config.json

{
  "_README": "POD variant of config.json for Jv2 stages 00-03 on a single Pro 6000 (96GB). Models pulled from HF; paths are pod-local (/workspace). it_base + instruct_fisher are NULL here — they belong to stage 04 (merge) which runs on AIBOX, not the pod. Finalize/verify bin paths + fisher params ON the pod (jaxx: expect quirks). Keep the local config.json as the AIBOX/stage-04 source of truth.",
  "_MODELS_ARE_REAL_J_INPUTS": "SET 2026-05-29 07:18Z (off the old Gv3/Equinox smoke placeholders). J-line = Iv-winner ⊕ toasty's glimmer-rp v0.1. Equinox is NOT a J input — it is already INSIDE the Iv-winner (Iv3 = Gv3-rebuilt + Equinox). REAL Jv2 style ratio is between A = the LOCKED Iv-winner (Iv3_rebuilt, scp'd to /workspace/jv2) and B = glimmer-on-Iv-winner (= Jv1; fold glimmer-rp v0.1 LoRA onto A on the pod). The separate Gv3/Equinox smoke (#74) is DROPPED from this run — first real run IS the smoke (README). #74 stays a later optional follow-up.",
  "run_label": "style_mask_j1_pod",
  "goal": "reduce slop + respect character details (personality, physical traits, voice)",
  "seed": 1729,

  "single_gpu": true,
  "cuda_devices": "0",
  "_gpu_note": "ONE 96GB card. run_all.sh serves with CUDA_VISIBLE_DEVICES=0,1 --tensor-split 1,1 (2-GPU aibox). On the pod: CUDA_VISIBLE_DEVICES=0 and DROP --tensor-split. See POD_RUNBOOK.md.",

  "models": {
    "A": {
      "name": "Iv_winner",
      "hf_repo": null,
      "_source": "Iv3 (Gv3-rebuilt + Equinox, thinkfisher 0.3/0.8) — scp'd from aibox, NOT pulled from HF.",
      "safetensors_dir": "/workspace/jv2/g4_31b_recipe_Iv3_rebuilt",
      "gguf": "/workspace/style_mask_j1/gguf/Iv_winner.Q4_K_M.gguf"
    },
    "B": {
      "name": "glimmer_on_Iv",
      "hf_repo": null,
      "_source": "= Jv1. Fold glimmer-rp v0.1 LoRA (/workspace/jv2/gemma4-31b-glimmer-rp-adapter) onto A (the Iv-winner) on the pod via fold_jv1.py -> this dir.",
      "safetensors_dir": "/workspace/jv2/Jv1",
      "gguf": "/workspace/style_mask_j1/gguf/glimmer_on_Iv.Q4_K_M.gguf"
    }
  },
  "it_base": null,
  "instruct_fisher": null,
  "instruct_layer_importance": null,
  "_stage04_inputs_note": "it_base=google/gemma-4-31B-it, instruct_fisher=g4_31b_it_fisher/g4-31b-it/fisher.pt (55GB local), instruct_layer_importance=...layer_importance.json — ALL aibox-local, used only in 04_merge_style.py. Never copied to the pod.",

  "eso5_prompts": "/workspace/style_mask_j1/prompts_used.jsonl",
  "n_prompts": 40,

  "generate": {
    "port": 8091,
    "ctx_size": 32768,
    "parallel": 8,
    "max_tokens": 1024,
    "temperature": 0.0,
    "workers": 8,
    "use_bf16_transformers": false
  },

  "fisher": {
    "backend": "pod_local",
    "_backend_note": "Run 02_extract_style_fisher.py directly on the pod GPU. On 96GB a single card likely fits without --cpu-offload; start with --gpu-memory-gib 90 and add --cpu-offload only if OOM. Verify on pod.",
    "max_length": 1024,
    "layers_per_batch": 8,
    "fisher_dtype": "bf16",
    "gpu_memory_gib": "90",
    "cpu_offload": false,
    "enable_thinking": false,
    "_enable_thinking_note": "Eso-5/style is thinking-OFF; extract style-Fisher thinking-off to match how the merge is evaluated."
  },

  "ratio": {
    "epsilon": 1e-8,
    "combined_floor_quantile": 0.5
  },

  "merge": {
    "mode": "v2_two_model_style",
    "density": null,
    "scale": null,
    "_note": "STAGE 04 — AIBOX ONLY. Not run on the pod. Fill density/scale from the #65 sweep winner before stage 04 runs on aibox."
  },

  "upload": {
    "ratio_repo": "Esobold/g4_31b_style_ratio_j1",
    "ratio_repo_private": true,
    "files": ["ratio/style_ratio.pt", "ratio/ratio_layer_summary.json"],
    "_note": "After stage 03, upload style_ratio.pt (~35GB) to the pre-created private Esobold repo, then pull HF->aibox for stage 04."
  },

  "disk": {
    "min_free_gb_before_download": 400,
    "_footprint": "Gv3 ~63GB + Equinox ~62GB safetensors (~125) + per-model transient bf16 GGUF ~62GB during convert (deleted after) + 4 style-Fishers ~35GB each (~140) + ratio ~35GB. df preflight before download AND before stage 02.",
    "_twisted_2026_05_29": "Sanity-check disk so it doesn't run out partway through."
  },

  "bin": {
    "_note": "SET ON POD — point at the pod's llama.cpp build + venv python. Placeholders below assume a /workspace build.",
    "python": "python",
    "llama_server": "/workspace/llama.cpp/build/bin/llama-server",
    "llama_quantize": "/workspace/llama.cpp/build/bin/llama-quantize",
    "llama_convert": "/workspace/llama.cpp/convert_hf_to_gguf.py"
  }
}