Skip to main content

Spatial Computing

Measure. Decide. Prove.

One objective for the whole infrastructure: get every job done correctly at the lowest cost, and prove it. Spatial Computing prices every way to run a job across compute, memory, runtime, AI and a fleet of machines, picks the cheapest, and records the estimate next to what really happened. A gain is claimed only when it is measured. It does not add RAM, cores or GPUs: it compares execution strategies on the hardware you have, trades one resource for another in numbers, and says plainly what it cannot do.

Status: v0.22.0, early-stage open-source toolkit. Everything here was built and measured on a single 2017 MacBook Pro (Intel i7-7660U, 2 cores, 16 GB). Nothing has been proven on a second machine or a data-center GPU. See Limitations and the site.

Français

Install

Python 3.10 or newer.

python3 -m venv .venv && source .venv/bin/activate
pip install spatial-computing
spatial audit

From source, to develop:

git clone https://github.com/abraham1116/spatial-computing && cd spatial-computing
pip install -e ".[dev]"
pytest

Optional Linux micro-VM route (needs QEMU and two files, checked by SHA-256):

mkdir -p vendor && cd vendor
curl -O https://dl-cdn.alpinelinux.org/alpine/v3.22/releases/x86_64/netboot/vmlinuz-virt
curl -o busybox-x86_64 https://busybox.net/downloads/binaries/1.35.0-x86_64-linux-musl/busybox
shasum -a 256 -c SHA256SUMS

Console (accounts, apps, API keys, jobs, billing)

spatial token init --master-key-file hub.key
spatial serve hub --master-key-file hub.key --port 7700 --usage-file usage.jsonl --revoked-file revoked.txt
spatial platform serve --master-key-file hub.key --hub 127.0.0.1:7700 --usage-file usage.jsonl --revoked-file revoked.txt
# open http://127.0.0.1:8090/ : sign up, create an app, create a node key and a client key
spatial agent --hub 127.0.0.1:7700 --token-file node.tok     # the machine appears in Fleet; submit jobs from Jobs

Production (hub, console, HTTPS): docker compose --env-file deploy/.env up -d --build. See docs/PLATFORM.md and docs/DEPLOY.md.

Use

spatial audit                       # hardware, current usage, limits of this machine
spatial status                      # honest status of every engine
spatial plan app.exe                # detected format + execution routes, with reasons
spatial run  app.elf 600 300        # run through the first feasible route (micro-VM here)
spatial bench all                   # redo every measurement and regenerate the results table

Tiered memory as a library:

from spatial.memory import TieredStore, FileBackend
st = TieredStore(hot_bytes=16 << 20, compressed_bytes=16 << 20,
                 far=FileBackend("/tmp/far.bin"), prefetch=16, keep_far_copy=True)
st.put(0, b"..." * 1000); st.get(0); print(st.stats())

What is in the box

Engine Level Evidence
Virtual accelerator fabric (one job, several devices: CPU, Metal GPU, NVIDIA when present) first components tests/test_fabric.py, benchmarks/bench_unified.py
Planner (the one objective: price every option across all engines, choose, run, check) operational prototype tests/test_unified.py, benchmarks/bench_unified.py
Fleet (agent on every machine + hub: measure, place, recommend, audit) operational prototype tests/test_fleet.py, benchmarks/bench_fleet.py
Compute Engine (detect, monitor, compare strategies, offload, cost) operational prototype tests/test_compute.py, benchmarks/bench_cpu.py
Tiered memory (hot / compressed / far) operational prototype tests/test_memory.py, benchmarks/bench_memory.py
Runtime Engine (inspect binaries, route, micro-VM, WASI, containers incl. ARM64 via emulation) operational prototype tests/test_runtime.py, benchmarks/bench_runtime.py
Portable kernels (one source, CPU/Metal/CUDA/HIP/Triton; market-vertical presets) first components tests/test_kernel.py, benchmarks/bench_kernel.py
Live GPU market prices (spatial market, read-only, no account) first components tests/test_market.py, docs/MARKET.md
AI, Intelligence first components (toy model) tests/test_engines.py, tests/test_domains.py

Measured on the reference machine (AC power): Numba JIT ×31.6 over pure Python with an identical result; 0.88 s to boot an isolated Linux micro-VM on macOS; −75 % local RAM for a skewed-access dataset at 107× the mean access latency. Full table and conditions: docs/RESULTATS.fr.md.

What it does not do

  • It does not add RAM. Compressed, on-disk or remote memory is typed far_memory, never ram, and a test enforces it.
  • It does not confuse virtualization with extra resources.
  • It never exposes your machine: servers bind to 127.0.0.1 and refuse anything else without an explicit --allow-remote.

Open core

The core in this repository is licensed under Apache-2.0. Advanced placement and cost policies are developed in a separate private package and are not part of this repository.

Documentation

The site has the full reference: https://spatial-computing-ecru.vercel.app (English and French). docs/LIMITATIONS.md is current and in English; the other in-repo reports are historical v0.1 notes in French, and English translations are welcome (see CONTRIBUTING).

Contributing

Issues and pull requests are welcome. Start with CONTRIBUTING.md and the ROADMAP. Report security problems privately, see SECURITY.md.

Metadata

Release files for spatial-computing 0.24.0

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for spatial-computing 0.24.0
File Size Uploaded Provenance Actions
spatial_computing-0.24.0.tar.gz 448.2 kB Without provenance Details for spatial_computing-0.24.0.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for spatial-computing 0.24.0
File Interpreter ABI Platform Actions
spatial_computing-0.24.0-py3-none-any.whl Python 3 none any Details for spatial_computing-0.24.0-py3-none-any.whl

Total release size: 857.6 kB

Release files / spatial_computing-0.24.0.tar.gz

Download URL spatial_computing-0.24.0.tar.gz
Size 448.2 kB
Tags Source
SHA-256 checksum
How to use checksums
3e6cc51af7303a8d1f8ce6fb8bc2a17d4289abd10355b07b950a533d7a7f7a86
BLAKE2b-256 checksum
How to use checksums
769e1a4d163ff7eb986df4261ba4fcf5f75b4bb5521d1ef0192e785eda19fe7d
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.13.5

Release files / spatial_computing-0.24.0-py3-none-any.whl

Download URL spatial_computing-0.24.0-py3-none-any.whl
Size 409.5 kB
Tags Python 3
SHA-256 checksum
How to use checksums
48ba91be53efd768f2d56f8a9cb208a987114b0d60b34c20a675bc2684d4e92e
BLAKE2b-256 checksum
How to use checksums
240c20b8c67bbcec46a0c78340d60b734d76d7bd9ada3e3a2767b210b24da0db
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.13.5

Security for spatial-computing 0.24.0

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

No provenance found

None of the files in this release were uploaded with provenance metadata.

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page