gpu.json

{
  "_comment": "Optional GPU4PySCF 1.8 VRAM/CPHF tuning. GPU4PySCF owns the numerical kernels and multi-GPU scheduler. Pass with: mpscf ... --gpu-json gpu.json",

  "vram_budget_fraction": 0.70,

  "_workspace": "Use larger gradient tiles during geometry optimization while retaining the conservative Hessian cap.",

  "gradient_vram_budget_fraction": 0.90,
  "hessian_vram_budget_fraction": 0.70,

  "_cphf": "Limit the device-0 Krylov history by solving fewer atoms per CPHF batch. Lower cphf_subspace_frac or raise cphf_est_cycles to reduce peak memory.",
  "cphf_subspace_frac": 0.50,
  "cphf_est_cycles": 20,

  "_xc": "Reuse the invariant ground-state XC kernel and keep one replica per GPU during the Hessian CPHF solve.",
  "cache_xc_kernel": true,
  "cache_xc_kernel_per_device": true,

  "_scaling": "Use NCCL when available, otherwise reduce within P2P islands before crossing PCIe roots. Dynamically schedule RKS CPHF grid chunks and overlap XC with the single-device DF gradient.",
  "multi_gpu_reduction": "auto",
  "multi_gpu_reduction_min_bytes": 1048576,
  "topology_aware": true,
  "topology_reorder_devices": true,
  "hessian_dynamic_grid": true,
  "hessian_grid_chunks_per_device": 2,
  "gradient_overlap_xc_jk": true
}