Skip to content

install.md

Latest commit

 

History

History
55 lines (39 loc) · 1.92 KB

install.md

File metadata and controls

55 lines (39 loc) · 1.92 KB
Raw

Installation details

Source tt-metal

install_dependencies.sh installs the host packages and SFPI compiler required by build_metal.sh. Run it once before the first source build. Keep TT_METAL_HOME exported at runtime and retain $TT_METAL_HOME/build_Release, where JIT-compiled kernels are loaded.

If a kernel compile fails with riscv-tt-elf-g++: error: unrecognized option '-ftt-...', update SFPI and clear the old kernel cache before rebuilding:

sudo ./install_dependencies.sh --sfpi
rm -rf ~/.cache/tt-metal-cache/

Some board and firmware combinations require an explicit mesh descriptor:

export TT_MESH_GRAPH_DESC_PATH=$TT_METAL_HOME/tt_metal/fabric/mesh_graph_descriptors/p150_mesh_graph_descriptor.textproto

Set it before constructing tt_atom.batch.MultiCard / MultiCardSim so worker processes inherit it.

Bundles and reference dependencies

Calculator(atoms) exports and caches the required model weights on first use. UMA caches one bundle per reduced composition, charge, spin, and task. Orb caches one bundle per checkpoint.

The exporters use a separate reference environment because fairchem requires numpy>=2 while the validated ttnn environment uses numpy<2:

python -m venv refenv
refenv/bin/pip install "fairchem-core>=2.10" "orb-models==0.5.5"

Pass refenv="/path/to/refenv/bin/python" to Calculator, or set TT_ATOM_REFENV. Cached runs do not need the reference environment.

orb-models==0.5.5 is required by the current exporter. Version 0.6 changed the pretrained checkpoint API and moved system_config.

For a manual export from a source checkout:

refenv/bin/python tools/export_weights.py \
  --uma-s-1 --xyz structure.xyz --task omol --out model.npz
refenv/bin/python tools/export_orb_weights.py \
  --ckpt conservative-inf-omat --out weights.npz

Load the resulting file with TTAtomCalculator("model.npz") or OrbCalculator("weights.npz").