rift-audit 0.1.1

Reduced Inference Fast Transformer (RIFT) audit tool — measures inference efficiency in Transistor Flip Equivalents

RIFT Audit Tool

Reduced Inference Fast Transformer audit tool — measures inference efficiency in Transistor Flip Equivalents (TFE) across CMOS, Photonic, and Quantum substrates.

Part of the Blundin Space RIFT Competition by QuanTM.ai.

Install

pip install rift-audit

# With HuggingFace model support (required for most models):
pip install rift-audit[transformers]

From source

git clone https://github.com/justinm3434/rift-audit.git
cd rift-audit
pip install -e ".[transformers,dev]"

Quick Start

CLI

# Audit GPT-2 from HuggingFace
rift-audit gpt2 --hellaswag-acc 30.0 --val-loss 3.28

# Save signed JSON report (required for submission)
rift-audit gpt2 --hellaswag-acc 30.0 --val-loss 3.28 --output report.json

# Custom sequence length
rift-audit gpt2 --hellaswag-acc 30.0 --val-loss 3.28 --seq-len 256

The tool will output something like:

Computing model weight hash...
Model hash: a3f8c1d2e4b5...
Auditing gpt2 (seq_len=128, device=cpu)...
Report saved to report.json (signed)

Python API

import torch
from transformers import AutoModelForCausalLM
from rift_audit import RIFTAuditor, compute_rift_scores

# Load model
model = AutoModelForCausalLM.from_pretrained("gpt2")
model.eval()

# Create dummy input
input_ids = torch.randint(0, 50257, (1, 128))

# Audit
auditor = RIFTAuditor()
result = auditor.audit(model, input_ids, model_name="gpt2")

# Score
scores = compute_rift_scores(result, hellaswag_accuracy=30.0, validation_loss=3.28)

print(f"RIFT-Silicon:   {scores.silicon:.4f}")
print(f"RIFT-Photonic:  {scores.photonic:.4f}")
print(f"RIFT-Quantum:   {scores.quantum:.4f}")
print(f"RIFT-Universal: {scores.universal:.4f}")

Scoring

RIFT Score = (HellaSwag_accuracy × 1000) / log10(TFE_per_token)

Higher is better. The numerator rewards intelligence. The denominator penalizes thermodynamic cost on a log scale.

Leaderboard Categories

Category	TFE Table	Description
RIFT-Silicon	v1.0 (CMOS)	Current hardware. Win this today.
RIFT-Photonic	v2.0 (Photonic)	Design for the substrate that's coming.
RIFT-Quantum	v3.0 (Quantum)	The long game. Radical architectures welcome.
RIFT-Universal	Geometric mean	The overall champion.

Report Integrity

Every report generated by rift-audit is cryptographically signed using HMAC-SHA256 keyed with a SHA-256 hash of the model's weights. This prevents score tampering — any modification to the JSON invalidates the signature.

Verify a report (signature only)

rift-verify report.json

Full verification (signature + model weights)

rift-verify report.json --model gpt2

Example output:

============================================================
  RIFT REPORT VERIFICATION — Full (signature + model hash)
============================================================

  RESULT: PASS
  Signature valid:    True
  Model hash matches: True
  Reason: Signature valid and model hash matches
============================================================

Python API

from rift_audit import verify_report, compute_model_hash

# Signature-only verification
result = verify_report(report)
print(result["valid"])        # True/False
print(result["reason"])

# Full verification (requires model)
from rift_audit.integrity import verify_model_against_report
result = verify_model_against_report(model, report)
print(result["model_hash_matches"])

TFE Conversion Table

The TFE table maps operations to equivalent transistor state transitions — a substrate-agnostic measure of computational thermodynamic cost. Values are calibrated against published energy data (Horowitz 2014 ISSCC, Stillmaker & Baas 2017).

See the full table and methodology at quantm.ai/competition.

Qualifying Requirements

To appear on the leaderboard, submissions must meet:

Metric	Threshold
HellaSwag accuracy	≥ 30.0%
OpenWebText validation loss	≤ 3.28

Submission

Submit to [e-mail] with subject [RIFT Submission] Your Model Name. See quantm.ai/submit for the full guide.

License

MIT