{
  "source": "local startup log; model-loading scope verified against installed vLLM model_runner.py",
  "model_loading_gib": 19.13,
  "model_loading_includes_draft": true,
  "kv_cache_pool_gib": 5.5,
  "kv_cache_capacity_tokens": 108081,
  "graph_capture_gib": 1.08,
  "peak_activation_gib": 2.21,
  "weights_plus_non_torch_gib": 23.52,
  "gpu_memory_utilization": 0.98,
  "max_model_len": 40960,
  "max_num_seqs": 3,
  "observed_vram_peak_gib": 29.42,
  "note": "different measurement times/scopes; overlapping categories must not be summed as an exact allocation ledger",
  "bf16_embedding_gib": 2.3681640625,
  "bf16_lm_head_gib": 2.3681640625,
  "bf16_embedding_and_head_shape": [
    248320,
    5120
  ],
  "bf16_payload_source": "current target-mtpfp8 safetensors header"
}