The daemon

nutorchd is the process that owns the tensor registry, the LibTorch context, GPU memory, and autograd graphs. Clients talk to it over a Unix socket. Its lifecycle is invisible plumbing: you never start it, and by default it cleans up after itself.

Auto-start and the idle TTL

Any torch command starts the daemon if it isn’t running. It shuts itself down after 1 hour of inactivity — and every tensor operation renews the lease, so the clock only runs while you’re idle. Tensors live exactly as long as the daemon: that is the memory-horizon contract. Export what you want to keep (see tensors); everything else returns to the GPU when the daemon retires.

The default TTL is configurable via NUTORCHD_TTL (e.g. 30m, 2h, none for no expiry).

Inspecting and controlling it

torch daemon status      # pid, ttl, idle time, time remaining,
                         # tensor count, memory held, socket, log
torch daemon ttl 4h      # change the idle TTL on the live daemon (none = forever)
torch daemon stop        # shut down now
torch daemon restart     # fresh daemon, empty registry
torch daemon start       # start without running an op

torch daemon status    # pid, ttl, idle time, time remaining,
                         # tensor count, memory held, socket, log
torch daemon ttl 4h    # change the idle TTL on the live daemon (none = forever)
torch daemon stop      # shut down now
torch daemon restart   # fresh daemon, empty registry
torch daemon start     # start without running an op

torch daemon status --json emits the same record as JSON for scripts.

A status report looks like:

version: 0.1.0
pid: 32799
uptime: 240s
device: mps
ttl: 3600s
idle: 12s
remaining: 3588s
tensors: 3
memory: ~24 bytes
socket: /tmp/…/nutorchd.sock
log: /tmp/…/nutorchd.log

Concurrency

Multiple shells can talk to one daemon at the same time. Connections are served concurrently (thread per connection); operations execute strictly serialized inside the daemon, so a parked or slow client never blocks anyone else’s commands.

GPU-only, by design

The daemon requires MPS and refuses to start without it. There is no device option anywhere in the API — every tensor lives on the Apple-silicon GPU.