The machine has an RTX 4060 Laptop GPU with 8 GiB VRAM and roughly 15 GiB host RAM. During RL, vLLM and the trainer both need to stay alive. Host memory also has to hold the orchestrator, environment processes, data and checkpoint staging.

Two copies of the policy

vLLM owns an inference representation of the policy. The trainer owns trainable parameters, gradients and optimizer state. FSDP2 supplies the training machinery, but this profile runs on one physical GPU; it does not gain another device’s memory by calling it distributed. The small presets reduce context, sequence slots, concurrent rollouts and temporary buffers. Activation checkpointing trades recomputation for activation memory. Chunked optimizer state and per-tensor checkpoint writes address different peaks: updating weights and saving them. These are tested configurations for specific workloads. They do not establish capacity for arbitrary batches, context lengths or 1B models.

The guard is part of the design

The launcher checks available host RAM, free GPU memory, disk space and existing compute jobs. Its systemd scope caps job RAM at 7 GiB and permits no job swap. The runtime guard stops the job at low available RAM or disk, an excessive sampled GPU allocation, or its deadline. A guard stop is a failed run. It is not an out-of-memory success or a timing sample. The GPU sampling interval can miss short allocation peaks, so a recorded maximum is a sampled maximum.

Which utilization number?

GPU utilization, memory bandwidth and model FLOP utilization measure different things. A GPU can report activity while waiting on memory or running inefficient launches. MFU is a percentage of estimated useful training FLOPs against the selected hardware peak; changing precision assumptions changes the denominator. The ten-update profiling record’s 70.7% sampled GPU utilization is explicitly not an MFU result. Launcher and guard · Capacity profiles and limits Continue with one RL update.