TT-Lang Documentation

User Guide

• Introduction
• Getting Started
  • Install from PyPI
  • Build from source for the simulator only
  • Docker quick start
  • Building without Docker
  • Functional simulator
  • Quick checks
  • Next steps
• Getting started with compiler-backed emulation
  • Host prerequisites
  • Obtain TT-Lang
  • Install the environment
  • Run a program
  • Run tests with the existing test framework
  • Validate and inspect the environment
  • Known limitations
  • Troubleshooting
• Simulation backends
  • Setup
  • Running
  • Per-node body execution
  • Blocks released without being pushed or popped
  • Unified bodies and the transfer handle
  • Float32 Promotion
  • Simulator statistics (tt-lang-sim-stats)
  • Debugging
  • Python profiling (tt-lang-sim-profile)
  • Dry-run mode
• Tour of TT-Lang
  • Overview
  • Key Concepts
  • Operation Basics
  • Dataflow Buffers
  • Reference
• Elementwise Tutorial
  • The Goal
  • Step 0 — TT-NN Baseline
  • Step 1 — Single Node, Single-Tile Block
  • Step 2 — Single Node, Multi-Tile Block
  • Step 3 — Multi-Node, Fixed Grid
  • Step 4 — Multi-Node, Full Grid
• Matmul Tutorial
  • The Goal
  • Step 0 — TT-NN Baseline
  • Step 1 — Single Node, Single-Tile Block
  • Step 2 — Single Node, Multi-Tile Block
  • Step 3 — Multi-Node, Fixed Grid
  • Step 4 — Multi-Node, Full Grid
  • Step 5 — Multi-Device, Shard M
  • Step 6 — Multi-Device, Shard K
  • Step 7 — Multi-Device, Shard K with All-Reduce
• Programming Guide
  • Compiler Options
  • External Functions
  • Operation Runtime Resources
  • Print Debugging
  • Performance Tools
  • Simulator
  • Examples
• Claude Skills
  • Example Workflow
  • Available Commands
• Testing
  • Compiler tests
  • Simulator tests
  • Python tests
  • Build hygiene
• Tools

Reference

• TT-Lang Spec
  • Specification Versions
  • Introduction
  • Operation function
  • Grid
  • Dataflow buffer
  • Block
  • Block states
  • Pipe
  • Tensor slice
  • Copy
  • Semaphore
  • Performance and debugging
  • Appendix A. Glossary
  • Appendix B. Block operators and math functions
  • Appendix C. Naming guidelines
  • Appendix D. Functionality matrix
  • Appendix E. Platform limitations
• Compiler Options
  • Code Generation Options
  • Compute Configuration
  • Environment Variables
  • Other Decorator Parameters
  • ttlang-opt Pass Reference
• External Functions
  • Scalar results
  • Logical-kernel selection
  • Dispatch-stable condition results
  • Template arguments
  • Function arguments
  • DFB dependencies, protocol effects, and interface access
  • Include directories
  • DFB synchronization ownership
• Operation Runtime Resources
  • Typed Records
  • External Fabric Managers
  • Validation and Specialization
  • Program Cache and Lifetimes
  • Emitted Runners and Simulator
• TT-Lang Debug Print Spec
  • Overview
  • Python API
  • In depth + code gen
  • Example: instrumented compute kernel
• Performance Tools
  • Using Claude for Performance Analysis
  • Environment Variables
  • Perf Summary
  • Auto-Profiling
  • User-Defined Signposts
  • Perfetto Trace Server

Contributor Guide

• Build System
  • Overview
  • Prerequisites
  • Build Modes
  • Installing
  • Building Documentation
  • Submodules
  • Uplifting Submodules
  • CMake Options
  • Build Architecture
  • Troubleshooting
• macOS: Building and Simulator-Based Runtime Testing
  • Three non-silicon execution paths
  • macOS host toolchain build (LLVM/MLIR only)
  • The hardware simulator (libttsim.so)
  • Runtime tests in a Lima VM
  • Reference: how tt-mlir CI uses the same simulator
  • Status
• Guidelines
• Contributor Guide
  • Workflow
  • Validation
  • Documentation
  • Adding Elementwise Operations
• Implementation Notes
  • Design Documents
  • Lowering Pipeline Traces