flopscope-server 0.8.0rc4

Backend server for flopscope client-server architecture

flopscope-server

Server-side runtime for the flopscope client-server architecture.

flopscope-server runs the heavy NumPy-backed flopscope library on behalf of remote flopscope-client processes that connect over ZMQ + msgpack. Use it when you want to isolate untrusted or sandboxed code (typically participant submissions in the ARC Whitebox Estimation Challenge) from the actual computation environment.

When to install this

You want this on the server side of a client-server flopscope deployment. It pulls in flopscope, numpy, pyzmq, and msgpack:

pip install flopscope-server

Equivalent extra on the main flopscope distribution (installs both at the same exact version):

pip install "flopscope[server]"

Quick start

Launch the server

# Over a UNIX domain socket (recommended for local / single-host deployments)
python -m flopscope_server --url ipc:///tmp/flopscope.sock

# Or over TCP
python -m flopscope_server --url tcp://127.0.0.1:15555

Connect a client

In a separate process (typically a sandboxed container) that has flopscope-client installed:

export FLOPSCOPE_SERVER_URL=ipc:///tmp/flopscope.sock

import flopscope as flops
import flopscope.numpy as fnp

with flops.BudgetContext(flop_budget=1_000_000):
    a = fnp.array([1.0, 2.0, 3.0])
    b = fnp.array([4.0, 5.0, 6.0])
    fnp.add(a, b)
    flops.budget_summary()

The client's first request performs a version handshake with the server; mismatched flopscope versions raise ConnectionError with both versions in the message. Keep flopscope-server and flopscope-client on the same release.

Architecture overview

┌─────────────────────────────┐         ZMQ + msgpack         ┌──────────────────────────────┐
│  flopscope-client process   │ ───────────────────────────▶  │  flopscope-server process    │
│  (sandbox; no numpy)        │                                │  (full flopscope + numpy)    │
│  `import flopscope as flops`│ ◀───────────────────────────  │  Executes ops, tracks budget │
└─────────────────────────────┘                                └──────────────────────────────┘

Every counted operation (matmul, einsum, FFT, reductions, …) is dispatched to the server, executed against the real library, and the result + remaining budget streamed back. The client never sees numpy.

Related

• flopscope — the NumPy-backed library hosted by this server
• flopscope-client — the lightweight proxy clients use to talk to a flopscope-server
• Documentation — full guides including client/server architecture and Docker recipes
• GitHub — source, CHANGELOG, contributor guide

License

MIT