{
  "model_revision": "ed40e7da625f594fc9c4990f5f66be24873c22d7",
  "recipe_trl_revision": "e498acc637084cffff14deb09ccdc423a766adb2",
  "target_modules_in_recipe": [
    "q_proj",
    "k_proj",
    "v_proj",
    "o_proj",
    "gate_proj",
    "up_proj",
    "down_proj"
  ],
  "rank": 32,
  "alpha": 64,
  "literal_peft_error": "Target module Inkling2LoopQueryProjection",
  "text_layers": 24,
  "routed_experts_per_layer": 128,
  "underlying_attention_linears": 96,
  "fused_gate_up_parameter_banks": 24,
  "down_parameter_banks": 24,
  "independent_gate_and_up_witness": {
    "gate_rank": 32,
    "up_rank": 32,
    "stacked_rank": 64
  },
  "learner_layout": {
    "peft_version": "0.21.2",
    "trainable_parameters": 1032585216,
    "banked_factor_tensors": 336,
    "attention_projection_pairs": 96,
    "expert_projection_pairs": 9216,
    "serving_factor_tensors": 18624,
    "scope": "Released multimodal configuration instantiated on the meta device, without loading weights."
  },
  "training_and_serving_fixture": {
    "learner_fork": "https://github.com/idoru/trl/commit/5e65f9f2548851e1355efb64f5f00059df5da99a",
    "reference_vllm_commit": "752a3a504485790a2e8491cacbb35c137339ad34",
    "vllm_cpu_image": "vllm/vllm-openai-cpu@sha256:6b301f040db8152dfb8ff55e06fd348aa5d0d9a311f58118160c7058262c8628",
    "training_device": "CPU",
    "training_torch": "2.13.0",
    "training_peft": "0.21.2",
    "training_transformers": "66f51e76efd8542232b8bdd40ade7d1edbaac46f with the published Inkling backport",
    "vllm_version": "0.25.1",
    "vllm_torch": "2.11.0+cpu",
    "training_checks": [
      "Routed projections, input gradients, routing-weight gradients and all six expert LoRA factor gradients matched an independent per-expert dense calculation.",
      "The real HarnessAsyncGRPOTrainer computed packed loss and gradients and applied an optimizer update to a synthetic token fixture. Its HTTP peer was mocked.",
      "A separate two-step fixture saved distinct PEFT checkpoints and versioned serving adapters. Reloading the first training checkpoint reproduced its logits exactly.",
      "57 learner, packing and upstream regression tests passed. Two existing upstream tests were marked expected failures."
    ],
    "reader": "LoRAModel.from_local_checkpoint and PackedLoRALayerWeights.pack_moe from the unmodified official CPU release",
    "versions": 2,
    "reader_dtypes": ["float32", "bfloat16"],
    "projection_pairs_per_version": 128,
    "factor_values_per_version_per_dtype": 458752,
    "max_factor_difference_after_reader_cast_and_scaling": 0.0,
    "combined_gate_up_rank_in_both_versions_and_dtypes": 64,
    "scope": "Small synthetic eight-layer Inkling model with four experts. The complete GPU model, generation server, public harness, prefix cache and production training updates are outside this check.",
    "verification_script": "jobs/check_inkling_vllm_loader.py"
  }
}
