fraqtl-runtime 0.1.1

fraQtl runtime — drop-in loader for fraQtl-compressed Hugging Face checkpoints. Production LLM inference with calibration-aware compression.

fraQtl

Runtime KV-cache and weight compression for production LLM inference.

Drop-in. No retraining. Calibration-aware.

---

What it is

fraqtl-runtime is the runtime loader for fraQtl-compressed model artifacts. It enables:

• Weight compression: load fraQtl-compressed Hugging Face checkpoints (e.g. fraQtl/Qwen3.6-35B-A3B-compressed) via standard transformers with trust_remote_code=True. The wheel ships the compiled loader that decodes the packed weights at load time.
• Runtime KV-cache compression (separate, in active validation): a llama.cpp-compatible runtime layer that compresses the V cache at runtime — independent of weight format.

---

Install

pip install fraqtl-runtime

That's the entire setup. No license token required for loading published artifacts.

---

Quick start

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

repo = "fraQtl/Qwen3.6-35B-A3B-compressed"
model = AutoModelForCausalLM.from_pretrained(
    repo, trust_remote_code=True,
    torch_dtype=torch.bfloat16, device_map="auto",
)
tok = AutoTokenizer.from_pretrained(repo)

ids = tok("The capital of France is", return_tensors="pt").to(model.device)
print(tok.decode(model.generate(**ids, max_new_tokens=20, do_sample=False)[0]))

trust_remote_code=True pulls a small stub from the model repo that imports the compiled loader from this wheel. You never write import fraqtl directly.

---

High-level approach

fraQtl combines two ideas:

1. Calibration-aware eigenbasis rotation — protect the input directions that matter for the deployment task; quantize the rest. The calibration corpus determines which directions are protected (this is FPT — fraQtl Pullback Theorem).
2. Per-row sign correction primitive — additional precision on top of low-bit quantization where it matters most for reasoning.

Both compose with standard quantization machinery (Lloyd-Max centroids, INT3 packing) and standard inference engines (HF transformers, llama.cpp).

---

Status

• Public weight-compression artifacts on Hugging Face: huggingface.co/fraQtl
• Runtime KV-cache compression layer: in active validation. Public benchmark numbers landing after H100 measurement lock and manual review.
• Methodology paper in preparation.

---

Links

• Site: fraqtl.ai
• Hugging Face: huggingface.co/fraQtl
• Diagnostic tool (open-source): fraqtl-diagnostic
• Contact: [e-mail]

---

License

Proprietary. The compressed model weights and loader are free to install and use for research and evaluation. Production / commercial use: contact fraQtl.