{
  "observation": "Start from a safe tp=4 layout and conservative batching.",
  "diagnosis": "The first pass should verify multi-GPU launch and avoid queueing collapse from over-batching.",
  "config_patch": {
    "env_patch": {},
    "flag_patch": {
      "tensor-parallel-size": 4,
      "max-num-seqs": 8,
      "max-num-batched-tokens": 32768,
      "gpu-memory-utilization": 0.85,
      "block-size": 64
    }
  },
  "expected_effects": [
    "Stable startup on 4x H20",
    "Low risk of OOM during the first binary-search probes"
  ],
  "why_not_previous_failures": "This is the initial baseline proposal."
}