# Tensor Parallelism Communication Tax Audit

## Description
Calculates All-Reduce collective communication volume and latency overhead across NVLink and PCIe interconnects to prevent negative TP scaling.

## Category
`serving`

## Tags
`tensor-parallelism`, `all-reduce`, `nvlink`, `pcie`, `gpu-cluster`

## When to Use
Evaluate whether increasing TP degree (TP=2, 4, 8) delivers positive scaling or is throttled by interconnect communication overhead.

## Operational Limitations
Models standard Ring All-Reduce collective communication. Does not simulate overlap with custom CUDA kernels.

## Verification Script
Run the deterministic verification checker:
```bash
python3 scripts/check.py examples/synthetic.json
```
Exit codes:
- `0`: Check passed and thresholds satisfied.
- `1`: Threshold breach or operational violation detected.
- `2`: Schema error or invalid usage.
