{
 "buffer_lines": [
  "load_tensors:   CPU_Mapped model buffer size =   322.07 MiB",
  "load_tensors:        CUDA0 model buffer size = 51945.51 MiB",
  "load_tensors:   CPU_Mapped model buffer size = 27465.95 MiB"
 ],
 "cuda0_mib": 51945.51,
 "cuda0_vs_expect_pct": -1.7,
 "device_lines": [
  "using device CUDA0 (NVIDIA RTX PRO 6000 Blackwell Workstation Edition)"
 ],
 "expect": {
  "cuda0_mib": "52844 +/- 2%",
  "ngram_cpu_mapped_mib": "27465.95 (+/- 0.5), named by the lazy-read line of per_layer_token_embd.weight",
  "offloaded": "49/49",
  "token_embd_cpu_mapped_mib": "<= 2400"
 },
 "lazy_lines": [
  "add: tensor per_layer_token_embd.weight (size = 27465 MiB) lazy read enabled"
 ],
 "ngram_cpu_mapped_mib": 27465.95,
 "ngram_line": "load_tensors:   CPU_Mapped model buffer size = 27465.95 MiB",
 "offload_lines": [
  "offloaded 49/49 layers to GPU"
 ],
 "token_embd_cpu_mapped_mib": 322.07,
 "token_embd_line": "load_tensors:   CPU_Mapped model buffer size =   322.07 MiB",
 "verdict": "whole",
 "why": "every layer on the card (CUDA0); the input lookup tables in host memory: token_embd 322.07 MiB and the n-gram table 27465.95 MiB, as registered"
}