docs/validation-k2-262k.json

{
  "profile": "sglang-fp8patch-262k-k2-w144-v1",
  "checkpoint": "sglang-mixed-fp8vision-v2",
  "hardware": "RTX PRO 6000 Blackwell Server Edition",
  "input_tokens": 261632,
  "output_tokens": 56,
  "decode_tokens_per_second": 106.68489902629287,
  "e2e_seconds": 187.1946626380086,
  "mtp": 1.0,
  "accepted_drafts": 38,
  "proposed_drafts": 38,
  "flashinfer_workspace_mib": 144,
  "monitor_peak_mib": 32044.0,
  "remaining_vs_32_gib_mib_monitor": 724.0,
  "monitor_scope": "Server startup, three vision probes, 261632-token text probe, quality suite; PRO 6000 device total NVML usage. Not concurrent image-plus-262K input.",
  "checks": {
    "prompt_count": true,
    "no_prefix_cache": true,
    "facts_correct": true,
    "under_30_8gib_sampled": true,
    "under_31_5gib_sampled": true,
    "under_32gib_sampled": true
  },
  "quality_passed": 6,
  "quality_total": 10,
  "limitations": [
    "Not a 5090D hardware acceptance.",
    "56-token synthetic retrieval output is not a sustained decode benchmark.",
    "6/10 strict quality suite and 3/3 synthetic image tests are not broad capability proof."
  ],
  "evidence_archive_sha256": "4CB732B74386F04A32CDA535C7986193676B780C961F4E15FC81F4C423E12462"
}