ENOMEM, fork: Cannot allocate memory or Cannot allocate memory means Linux refused a memory-related operation. That can happen because usable memory or swap is exhausted, a container or service limit was reached, or the system’s commit policy rejected the allocation.

On AI servers, model weights, KV cache, request concurrency, tokenization, retrieval workers and containers can compete for different memory pools. Do not assume every allocation failure is GPU VRAM.

Preserve evidence before restarting

free -h
cat /proc/meminfo | grep -E 'MemAvailable|SwapFree|CommitLimit|Committed_AS'
ps -eo pid,ppid,rss,vsz,cmd --sort=-rss | head -20
journalctl -k --since '-30 min' | grep -Ei 'oom|killed process|memory cgroup'

Check whether the kernel killed a process, the allocation was rejected without a kill, or a cgroup enforced its own limit.

Check the actual execution boundary

For systemd:

systemctl show SERVICE --property=MemoryCurrent,MemoryMax,MemoryHigh

For containers:

docker stats --no-stream
cat /sys/fs/cgroup/memory.current 2>/dev/null
cat /sys/fs/cgroup/memory.max 2>/dev/null

For Kubernetes, inspect limits and last termination state. Use Kubernetes OOMKilled troubleshooting when the container was killed.

AI workload causes

• model or quantization no longer fits host RAM;
• context length or batch size increased KV-cache pressure;
• concurrent workers each loaded their own model copy;
• CPU offload moved pressure from VRAM to RAM;
• an index build or embedding job overlapped inference;
• a memory leak accumulated across requests;
• a rollout temporarily ran old and new replicas together.

Correlate the error with release, traffic and workload metrics. The broad AI out-of-memory foundation explains RAM versus VRAM and container OOM behavior.

Recover without hiding the cause

Reduce load or stop the offending workload through the normal operational path. Add temporary swap only when its latency and data-security implications are acceptable; swap is not a capacity plan for active model inference.

Do not change vm.overcommit_memory blindly. Linux commit policies determine whether allocations may exceed backed memory, and changing them can trade immediate ENOMEM for later OOM termination. Review CommitLimit, Committed_AS and workload requirements first.

Long-term fixes include:

• smaller or quantized models;
• bounded context, batches and concurrency;
• shared model processes instead of duplicate workers;
• explicit container reservations and limits;
• staged rollouts with capacity headroom;
• leak detection and per-request memory telemetry;
• separate capacity for indexing and background jobs.

Verify

Repeat the failed operation under controlled load, confirm memory returns after completion and monitor both average and peak use. Record which model, context, concurrency and container limits produced the incident.

Connect remediation to AI Operations and the broader AI out-of-memory foundation. A restart restores availability; evidence and capacity controls prevent recurrence.