How tenstorrent/tt-metal works
TT-Metal is Tenstorrent’s software stack: TT-NN provides Python/C++ neural-network operations, TT-Metalium exposes low-level hardware programming, and model implementations compose these for inference and generation. A typical user supplies inputs to a model pipeline, which runs neural-network stages on device and returns generated output; other documented capabilities include model demos, training, and multi-device operation. The graph reflects README claims and the supplied source index/excerpts; unsampled subsystem links are omitted where their callers are not evidenced.
Main entry points: dispatch_query_manager.cpp Dispatch Queries · dispatch_engine_cores.cpp Dispatch Core Selection · pipeline_minimax_h3.py MiniMax-H3 Pipeline · pipeline_ltx.py LTX Pipeline
Components 34
Model Pipelines
- MiniMax-H3 Pipeline
models/tt_dit/pipelines/minimax_h3/pipeline_minimax_h3.py - LTX Pipeline
models/tt_dit/pipelines/ltx/pipeline_ltx.py - Hunyuan Pipeline
models/experimental/hunyuan_image_3_0/ttnn/pipeline.py - CosyVoice2 Pipeline
models/experimental/cosyvoice2/tt/pipeline.py - Ideogram Pipeline
models/tt_dit/pipelines/ideogram4/pipeline_ideogram4.py
Diffusion Components
- DiT Parallel Config
models/tt_dit/parallel/config.py - DiT Parallel Manager
models/tt_dit/parallel/manager.py - Tensor Utilities
models/tt_dit/utils/tensor.py - Tracing Utilities
models/tt_dit/utils/tracing.py - Qwen3-VL Encoder
models/tt_dit/encoders/qwen3vl/model_qwen3vl.py - Audio Operations
models/tt_dit/layers/audio_ops.py - LTX VAE
models/tt_dit/models/vae/vae_ltx.py - MiniMax VAE
models/tt_dit/models/vae/minimax_h3/vae_minimax_h3.py
TT-NN Runtime
- TT-NN API
ttnn/ttnn/__init__.py - Tensor Storage
ttnn/api/ttnn/tensor/storage.hpp - Storage Implementation
ttnn/core/tensor/storage.cpp - Pipeline Binding
ttnn/cpp/ttnn-nanobind/pipeline_module_nanobind.cpp
Metalium Runtime
- Metal Context
tt_metal/impl/context/metal_context.hpp - Dispatch Queries
tt_metal/impl/dispatch/dispatch_query_manager.cpp - Dispatch Core Selection
tt_metal/llrt/dispatch_engine_cores.cpp - Dispatch Query Interface
tt_metal/impl/dispatch/dispatch_query_manager.hpp - Core Coordinates
tt_metal/common/core_coord.hpp - Dispatch Core Config
tt_metal/impl/dispatch/dispatch_core_common.hpp - Dispatch Core Manager
tt_metal/impl/dispatch/dispatch_core_manager.hpp - Core Descriptor
tt_metal/llrt/core_descriptor.hpp - Hardware Abstraction
tt_metal/llrt/hal.hpp - Runtime Options
tt_metal/llrt/rtoptions.hpp - Metal Environment
tt_metal/impl/context/metal_env_impl.hpp - Blitz Pipeline API
tt_metal/api/internal/blitz_decode_pipeline.hpp - Blitz Pipeline
tt_metal/distributed/experimental/blitz_decode_pipeline.cpp - Fabric Pipeline Builder
tt_metal/fabric/pipeline_builder.cpp
Distributed Training
- Pipeline-Parallel Llama
tt-train/sources/ttml/models/distributed/pipeline_parallel_llama.cpp - Pipeline Communication Ops
tt-train/sources/ttml/ops/distributed/pipeline_parallel_comm_ops.cpp
Other components
- Model User
Connections 39
- Model User → MiniMax-H3 Pipeline: requests video/audio
- Model User → LTX Pipeline: requests generation
- Model User → Hunyuan Pipeline: requests denoising
- Model User → CosyVoice2 Pipeline: requests speech
- Model User → Ideogram Pipeline: requests generation
- MiniMax-H3 Pipeline → Qwen3-VL Encoder: uses encoder
models/tt_dit/pipelines/minimax_h3/pipeline_minimax_h3.py - MiniMax-H3 Pipeline → Audio Operations: uses audio ops
models/tt_dit/pipelines/minimax_h3/pipeline_minimax_h3.py - MiniMax-H3 Pipeline → MiniMax VAE: uses VAE
models/tt_dit/pipelines/minimax_h3/pipeline_minimax_h3.py - LTX Pipeline → LTX VAE: uses VAE
models/tt_dit/pipelines/ltx/pipeline_ltx.py - LTX Pipeline → DiT Parallel Config: uses parallel config
models/tt_dit/pipelines/ltx/pipeline_ltx.py - LTX Pipeline → DiT Parallel Manager: uses manager
models/tt_dit/pipelines/ltx/pipeline_ltx.py - Ideogram Pipeline → Qwen3-VL Encoder: uses encoder
models/tt_dit/pipelines/ideogram4/pipeline_ideogram4.py - Ideogram Pipeline → DiT Parallel Config: uses parallel config
models/tt_dit/pipelines/ideogram4/pipeline_ideogram4.py - Ideogram Pipeline → DiT Parallel Manager: uses manager
models/tt_dit/pipelines/ideogram4/pipeline_ideogram4.py - Hunyuan Pipeline → TT-NN API: creates device tensors
models/experimental/hunyuan_image_3_0/ttnn/pipeline.py - CosyVoice2 Pipeline → TT-NN API: uses tensor API
models/experimental/cosyvoice2/tt/pipeline.py - LTX Pipeline → TT-NN API: uses tensor API
models/tt_dit/pipelines/ltx/pipeline_ltx.py - MiniMax-H3 Pipeline → TT-NN API: uses tensor API
models/tt_dit/pipelines/minimax_h3/pipeline_minimax_h3.py - Ideogram Pipeline → Tensor Utilities: uses tensor helpers
models/tt_dit/pipelines/ideogram4/pipeline_ideogram4.py - Hunyuan Pipeline → Tensor Utilities: uses tensor helpers
- TT-NN API → Tensor Storage: exposes tensor runtime
- Storage Implementation → Tensor Storage: implements storage
ttnn/core/tensor/storage.cpp - Pipeline Binding → Blitz Pipeline API: binds pipeline API
ttnn/cpp/ttnn-nanobind/pipeline_module_nanobind.cpp - Blitz Pipeline → Blitz Pipeline API: implements API
tt_metal/distributed/experimental/blitz_decode_pipeline.cpp - Dispatch Queries → Dispatch Query Interface: implements interface
tt_metal/impl/dispatch/dispatch_query_manager.cpp - Dispatch Queries → Dispatch Core Manager: queries placement
tt_metal/impl/dispatch/dispatch_query_manager.cpp - Dispatch Queries → Dispatch Core Selection: queries core layout
tt_metal/impl/dispatch/dispatch_query_manager.cpp - Dispatch Queries → Metal Context: reads environment
tt_metal/impl/dispatch/dispatch_query_manager.cpp - Dispatch Queries → Dispatch Core Config: uses dispatch config
tt_metal/impl/dispatch/dispatch_query_manager.cpp - Dispatch Queries → Core Descriptor: uses core data
tt_metal/impl/dispatch/dispatch_query_manager.cpp - Dispatch Queries → Hardware Abstraction: queries hardware
tt_metal/impl/dispatch/dispatch_query_manager.cpp - Dispatch Core Selection → Core Coordinates: uses coordinates
tt_metal/llrt/dispatch_engine_cores.cpp - Dispatch Core Selection → Core Descriptor: selects engine cores
tt_metal/llrt/dispatch_engine_cores.cpp - Dispatch Core Selection → Metal Environment: reads runtime state
tt_metal/llrt/dispatch_engine_cores.cpp - Dispatch Core Selection → Metal Context: uses context
tt_metal/llrt/dispatch_engine_cores.cpp - Dispatch Core Selection → Dispatch Core Config: uses config
tt_metal/llrt/dispatch_engine_cores.cpp - Dispatch Core Selection → Runtime Options: checks options
tt_metal/llrt/dispatch_engine_cores.cpp - Fabric Pipeline Builder → Metal Context: uses context
tt_metal/fabric/pipeline_builder.cpp - Pipeline-Parallel Llama → Pipeline Communication Ops: uses communication ops
tt-train/sources/ttml/models/distributed/pipeline_parallel_llama.cpp