model:
  repository: "ByteDance/Ouro-1.4B"
  revision: "e3b1e0993b1231a51d6069a870476dda4162c00a"
  config_sha256: "3ad25a6a186865de4cbda24d37c514e8488d8fdf4536fa47c10ec0efd5c8970a"
  modeling_ouro_sha256: "59e7523e9fe23f75bde78bb914d1a4d24f3997ae02bca510bbaef947356c447b"
  physical_layers: 24
  recurrent_loops: 4
  hidden_size: 2048
  intermediate_size: 5632
  dtype: bfloat16

groups:
  attention_qkv:
    hf_weights: [self_attn.q_proj, self_attn.k_proj, self_attn.v_proj]
    vllm_consumer: self_attn.qkv_proj
    transform_site: self_attn.ln_trans
  attention_output:
    hf_weights: [self_attn.o_proj]
    vllm_consumer: self_attn.o_proj
    transform_site: self_attn.o_trans
  mlp_up_gate:
    hf_weights: [mlp.gate_proj, mlp.up_proj]
    vllm_consumer: mlp.gate_up_proj
    transform_site: mlp.up_gate_trans
  mlp_down:
    hf_weights: [mlp.down_proj]
    vllm_consumer: mlp.down_proj
    transform_site: mlp.down_trans

runtime_status:
  adapter_contract: implemented
  vllm_hook_integration: not_implemented
  smoke_generation: not_run

required_vllm_integration_points:
  artifact_load: "OuroForVLLM.__init__ must load and validate the isolated LoopQ artifact."
  qkv_input: "OuroAttention.forward immediately before self.qkv_proj."
  attention_output_input: "OuroAttention.forward immediately before self.o_proj."
  mlp_up_gate_input: "OuroMLP.forward immediately before self.gate_up_proj."
  mlp_down_input: "OuroMLP.forward immediately before self.down_proj."
  weight_load: >-
    OuroForVLLM.load_weights must install the offline folded/RTN W4 projection
    representation without invoking the NVFP4 policy.
  selected_weight_dispatch: >-
    Implemented for eager validation: retain one shared base and four packed
    variants only for each of the four SLT-selected groups, dispatching by
    recurrence index 0..3.
  transition: >-
    OuroModel.forward must apply CTA after completed loops 0, 1, and 2 only,
    after the recurrent boundary norm and before the next loop consumes state.
  smoke_gate: >-
    Load a calibrated artifact and complete vLLM generation with all foreign
    quantization environment policies disabled.
  smoke_command: >-
    CUDA_VISIBLE_DEVICES=0 PYTHONPATH=./loopQ
    python loopQ/scripts/smoke_ouro_vllm.py
    --artifact models/loopq/ouro_w4a4.pt
    --gpu 0