Metadata-Version: 2.4
Name: coolingcube-tp
Version: 0.1.0
Summary: Tensor Parallel feasibility analyser — should you use TP, how many GPUs, does DFlash break scaling?
Author-email: Devon Langford <CoolingCubeInfo@proton.me>
Project-URL: Homepage, https://coolingcube.cc
Keywords: tensor-parallel,inference,llm,gpu,exllama,vllm
Classifier: Programming Language :: Python :: 3
Classifier: License :: OSI Approved :: MIT License
Requires-Python: >=3.8
Description-Content-Type: text/markdown

# coolingcube-tp

Tensor Parallel feasibility analyser. Answers the question every multi-GPU inference user has:

> Should I use TP? How many GPUs? Does DFlash break scaling?

## Install

```bash
pip install coolingcube-tp
```

## Usage

```bash
# RTX5090 + RTX PRO 6000, 32B model, DFlash enabled
coolingcube-tp --gpus rtx5090 rtx_pro_6000 --model 32 --dflash 8

# 4x RTX4090, 70B model, no DFlash
coolingcube-tp --gpus rtx4090 rtx4090 rtx4090 rtx4090 --model 70

# List all supported GPUs
coolingcube-tp --list-gpus
```

## Python API

```python
from coolingcube_tp import analyse_tp

result = analyse_tp(
    gpus=["rtx5090", "rtx_pro_6000"],
    model_params_b=32,
    dflash_draft=8,
    interconnect="pcie"
)

print(result["verdict"])
print(result["recommendation"])
```

## How it works

NFF-IMC (Non-Feasible Finite — Inverse Mass Control) applied to inference TP:

- Computes per-GPU compute time from real TFLOPS and bandwidth specs
- Models AllReduce communication overhead per layer
- Applies DFlash draft amplification when enabled
- Returns speedup vs single GPU and a clear verdict

Free. No API calls. Runs locally.

coolingcube.cc
