model:
  hf_repo: deepreinforce-ai/Ornith-1.0-35B
  release_repo: LordNeel/Ornith-1.0-35B-GGUF-llamacpp-tp1
  served_name: Ornith-1.0-35B

policy:
  serving_parallelism: tp1_only
  tensor_parallel_size: 1
  backend: llama.cpp
  reasoning: "off"
  gpu_usage_note: Use one independent model copy per GPU. Do not split one serving replica across GPUs.
  forbidden:
    - tensor_parallel_size_gt_1
    - multi_gpu_serving

default_profile: llamacpp_q4_k_m_coding_context
lowest_memory_profile: llamacpp_q3_k_m_short_context
middle_footprint_profile: llamacpp_iq4_xs_short_context
adaptive_mtp_profile: llamacpp_iq4_xs_mtp_graft_adaptive

profiles:
  - id: llamacpp_q4_k_m_short_context
    status: recommended_speed_benchmark
    backend: llama.cpp
    tensor_parallel_size: 1
    artifact: ornith-1.0-35b-Q4_K_M.gguf
    quant: Q4_K_M
    command:
      script: scripts/serve_llamacpp_gpu0.sh
      env:
        QUANT: Q4_K_M
        PORT: "8002"
        CTX_SIZE: "8192"
        PARALLEL: "16"
        BATCH_SIZE: "4096"
        UBATCH_SIZE: "512"
        REASONING: "off"
    measured:
      loaded_vram_gib: 21.31
      c1_output_tps: 243.3
      c16_output_tps: 655.6
      c16_p95_ttft_ms: 650.0
      behavior_gate: pass
    evidence:
      benchmark: benchmarks/llamacpp-quant-benchmarks.md
      behavior: probes/behavior-llamacpp-q4-script-default.json
    notes:
      - Short-context benchmark profile exposes about 512 tokens per slot because CTX_SIZE is split across PARALLEL slots.

  - id: llamacpp_q4_k_m_coding_context
    status: recommended_default
    backend: llama.cpp
    tensor_parallel_size: 1
    artifact: ornith-1.0-35b-Q4_K_M.gguf
    quant: Q4_K_M
    command:
      script: scripts/serve_llamacpp_gpu0.sh
      env:
        QUANT: Q4_K_M
        PORT: "8002"
        CTX_SIZE: "131072"
        PARALLEL: "16"
        BATCH_SIZE: "4096"
        UBATCH_SIZE: "512"
        REASONING: "off"
    measured:
      loaded_vram_gib_short_context: 21.31
      c16_output_tps_short_context: 655.6
      behavior_gate: pass
    evidence:
      benchmark: benchmarks/llamacpp-quant-benchmarks.md
      behavior: probes/behavior-llamacpp-q4-script-default.json
    notes:
      - Practical coding-serving profile with 16 slots and about 8k context per slot.
      - Throughput numbers come from the matched short-context sweep.

  - id: llamacpp_q3_k_m_short_context
    status: lowest_memory_validated
    backend: llama.cpp
    tensor_parallel_size: 1
    artifact: ornith-1.0-35b-Q3_K_M.gguf
    quant: Q3_K_M
    command:
      script: scripts/serve_llamacpp_gpu0.sh
      env:
        QUANT: Q3_K_M
        PORT: "8002"
        CTX_SIZE: "8192"
        PARALLEL: "16"
        BATCH_SIZE: "4096"
        UBATCH_SIZE: "512"
        REASONING: "off"
    measured:
      gguf_gib: 15.61
      loaded_vram_gib: 17.27
      c1_output_tps: 240.5
      c16_output_tps: 493.0
      mean_kld_nats_vs_bf16_top64: 0.3620
      behavior_gate: pass
    evidence:
      benchmark: benchmarks/llamacpp-quant-benchmarks.md
      kld: benchmarks/kld-quant-vs-bf16-top64.md
      behavior: probes/behavior-llamacpp-q3-k-m-parallel16.json

  - id: llamacpp_iq4_xs_short_context
    status: middle_footprint_validated
    backend: llama.cpp
    tensor_parallel_size: 1
    artifact: ornith-1.0-35b-IQ4_XS.gguf
    quant: IQ4_XS
    command:
      script: scripts/serve_llamacpp_gpu0.sh
      env:
        QUANT: IQ4_XS
        PORT: "8002"
        CTX_SIZE: "8192"
        PARALLEL: "16"
        BATCH_SIZE: "4096"
        UBATCH_SIZE: "512"
        REASONING: "off"
    measured:
      gguf_gib: 17.64
      loaded_vram_gib: 19.34
      c1_output_tps: 234.1
      c16_output_tps: 476.0
      mean_kld_nats_vs_bf16_top64: 0.1426
      behavior_gate: pass
    evidence:
      benchmark: benchmarks/llamacpp-quant-benchmarks.md
      kld: benchmarks/kld-quant-vs-bf16-top64.md
      behavior: probes/behavior-llamacpp-iq4-xs-parallel16.json

  - id: llamacpp_iq4_xs_mtp_graft_adaptive
    status: experimental_adaptive_mtp
    backend: llama.cpp
    tensor_parallel_size: 1
    artifact: ornith-1.0-35b-IQ4_XS-MTP-graft-headQ6.gguf
    quant: IQ4_XS_with_Q6_MTP_head
    command:
      script: scripts/serve_llamacpp_gpu0.sh
      env:
        MODEL_PATH: ornith-1.0-35b-IQ4_XS-MTP-graft-headQ6.gguf
        ALIAS: Ornith-1.0-35B-GGUF-IQ4_XS-MTP
        PORT: "8002"
        CTX_SIZE: "8192"
        PARALLEL: "16"
        BATCH_SIZE: "4096"
        UBATCH_SIZE: "512"
        REASONING: "off"
        SPEC_TYPE: draft-mtp
        SPEC_DRAFT_N_MAX: "2"
        SPEC_DRAFT_N_MIN: "0"
        SPEC_DRAFT_BACKEND_SAMPLING: "1"
        LLAMA_SPEC_MAX_DRAFTING_SLOTS: "1"
        LLAMA_MTP_FAST_BACKEND_SAMPLE: "1"
        LLAMA_MTP_DRAFT_TOP_K: "1"
        LLAMA_MTP_DRAFT_TOP_P: "1"
        LLAMA_MTP_DRAFT_TEMP: "1"
    measured:
      c1_ar_tps: 221.80
      c1_mtp_adaptive_tps: 319.53
      c1_acceptance_by_position:
        - 0.953
        - 0.865
      c16_ar_tps: 568.57
      c16_adaptive_tps: 564.48
      c16_threshold1_tps: 592.45
    evidence:
      benchmark: benchmarks/llamacpp-iq4-xs-mtp-graft-adaptive-profile.md
      patch: patches/llamacpp-mtp-adaptive-serving-and-recurrent-rollback.patch
    caveats:
      - Always-on MTP is not recommended at saturated c16 because it was slower than target-only decoding.
      - Adaptive throttling uses MTP for low-concurrency requests and disables it under saturated batches.