{
 "file": "gpu-ql3090-24g-nemo-q4-k-m-control-3090-q4_K_M-ladder.json",
 "model_tag": "mistral-nemo:12b-instruct-2407-q4_K_M",
 "arm": "control-3090",
 "box_class": "gpu-ql3090-24g",
 "quantization": "Q4_K_M",
 "store_gib": 6.96,
 "largest_whole": 131072,
 "topped_out": true,
 "stopped_at": null,
 "errors": [],
 "ups_before": 12.0,
 "ups_after": 33.0,
 "windows": [
  {
   "num_ctx": 4096,
   "whole": true,
   "verdict": "fits: fully resident on the card",
   "vram_delta_mib": {
    "<withheld: GPU UUID of board A>": 7535.0
   },
   "runs": 3,
   "decode_median": 88.203,
   "ttft_ms_median": 481.8,
   "board_w_median": 259.65,
   "j_per_1k_median": 2948.31,
   "sw_power_cap_fraction_median": 1.0,
   "temp_max_c": 48.0,
   "spread": 0.0029,
   "decode_range": 0.252,
   "throttle": [],
   "stop": null,
   "gate_refusals": 0,
   "gates_failed": [],
   "passes": true,
   "headline_decode": 88.203
  },
  {
   "num_ctx": 8192,
   "whole": true,
   "verdict": "fits: fully resident on the card",
   "vram_delta_mib": {
    "<withheld: GPU UUID of board A>": 470.0
   },
   "runs": 3,
   "decode_median": 87.944,
   "ttft_ms_median": 478.2,
   "board_w_median": 262.11,
   "j_per_1k_median": 2980.41,
   "sw_power_cap_fraction_median": 1.0,
   "temp_max_c": 54.0,
   "spread": 0.0031,
   "decode_range": 0.276,
   "throttle": [],
   "stop": null,
   "gate_refusals": 0,
   "gates_failed": [],
   "passes": true,
   "headline_decode": 87.944
  },
  {
   "num_ctx": 16384,
   "whole": true,
   "verdict": "fits: fully resident on the card",
   "vram_delta_mib": {
    "<withheld: GPU UUID of board A>": 696.0
   },
   "runs": 3,
   "decode_median": 87.62,
   "ttft_ms_median": 507.83,
   "board_w_median": 261.57,
   "j_per_1k_median": 2983.24,
   "sw_power_cap_fraction_median": 1.0,
   "temp_max_c": 59.0,
   "spread": 0.0035,
   "decode_range": 0.31,
   "throttle": [],
   "stop": null,
   "gate_refusals": 0,
   "gates_failed": [],
   "passes": true,
   "headline_decode": 87.62
  },
  {
   "num_ctx": 32768,
   "whole": true,
   "verdict": "fits: fully resident on the card",
   "vram_delta_mib": {
    "<withheld: GPU UUID of board A>": 1392.0
   },
   "runs": 3,
   "decode_median": 87.234,
   "ttft_ms_median": 458.02,
   "board_w_median": 262.63,
   "j_per_1k_median": 3012.02,
   "sw_power_cap_fraction_median": 1.0,
   "temp_max_c": 63.0,
   "spread": 0.001,
   "decode_range": 0.086,
   "throttle": [],
   "stop": null,
   "gate_refusals": 0,
   "gates_failed": [],
   "passes": true,
   "headline_decode": 87.234
  },
  {
   "num_ctx": 49152,
   "whole": true,
   "verdict": "fits: fully resident on the card",
   "vram_delta_mib": {
    "<withheld: GPU UUID of board A>": 1432.0
   },
   "runs": 3,
   "decode_median": 87.003,
   "ttft_ms_median": 460.24,
   "board_w_median": 264.12,
   "j_per_1k_median": 3030.27,
   "sw_power_cap_fraction_median": 1.0,
   "temp_max_c": 64.0,
   "spread": 0.0026,
   "decode_range": 0.23,
   "throttle": [],
   "stop": null,
   "gate_refusals": 0,
   "gates_failed": [],
   "passes": true,
   "headline_decode": 87.003
  },
  {
   "num_ctx": 65536,
   "whole": true,
   "verdict": "fits: fully resident on the card",
   "vram_delta_mib": {
    "<withheld: GPU UUID of board A>": 1350.0
   },
   "runs": 3,
   "decode_median": 86.984,
   "ttft_ms_median": 512.36,
   "board_w_median": 263.46,
   "j_per_1k_median": 3027.95,
   "sw_power_cap_fraction_median": 1.0,
   "temp_max_c": 63.0,
   "spread": 0.0005,
   "decode_range": 0.046,
   "throttle": [],
   "stop": null,
   "gate_refusals": 0,
   "gates_failed": [],
   "passes": true,
   "headline_decode": 86.984
  },
  {
   "num_ctx": 98304,
   "whole": true,
   "verdict": "fits: fully resident on the card",
   "vram_delta_mib": {
    "<withheld: GPU UUID of board A>": 2880.0
   },
   "runs": 3,
   "decode_median": 86.934,
   "ttft_ms_median": 481.49,
   "board_w_median": 264.36,
   "j_per_1k_median": 3038.02,
   "sw_power_cap_fraction_median": 1.0,
   "temp_max_c": 64.0,
   "spread": 0.0019,
   "decode_range": 0.167,
   "throttle": [],
   "stop": null,
   "gate_refusals": 0,
   "gates_failed": [],
   "passes": true,
   "headline_decode": 86.934
  },
  {
   "num_ctx": 131072,
   "whole": true,
   "verdict": "fits: fully resident on the card",
   "vram_delta_mib": {
    "<withheld: GPU UUID of board A>": 2880.0
   },
   "runs": 3,
   "decode_median": 86.883,
   "ttft_ms_median": 493.09,
   "board_w_median": 261.21,
   "j_per_1k_median": 3005.62,
   "sw_power_cap_fraction_median": 1.0,
   "temp_max_c": 64.0,
   "spread": 0.0007,
   "decode_range": 0.062,
   "throttle": [],
   "stop": null,
   "gate_refusals": 0,
   "gates_failed": [],
   "passes": true,
   "headline_decode": 86.883
  }
 ],
 "resident_named_card_growth_mib": 7535.0,
 "resident": true,
 "passing_windows": [
  4096,
  8192,
  16384,
  32768,
  49152,
  65536,
  98304,
  131072
 ]
}