logitly 0.2.2

Turn compatible causal LLMs into fast, bounded decision engines.

Logitly turns state, a question, and named choices into a probability distribution. It performs one prefill forward pass, reads the model's original next-token logits for fixed labels, and normalizes only the supplied choices.

pip install "logitly[transformers]"

Quick Start

from logitly import DecisionModel

with DecisionModel.from_pretrained("lfm") as model:
    result = model.choice(
        state={"amount": 9700, "device": "unknown"},
        question="What action should be taken?",
        choices={
            "close": "Close as benign",
            "review": "Request analyst review",
            "block": "Block immediately",
        },
    )

print(result.choice)
print(result.confidence)
print(result.probabilities)

No answer tokens are generated. Logitly does not fine-tune the model, replace its LM head, or run a completion loop.

Decision Primitives

model.noul(state, question)                  # yes/no probabilities
model.choice(state, question, choices)       # named choice distribution
model.score(state, question, levels)         # ordinal distribution + mean

model.decide_many(requests, batch_size=...) evaluates mixed decisions in a batch while preserving the same result schemas.

How It Works

state + question + choices
            ↓
      frozen causal LLM
            ↓
   original LM-head logits
            ↓
 select A/B/C/… label logits
            ↓
          softmax
            ↓
 named probability distribution

Logitly gathers the final-position vocabulary logit for each valid label and computes a softmax over only that restricted set. Every model profile verifies that its labels are distinct, single-token continuations at the exact native assistant-answer boundary before inference begins.

Models and Runtimes

Compatible Hugging Face model IDs and local checkpoints are also accepted. The Python package provides Transformers, vLLM, and llama.cpp runtimes behind the same API.

pip install "logitly[transformers,quantized]"
pip install "logitly[vllm]"
pip install "logitly[llama-cpp]"
pip install "logitly[browser]"

Transformers and vLLM should use separate environments because their pinned runtime dependencies differ. Python inference currently requires an NVIDIA CUDA GPU.

CLI

logitly validate lfm
logitly playground lfm --port 8000
logitly benchmark lfm --size 128 --output results/lfm

Documentation

• Source and full guide
• Library and runtime guide
• Apple runtime guide

Scope

confidence is the maximum probability in the restricted distribution, not an automatic correctness guarantee. Validate outcomes and calibration on data representative of your application.

License

Logitly is available under the MIT License.