and tells you whether your quant is safe to deploy for coding, math, reasoning or chat.

## Install

```bash
pip install coolingcube-q
```

## Usage

```python
from coolingcube_q import analyse_quant

# remichu_sm's real measurement: Step3.7 5.05bpw
result = analyse_quant(
    kld=0.0056,
    top1_agreement=0.9594,
    task="coding",
    bpw=5.05,
    model_params_b=27
)

print(result["verdict"])       # Feasible — safe to deploy
print(result["recommendation"])
```

## Get your measurements

From ExLlamaV3 eval scripts:

```bash
# KLD vs BF16 base
python eval/model_diff.py -ma ./quantized_model -mb ./bf16_model -r 5

# Perplexity across bitrates
python eval/compare_q.py --help
```

## Tasks

- `coding` — code generation, agentic coding (most demanding)
- `math` — mathematical reasoning
- `reasoning` — thinking models, chain of thought
- `general` — mixed use
- `chat` — conversation (most forgiving)

## Check all tasks at once

```bash
coolingcube-q --kld 0.0056 --top1 0.9594 --bpw 5.05 --model 27 --all-tasks
```

coolingcube.cc
