reports/memory_estimate.json

{
  "expert_params": 7751073792,
  "non_expert_params": 716783040,
  "stored_weights_gb": {
    "experts_int8": 7.751073792,
    "non_experts_bf16": 1.43356608,
    "scale_metadata_estimate": [
      0.25,
      0.5
    ],
    "total_estimate": [
      9.43,
      9.68
    ]
  },
  "training_gradient_pressure_gb": {
    "expert_grad_float32_if_naive_ste": 31.004295168,
    "expert_grad_bf16_if_supported": 15.502147584,
    "expert_grad_int8_sign_if_custom_optimizer": 7.751073792
  },
  "blocker": "A full-model value_and_grad path must store expert gradients. Naive STE float gradients alone are ~31 GB for experts."
}