pip install py-infinigram
Copy

git clone https://github.com/queelius/infinigram.git
cd infinigram
pip install -e .
Copy

from infinigram import Infinigram

# Create model from text (byte-level, automatic UTF-8 encoding)
model = Infinigram(b"the cat sat on the mat the cat ran")

# Predict next byte probabilities
probs = model.predict(b"the cat", top_k=5)
# {32: 1.0}  # space byte (32) has highest probability

# Find longest matching suffix
position, length = model.longest_suffix(b"the cat")
print(f"Matched {length} bytes at position {position}")

# String input works too (auto-encoded to UTF-8)
probs = model.predict("the cat", top_k=5)
Copy

# Build and save a model
model = Infinigram.build("The quick brown fox jumps over the lazy dog.", "my_model")

# Load it later (instant — uses mmap, not RAM)
model = Infinigram.load("my_model")
probs = model.predict("The quick")
Copy

infinigram-repl

# Quick tour:
infinigram> ds demo              # Create dataset
infinigram> add the cat sat      # Add document
infinigram> predict the cat      # Get predictions
infinigram> complete the --max 10  # Generate text
infinigram> config               # Show settings
infinigram> help                 # All commands
Copy

infinigram-serve --port 8000
Copy

model = Infinigram(b"abcdef abcdef abcxyz")

pos, length = model.longest_suffix(b"abcd")   # length=4
pos, length = model.longest_suffix(b"abc")    # length=3
pos, length = model.longest_suffix(b"zzz")    # length=0 (no match)
Copy

# Standard: longest suffix match
probs = model.predict(b"the cat", top_k=10)

# Weighted: combine predictions from multiple suffix lengths
probs = model.predict_weighted(b"the cat", weight_fn=lambda k: k**2)

# Backoff: Stupid Backoff smoothing (Brants et al., 2007)
probs = model.predict_backoff(b"the cat", backoff_factor=0.4)
Copy

from infinigram import Infinigram, TokenProbabilityAdapter
import tiktoken

model = Infinigram(b"The quick brown fox jumps over the lazy dog.")
tokenizer = tiktoken.encoding_for_model("gpt-4")
adapter = TokenProbabilityAdapter(model, tokenizer)

# Compute token probability via byte-level chain rule:
# P(token|ctx) = product of P(byte_i | ctx, byte_1, ..., byte_{i-1})
prob = adapter.token_probability("The quick brown", tokenizer.encode(" fox")[0])

# Mix with LLM probabilities: P_final = alpha * P_LLM + (1-alpha) * P_corpus
llm_probs = {token_id: 0.3}
mixed = adapter.mix_probabilities("The quick brown", llm_probs, alpha=0.7)
Copy

from infinigram import Infinigram
from infinigram.corpus_utils import build_corpus_with_augmentation

docs = ["Hello World", "Goodbye World"]
corpus = build_corpus_with_augmentation(docs, augmentations=[str.lower, str.upper])
model = Infinigram(corpus)
Copy

model = Infinigram(b"hello")
model.update(b" world")  # Appends and rebuilds suffix array
print(model.n)  # 11
Copy

Infinigram(
    corpus_or_path,           # bytes, List[int], str, or Path
    max_length: int = None,   # Max suffix length (None = unlimited)
    min_count: int = 1,       # Min frequency threshold
)
Copy

CLI & REPL              infinigram-repl, infinigram-serve
REST API                FastAPI, OpenAI-compatible /v1/completions
Python API              Infinigram class (predict, search, continuations)
Suffix Array Engine     MmapSuffixArray, ChunkedMmapSuffixArray (pydivsufsort)
Copy

pytest tests/                                    # All tests
pytest tests/ -v                                 # Verbose
pytest tests/ --cov=infinigram --cov-report=html # Coverage
pytest tests/test_infinigram.py -k "cache"       # Keyword filter
Copy