SPECTER LABS

TT Backend

Tenstorrent execution for Flow Lenia. The user-facing entrypoint is LeniaCLI tt run; Lenia Studio remains the replay and inspection surface.

Modes

  • single: one simulation on one TT device.
  • fleet: independent simulations split across TT devices.
  • mesh: one simulation over a TTNN/TT-Lang mesh. Reintegration is row-sharded; the spectral front half is still replicated/gathered.

Mesh mode is correctness-enabled. It is not yet the default performance path.

Run

Remote runs require --remote-root or LENIA_TT_REMOTE_ROOT pointing at this dossier on the target host.

export LENIA_TT_REMOTE_ROOT="$(dispatch workspace plan --on quietbox --project specter-labs --json | jq -r .remote_cwd)"

LeniaCLI tt run \
  --host quietbox \
  --config configs/base/paper_base_2c_128.json \
  --output tmp/tt-runs/orbitum-128 \
  --execution-mode single \
  --tt-card-num 0 \
  --steps 300 \
  --frame-every 5

LeniaCLI tt run \
  --host quietbox \
  --config configs/base/paper_base_2c_128.json \
  --output tmp/tt-runs/fleet-128 \
  --execution-mode fleet \
  --device-list 0,1,2,3 \
  --tt-card-list 0,1,2,3 \
  --batch-size 4 \
  --steps 300 \
  --frame-every 10

Each run writes tt_run.json with backend mode, device selection, seeds, timing, and final mass summaries.

Device Rules

QuietBox N300 cards expose two Wormhole chips per card. Mesh shapes count chips, so four cards are --mesh-shape 1,8.

The TT-Lang dist container maps single mode to /dev/tenstorrent/0. fleet and mesh need explicit reservations:

dispatch run --on quietbox --device wormhole:0,1 -- ...
dispatch run --on quietbox --device wormhole:all -- ...

LeniaCLI --host and dispatch run are separate launchers. Use LeniaCLI tt run --host quietbox for Studio/export artifacts. Use dispatch run for backend profiling when dispatch owns the Wormhole reservation; in that mode the command is already running in the remote workspace.

Developer Tools

tt_backend/devtools/ is for bringup, profiling, and hardware debugging. It is not the user workflow.

python devtools/run.py --backend tt --execution-mode mesh --device-list 0,1,2,3 --mesh-shape 1,8 ...
python devtools/bench.py --backend tt --execution-mode single --grid-sizes 128,256 --batch-sizes 1 ...
python devtools/probe_mesh_halo.py --mesh-shape 1,2 --tt-visible-devices 0 --size 256 --planes 2 --assemble
python devtools/probe_mesh_dft.py --mesh-shape 1,4 --tt-visible-devices 0,1 --size 512 --planes 20 --mode partition-complex-dft

--mesh-dft enables the experimental mesh-partitioned DFT path for devtools/run.py and devtools/bench.py. It is correctness-validated but slower than the default front half.

Accepted performance facts and rejected probes are tracked in TT Backend Performance.