{
  "format": "companion-forge-runtime/v4",
  "target": {
    "gpu": "NVIDIA L4",
    "compute_capability": "8.9",
    "flavor": "l4x1",
    "cuda": "12.8",
    "tensorrt": "11.2.1.2"
  },
  "runtime_order": {
    "flux2-klein-4b": [
      "tensorrt-512",
      "pytorch-bf16"
    ],
    "dinov2": [
      "tensorrt",
      "onnxruntime-cuda",
      "pytorch"
    ],
    "dsine": [
      "tensorrt",
      "pytorch",
      "onnxruntime-cuda"
    ],
    "anigen-ss-flow-solo": [
      "tensorrt",
      "pytorch"
    ],
    "anigen-ss-decoder": [
      "tensorrt",
      "pytorch"
    ],
    "anigen-slat-flow-core": [
      "tensorrt+spconv-shell",
      "pytorch-spconv"
    ],
    "anigen-skin-decoder": [
      "tensorrt",
      "pytorch"
    ],
    "anigen-slat-decoder-backbone": [
      "tensorrt-custom-sparse",
      "pytorch-spconv-lazy-fallback"
    ]
  },
  "components": {
    "dinov2": {
      "onnx": "production",
      "trt": "production",
      "engine": "engines/l4-sm89/dinov2.plan",
      "input": [
        1,
        3,
        518,
        518
      ],
      "precision": "mixed-fp16",
      "validation": "dinov2_trt_validation.json"
    },
    "dsine": {
      "onnx": "validated",
      "trt": "production",
      "engine": "engines/l4-sm89/dsine.plan",
      "image_input": [
        1,
        3,
        544,
        544
      ],
      "intrinsics_input": [
        1,
        3,
        3
      ],
      "fallback": "pytorch-native",
      "validation": "dsine_trt_validation.json",
      "precision": "mixed-fp16"
    },
    "flux2-klein-4b": {
      "onnx": "validated",
      "trt": "production",
      "precision": "fp16-static / bf16-pipeline-facade",
      "profile_512": {
        "hidden_states": [
          1,
          1024,
          128
        ],
        "encoder_hidden_states": [
          1,
          512,
          7680
        ],
        "img_ids": [
          1,
          1024,
          4
        ],
        "txt_ids": [
          1,
          512,
          4
        ]
      },
      "engine": "engines/l4-sm89/flux2-klein-4b-512.plan",
      "validation": "flux2_trt_validation.json",
      "resolution": 512,
      "fallback": "pytorch-bf16"
    },
    "anigen-ss-flow-solo": {
      "onnx": "production",
      "opset": 23,
      "trt": "production",
      "onnx_model": "onnx/anigen/ss-flow-solo/model.onnx",
      "engine": "engines/l4-sm89/anigen-ss-flow-solo.plan",
      "precision": "mixed-fp16 / fp32-io",
      "profile": {
        "x": [
          1,
          8,
          16,
          16,
          16
        ],
        "x_skl": [
          1,
          4,
          16,
          16,
          16
        ],
        "cond": [
          1,
          1374,
          1024
        ]
      },
      "validation": "anigen_ss-flow-solo_trt_validation.json",
      "speedup_l4": 1.282
    },
    "anigen-ss-decoder": {
      "onnx": "production",
      "opset": 23,
      "trt": "production",
      "onnx_model": "onnx/anigen/ss-decoder/model.onnx",
      "engine": "engines/l4-sm89/anigen-ss-decoder.plan",
      "precision": "mixed-fp16 / fp32-io",
      "profile": {
        "z": [
          1,
          8,
          16,
          16,
          16
        ],
        "z_skl": [
          1,
          4,
          16,
          16,
          16
        ]
      },
      "validation": "anigen_ss-decoder_trt_validation.json",
      "speedup_l4": 2.017
    },
    "anigen-slat-flow-core": {
      "onnx": "production",
      "opset": 18,
      "trt": "production",
      "onnx_model": "onnx/anigen/slat-flow-core/model.onnx",
      "engine": "engines/l4-sm89/anigen-slat-flow-core.plan",
      "precision": "fp16",
      "dynamic_profile": {
        "n_geo": [
          128,
          4096,
          16384
        ],
        "n_skl": [
          16,
          1024,
          8192
        ]
      },
      "shell_weights": "weights/anigen/slat-flow-shell/shell.safetensors",
      "shell_bytes": 188085736,
      "original_checkpoint_bytes": 2457468473,
      "native_shell": [
        "SparseConv3d",
        "SparseDownsample",
        "SparseUpsample"
      ],
      "validation": "anigen_slat-flow-core_trt_validation.json",
      "speedup_l4": 7.302
    },
    "anigen-skin-decoder": {
      "onnx": "production",
      "opset": 18,
      "trt": "production",
      "onnx_model": "onnx/anigen/skin-decoder/model.onnx",
      "engine": "engines/l4-sm89/anigen-skin-decoder.plan",
      "precision": "fp32-io / TRT optimized",
      "dynamic_profile": {
        "vertices": [
          128,
          25000,
          300000
        ],
        "joints": [
          2,
          24,
          64
        ]
      },
      "validation": "anigen_skin-decoder_trt_validation.json",
      "speedup_l4": 6.805,
      "production_observed_vertices": 148798
    },
    "anigen-slat-decoder-backbone": {
      "status": "production",
      "onnx": "production",
      "trt": "production",
      "backend": "TensorRT 11.2 Python IPluginV3 + spconv Native + FlashAttention",
      "precision": "fp16 branches/upsample internals + fp32 output heads/mesh",
      "onnx_domain": "com.companionforge",
      "plugin_namespace": "companionforge",
      "branches": {
        "geo": "engines/l4-sm89/slat-dae-custom/geo.plan",
        "skin": "engines/l4-sm89/slat-dae-custom/skin.plan",
        "skl": "engines/l4-sm89/slat-dae-custom/skl.plan"
      },
      "heads": {
        "geo": "engines/l4-sm89/slat-dae-custom/geo-head.plan",
        "skin": "engines/l4-sm89/slat-dae-custom/skin-head.plan",
        "skl": "engines/l4-sm89/slat-dae-custom/skl-head.plan"
      },
      "mesh_engine": "engines/l4-sm89/custom-ops/SparseMeshTopologyExtract.plan",
      "dynamic_profile": {
        "branch_tokens_max": 32768,
        "head_tokens_max": 8192,
        "mesh_sparse_cubes_max": 524288
      },
      "custom_ops": [
        "SparseConv3D",
        "SparseWindowAttention",
        "SparseDownsample",
        "SparseUpsample",
        "SparseSubdivide",
        "SparseMeshTopologyExtract"
      ],
      "validation": [
        "bench/slat_dae_custom_validation.json",
        "bench/sparse_mesh_topology_validation.json",
        "bench/full_trt_e2e_validation.json"
      ],
      "native_checkpoint": "lazy-fallback-only",
      "original_decoder_checkpoint_bytes": 809000000
    },
    "rembg-birefnet": {
      "onnx": "upstream-native",
      "runtime": "onnxruntime-gpu-when-fallback-needed",
      "note": "studio references use fast alpha-key before this fallback"
    }
  }
}