Explorer

Explorer is an interactive GUI tool from TT-MLIR for visualizing and experimenting with model graphs (including Tenstorrent's MLIR dialects), compiling and executing your model on Tenstorrent hardware.

What is Explorer?

Explorer is a visual debugging and performance analysis tool that allows you to:

• Visualize MLIR graphs: Inspect your model graph with hierarchical visualization
• Compile and execute your model: Compile your model to Tenstorrent hardware and execute it
• Debug performance: Identify bottlenecks and see affects of optimizations on runtime performance

Building with Explorer

Explorer is only available when building TT-XLA from source. It is not included in pre-built wheels. It is disabled by default in TT-XLA. You can enable it by building with the TTXLA_ENABLE_EXPLORER CMake option:

cmake -G Ninja -B build -DCMAKE_BUILD_TYPE=Release -DTTXLA_ENABLE_EXPLORER=ON
cmake --build build

Note: Enabling Explorer also enables Tracy performance tracing (TTMLIR_ENABLE_PERF_TRACE), which may slow down execution. For production deployments or performance benchmarking, consider building with -DTTXLA_ENABLE_EXPLORER=OFF.

Using Explorer

After building with Explorer enabled, launch the tool by running:

tt-explorer

This will start the interactive GUI for analyzing your model's compilation and execution.

Example graph to try out

module attributes {} {
  func.func @forward(%arg0: tensor<64x128xbf16>, %arg1: tensor<64x128xbf16>, %arg2: tensor<64x128xbf16>, %arg3: tensor<64x128xbf16>) -> tensor<64x128xbf16> {
    %0 = ttir.empty() : tensor<64x128xbf16>
    %1 = "ttir.add"(%arg0, %arg1, %0) : (tensor<64x128xbf16>, tensor<64x128xbf16>, tensor<64x128xbf16>) -> tensor<64x128xbf16>
    %2 = ttir.empty() : tensor<64x128xbf16>
    %3 = "ttir.add"(%arg2, %arg3, %2) : (tensor<64x128xbf16>, tensor<64x128xbf16>, tensor<64x128xbf16>) -> tensor<64x128xbf16>
    %4 = ttir.empty() : tensor<64x128xbf16>
    %5 = "ttir.add"(%1, %3, %4) : (tensor<64x128xbf16>, tensor<64x128xbf16>, tensor<64x128xbf16>) -> tensor<64x128xbf16>
    %6 = ttir.empty() : tensor<64x128xbf16>
    %7 = "ttir.relu"(%5, %6) : (tensor<64x128xbf16>, tensor<64x128xbf16>) -> tensor<64x128xbf16>
    return %7 : tensor<64x128xbf16>
  }
}

Learn More

For detailed documentation on how to use Explorer, including tutorials and advanced features, see the TT-MLIR Explorer Documentation.

Explorer is based on Google's Model Explorer with added support for Tenstorrent hardware compilation and execution.