Skip to content

Concepts

[[toc]]

Understand how Fuse models tensor programs, how execution policies work, and where each backend fits.

Program, IR, and evaluation

Fuse parses .fuse files into an intermediate representation (IR) consisting of equations, sources, sinks, and exports. Programs compile down to backend-specific runners.

  • Parser — Converts the DSL into a structured IR, validating syntax and index semantics.
  • IR — Captures tensor expressions, projections, and metadata needed for execution.
  • Evaluators — Execute the IR using NumPy (always available) or optional Torch/JAX integrations.

Refer to fuse.core.ir for IR data structures.

Execution configuration

ExecutionConfig objects let you control precision, device selection, demand/fixpoint modes, and Monte Carlo projection strategies.

Configure execution

from pathlib import Path

from fuse import ExecutionConfig, Program

program = Program(Path("examples/01_attention_block.fuse").read_text())
config = ExecutionConfig(device="cpu", demand_mode=True)
runner = program.compile(backend="numpy", config=config)

Policies supplement configuration:

  • RuntimePolicies — Compose caching behaviour, manifest weight stores, and quantisation options.
  • Temperature schedules — Drive annealing-style loops with make_schedule helpers.
  • Explainability hooks — Program.explain() emits structured traces per iteration.

Backends at a glance

All backends share the parser, IR, caching, and policy layers. Key differences lie in execution engines:

  • NumPy evaluator — Reference implementation with complete DSL coverage.
  • Torch FX backend — Lowers to torch.fx.GraphModule objects and integrates with torch.compile.
  • JAX backend — Emits pure JAX functions with optional XLA compilation.

See the backend matrix for feature-by-feature coverage.

DSL building blocks

The DSL is designed to resemble algebraic equations while supporting tensor-specific operations:

  • Equations with projections (=, +=, max=, avg=) define computation graphs.
  • Sources map filenames to tensors; sinks export results.
  • Boolean tensors use parentheses (Fact(i, j)), and dotted indices indicate reduction axes.
  • Builtins cover activation functions, attention, concatenation, and reduction helpers.

Consult the DSL reference for the full grammar and examples.

Migration to structured syntax (v2)

The legacy parser remains the default to keep existing programs working. You can opt into a structured grammar that adds expressions and blocks by constructing Program with parser='v2'. See the migration guide for an incremental, no‑rewrite path.