Τομί (Tomii)

Website | Guide | Reference

Task-graph framework for streaming pipelines, MIMO workloads, and agent-tuneable applications.

NOTE Tomii is a research prototype under active development. Interfaces and performance characteristics may change between releases; it is best suited for research and experimentation today.

Tomii implements Tripartite Decoupling separating the three following artifacts: 1) Computation structure, 2) Kernel Implementation, 3) Runtime. Each artifact can be modified independently and Tomii is responsible for combining them at compile (transform kernels into tomii-suitable format) or execution time (parse computation structure description).

Tomii is not a general-purpose Taskflow or TBB replacement. For pure single-frame micro-task DAGs where dispatch overhead dominates, Taskflow is faster. Tomii's advantage is in workloads that benefit from concurrent frames, generational slot reuse, and structured graph surfaces — particularly MIMO-class packet-driven pipelines and agent-driven automated optimisation loops.

Each benchmark under bench/<name>/ is verifier-gated and reproducible from this repository — run the script in its directory; see bench/*/README.md for the topology, metric definition, and run instructions. Measured comparison numbers are published in the project documentation, not in these READMEs.

Python API (recommended)

import tomii as tm

app = tm.Graph()

buf_size    = app.var("buf_size", 100)
fft_planner = app.var("fft_planner", func="fft_planner", args=[buf_size])

gen_vec     = app.node("gen_vec",     func="generate_vector", factor=200, args=[buf_size])
compute_fft = app.node("compute_fft", func="compute_fft",     factor=200,
                       args=[fft_planner, gen_vec.out()])
vec_mat     = app.node("vec_mat",     func="vec_to_mat",      factor=200,
                       args=[gen_vec.out(), compute_fft.wait()])

app.build(func_path="plugin/src/lib.rs", plugin_manifest="plugin/Cargo.toml")
app.run(workers=4, slots=2, timing="timing.csv")

Install

# From PyPI — the distribution is named `tomii-rt`; the import name is `tomii`.
# 1.2.0 ships wheels for Linux (x86_64, arm64) and macOS (Intel, Apple
# Silicon), CPython 3.9–3.14 including free-threaded 3.13t/3.14t.
# (Windows wheels are planned for a later release.)
pip install tomii-rt          # then:  import tomii

# From source (for development)
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

The Rust crates are published on crates.io: tomii-core, tomii-types, tomii-macro, tomii-converter (cargo add tomii-core).

Graph API

Type system

tm.i32(-5) / tm.f32(3.14) / tm.String("hello") / tm.bool_(True)
tm.Complex64(1.0, -0.5) / tm.Vec("f32", [1.0, 2.0, 3.0])

Python int → usize, float → f64 by default.

Loops and conditions

loop_node = app.node("proc", func="process", factor=200,
                     loop=Loop("iter", factor=loop_factor))

cond_node = app.node("filter", func="filter_fn", factor=200,
                     condition=Condition(operation="Eq", value=1, value_type="usize",
                                        func="check_fn", args=[some_var]))

Build options

app.build(func_path="plugin/src/lib.rs",      # Rust source — auto-generates wrappers
          # or func_path="include/plugin.h",  # C header with // @tomii_export annotations
          plugin_manifest="plugin/Cargo.toml",
          release=True)

Run options

app.run(workers=8, system_threads=2, slots=4, max_frames=100,
        timing="timing.csv", report="report.json",
        inline_continuation=True, coalesce_barriers=True)

Agent-native interfaces

Tomii exposes structured discovery and diagnostic interfaces designed for LLM agents.

Discovery

python -m tomii --list-knobs           # all graph.run() options
python -m tomii --list-knobs-json      # machine-readable JSON with search hints
python -m tomii --schema               # JSON schema for the graph construction API

Structured performance report

Pass report="report.json" to app.run() for a JSON performance report after each run:

Agent Skills

SKILLS/ contains structured workflow skills covering the full optimisation lifecycle — from project discovery to graph coarsening:

./SKILLS/install-skills.sh            # installs to .claude/skills/ in CWD

Graph visualisation

python -m tomii --visualize examples/stream-analytics/graph.json       # browser view
python -m tomii --visualize examples/stream-analytics/graph.json --edit # browser edit
python -m tomii --visualize graph.json --ascii                          # terminal ASCII

The web UI renders a colour-coded DAG (Dagre layout): green for compute nodes, orange-bordered for conditional, gray for post-nodes. Edges are styled by type ($res solid blue, $dep dashed, $barrier thick orange). Export Python downloads a ready-to-run script from the current graph.

JSON + CLI workflow

cargo run -p tomii-core --bin main -- \
  --json graph.json --dylib plugin.so \
  --workers 4 --slots 2 --max-frames 100

Key flags: --workers, --slots, --system-threads, --timing, --report, --fifo, --custom, --inline-continuation, --coalesce-barriers, --resolution-strategy multi-slot-batch.

Full flag list: cargo run -p tomii-core --bin main -- --help

Pluggable scheduler

Implement the TaskScheduler trait (defined entirely in tomii-types — no tomii-core dependency required) and load it at runtime:

cargo build --release -p scheduler-plugin
cargo run -p tomii-core --bin main -- \
  --scheduler-plugin target/release/libscheduler_plugin.so \
  --json graph.json --dylib plugin.so

See examples/scheduler-plugin/ for a minimal FIFO example and tomii-core/PLUGIN_SCHEDULER_API.md for the stability contract.

Flagship examples

Example #1 — Public 4×4 MIMO uplink benchmark

cd bench/mimo-bench
# requires Intel MKL, Agora sender (see README)
python tomii/run_bench.py --workers 4 --slots 4 --frames 200
python taskflow/run_bench.py --workers 4 --slots 4 --frames 200

Tomii dispatches FFT tasks as each UDP packet arrives; Taskflow must collect all packets before submitting the full DAG. This packet-overlap is the structural advantage the benchmark measures. See bench/mimo-bench/README.md for topology and metric.

Example #2 — Multi-frame pipeline S-scaling

cd bench/pipeline-bench
python tomii/run_bench.py --workers 4 --slots 16 --frames 200
python taskflow/run_bench.py --workers 4 --slots 16 --frames 200

Sweeps concurrent-slot count (S) to show how per-frame scheduling overhead amortises as multi-slot reuse takes effect. Run run_bench.py to generate the sweep CSV locally (run outputs are not committed). See bench/pipeline-bench/README.md for the topology and metric.

Tomii runs use --custom --coalesce-barriers --inline-continuation (hardcoded in run_bench.py); these are the recommended flags for streaming workloads. Taskflow uses default tf::Executor.

Ergonomics #1 — Agent-native graph tuning

cd examples/agent-tuning
bash run_all.sh 50   # runs all 4 arms (random, Bayesian, grid, Claude)

Four optimisation arms (random, Bayesian, grid, Claude) compete over the stream-analytics knob space with the same budget (50 iterations) and verifier. The agent converges without being given source code or documentation; an edit that drops a barrier or removes a $dep edge fails the verifier and is rejected. See examples/agent-tuning/README.md.

Ergonomics #2 — Polyglot plugin showcase

The same DAG (FFT + matrix compute) runs with Rust, C, and Python kernels:

python examples/matrix-compute/run_bench.py --workers 4
python examples/matrix-compute-C/run_bench.py --workers 4
python examples/matrix-compute-python/run_bench.py --workers 4

No source changes to the runtime — the plugin boundary is a C ABI, language-agnostic by design. See examples/README.md for the full capability matrix.

Architecture

Workspace crates:

• tomii-core — runtime, scheduler, graph engine, network receiver
• tomii-types — CmTypes enum + stable scheduler API types (SchedulerPriority, SchedulerWorkerRange, CoreSpec)
• tomii-converter — code-generation: wraps Rust/C plugin headers into wrappers.rs/reg.rs
• tomii-macro — procedural macros for plugin wrapping (WIP)
• tomii/ — Python API package

Key runtime modules (tomii-core/src/runtime/):

• resolution_loop.rs — main resolution loop, batch draining
• batch_resolution.rs — four-phase batch processing (Phases 1–3)
• resolution_strategy.rs — ResolutionStrategy trait + MultiSlotBatchStrategy
• task_execution.rs — worker task execution, inline-continuation fast path
• successor.rs — successor collection and dependency propagation
• shared_data.rs — SharedData, SlotData, borrow bundles
• slot_lifecycle.rs — slot completion detection
• ARCHITECTURE.md — detailed threading model, memory ordering, invariants

Threading model:

• Worker threads (Rayon or custom pool) with CPU affinity — kernel execution
• System thread(s) — dependency propagation and slot lifecycle
• Network receiver threads — low-latency UDP/TCP packet ingestion (feature network)
• Async recorder thread — non-blocking timing output

See tomii-core/src/runtime/ARCHITECTURE.md for the performance envelope (intrinsic costs, target workload classes, and workloads Tomii explicitly does not target).

Build

cargo build --release
make schema   # regenerate Python bindings after changing json_structs.rs

Note for bench/mimo-bench/ builds: the MIMO bench links Intel MKL and Agora libs. Run source examples/mimolib/scripts/export.sh before building to set the required library paths. See bench/mimo-bench/README.md for the full dependency list.

Environment variables

• FUNC_PATH — path to plugin header or Rust source (required for tomii-converter)
• WRAP_PATH / REG_PATH — wrapper/registry files (optional, auto-generated)

License

Apache License 2.0