sutra-compiler

The compiler for the Sutra programming language — a purely functional language whose primitives are tensor operations on frozen-LLM embedding vectors.

Install

pip install sutra-dev                     # validator + codegen only
pip install "sutra-dev[runtime]"          # adds torch so --emit / --run can execute the generated module
pip install "sutra-dev[runtime,embed]"    # adds the embedding client for LLM-substrate codebooks

Requires Python 3.11+.

Use

After install, the sutrac command is on your $PATH.

# Validate a .su source file.
sutrac path/to/program.su

# Validate every .su file under a directory.
sutrac path/to/directory

# Compile to self-contained PyTorch Python and print to stdout.
sutrac --emit path/to/program.su

# Compile and execute in one step (requires the [runtime] extra installed).
sutrac --run path/to/program.su

# Step-by-step compilation review (source → AST → inline → simplify → emit).
sutrac --review path/to/program.su

# Cross-file class-name casing check.
sutrac --consistency path/to/directory

sutrac --help lists every flag.

What is Sutra

A .su source file looks like TypeScript — functions, classes, && / ||, string and numeric literals — but every value is a vector in a frozen LLM embedding space, and every operation is a tensor op on those vectors. The compiler emits self-contained Python that calls a small runtime; chains of bind / unbind / bundle / similarity reduce to chains of matrix multiplies that the simplifier folds into cached matrices at compile time.

The result is a program where there are no Python branches, no host-side if / while on data values, and no JIT — just one straight-line tensor expression.

For a full introduction, see sutra.topazcomputing.com. For the language specification and design rationale, see the Sutra repository.

Status

Research-grade. Versions before 1.0 may break source compatibility. The grammar is stable; the codegen and the standard library still move.