Statevector memory budget ignores container (cgroup) memory limits #264

Hi! A heads-up for running tsim in containers or under job schedulers.

tsim.statevector.shots.memory_budget() sizes chunks and parts from psutil.virtual_memory().available // 2, which reports the host's free memory. Inside a cgroup with a lower limit (Docker, Kubernetes, Slurm, systemd-run -p MemoryMax=...), tsim can therefore plan far more memory than the process is allowed. The kernel then OOM-kills it instead of tsim falling back to smaller chunks.

For example, the fuzzing run ran this circuit under a 6 GB MemoryMax scope: a 2-qubit circuit with 61 measurements and one width-16 component, which needs about 3 MB per shot. sample(1000) planned for about 28 GB and was OOM-killed.

Possible fix: cap the budget by the cgroup limit when one is set. That means memory.max minus memory.current on cgroup v2, or memory.limit_in_bytes on v1. An environment variable or argument to override the budget would also help.

This predates the 1.0.0 candidate (#259); it isn't a regression.

🤖 Generated with Claude Code