compression_config.json

{
  "format": "gemma4-pmra-orbitquant-policy-v1",
  "created_utc": "2026-05-14T01:13:53.078680+00:00",
  "base_model": "google/gemma-4-E2B-it",
  "source_result": "C:\\Users\\power\\Documents\\Codex\\2026-05-12\\i-m-on-a-mission-to\\orbitquant\\results\\modal_gemma4_pmra_orbit_stack_trim128_safe3_latest.json",
  "source_profile": "trim128_safe3",
  "static_weight_state": {
    "method": "PMRA",
    "variant": "c2_calib_knapsack_mixed",
    "payload_bpw": 5.32661328956758,
    "payload_bytes": 3094396044
  },
  "selected_policy": {
    "name": "safe3",
    "total_buses": 10,
    "kv_layers": [
      {
        "layer": 33,
        "bits": 3,
        "rotation": "hadamard",
        "alpha": 0.75
      },
      {
        "layer": 28,
        "bits": 3,
        "rotation": "hadamard",
        "alpha": 0.75
      },
      {
        "layer": 30,
        "bits": 3,
        "rotation": "hadamard",
        "alpha": 0.75
      },
      {
        "layer": 16,
        "bits": 3,
        "rotation": "hadamard",
        "alpha": 0.75
      },
      {
        "layer": 18,
        "bits": 3,
        "rotation": "hadamard",
        "alpha": 0.75
      },
      {
        "layer": 11,
        "bits": 3,
        "rotation": "hadamard",
        "alpha": 0.75
      },
      {
        "layer": 15,
        "bits": 3,
        "rotation": "hadamard",
        "alpha": 0.75
      }
    ],
    "mlp_choices": [
      {
        "layer": 20,
        "bits": 2,
        "primitive": "plus",
        "rotation": "preperm_activation_max_hadamard",
        "alpha": 0.375,
        "block_size": 512
      },
      {
        "layer": 19,
        "bits": 2,
        "primitive": "plus",
        "rotation": "preperm_activation_max_hadamard",
        "alpha": 0.375,
        "block_size": 512
      },
      {
        "layer": 6,
        "bits": 2,
        "primitive": "plus",
        "rotation": "preperm_boundary_rms_hadamard",
        "alpha": 0.375,
        "block_size": 512
      }
    ]
  },
  "evaluation": {
    "tokens": 24058,
    "dataset": "wikitext",
    "dataset_config": "wikitext-2-raw-v1",
    "prompt_count": 128,
    "calibration_prompt_count": 24,
    "eval_max_length": 192,
    "calib_max_length": 192,
    "pmra_nll": 12.818462451820967,
    "q3_k_s_nll": 18.04630691050413,
    "stack_nll": 12.834082924860832,
    "delta_nll_vs_pmra": 0.015620473039865246,
    "delta_nll_vs_q3_k_s": -5.212223985643297,
    "last_logit_mse_to_fp16": 67.99472899734974,
    "top10_overlap_to_fp16": 0.13593750000000002,
    "runtime_savings_estimate": {
      "context_length": 8192,
      "mlp_lifetime_tokens": 64,
      "kv_saved_mib": 45.5,
      "mlp_saved_mib": 3.28125,
      "total_saved_mib": 48.78125,
      "mlp_dims": [
        12288,
        12288,
        6144
      ]
    }
  },
  "method": {
    "kv_cache": "Rotate K/V activations with Hadamard before 3-bit scalar quantization.",
    "mlp_intermediate": "Rotate Gemma4 MLP intermediate activations before calibrated 2-bit quantization, then invert before down_proj.",
    "calibration": "Hadamard-plus MLP choices reconstruct prepermutation order from calibration activations and down-proj weights."
  }
}