Hardware and version boundaries
Actual training allocations:
- ep2: 1 node(s), 4 × NVIDIA A100-SXM4-40GB, 40960 MiB per GPU. Scheduler requested 64 CPUs and 440 GB RAM per node; this is not measured usage. CPU model: not captured in preserved allocation/binding record.
- ep6: 2 node(s), 8 × NVIDIA A100-SXM4-40GB, 40960 MiB per GPU. Scheduler requested 64 CPUs and 440 GB RAM per node; this is not measured usage. CPU model: not captured in preserved allocation/binding record.
Actual evaluation inference used NVIDIA A100-SXM4-40GB (42406903808 reported bytes), PyTorch 2.8.0+cu128, CUDA 12.8, one model replica per reported server; six max clients. Episode workers: epoch2 4/6 across recorded batches; epoch6 6. Rendering used RTX 5090 32GB.
A pre-evaluation workstation snapshot dated 2026-09-18 reports Intel i9-12900KF, 24 logical CPUs, 62 GiB RAM, RTX 5090 32607 MiB, Ubuntu 22.04.5 and driver 580.178.04. That older snapshot is not proof that the driver/CPU/RAM was rechecked at every rollout. Training CPU model, actual memory usage and inference-server CPU/RAM are not recovered in the supplied facts. Do not substitute hardware from later experiments.
No hardware was exercised for this release; these statements describe historical runs. 原始记录保存在本地。