Tenstorrent software stack overview (TT-Forge, TT-NN, TT-Metalium, TT-Lang)

Vendor's own statement of the layer boundaries, which is what a deployment planner needs: TT-Forge 'is Tenstorrent's Multi-Level Intermediate Representation (MLIR)-based compiler. It bridges high-level machine learning frameworks with the Tenstorrent software stack', compiling models from PyTorch, JAX and TensorFlow and lowering to TT-Metalium; TT-NN is the pre-packaged optimized operator library on top; TT-Metalium is the low-level kernel programming model. TT-Forge's own framing is 'an automated, general path to run many types of model architectures without requiring custom kernel development' - vendor-general, not tuned per model. Note the existing engines/niche-tt-metal record covers TT-Metalium/TT-NN as a programming model; this source is what the new compiler record needs for the graph-compiler layer specifically.