taurscribe_granite_nar_manifest.json

{
  "format": "taurscribe-granite-nar-onnx-bundle",
  "format_version": 1,
  "source_model": "ibm-granite/granite-speech-4.1-2b-nar",
  "export_dtype": "int4-matmulnbits-weights",
  "fixed_encoder_frames": 800,
  "opset": 20,
  "execution_provider_preference": [
    "DmlExecutionProvider",
    "CPUExecutionProvider"
  ],
  "graphs": {
    "encoder": {
      "file": "encoder.onnx",
      "inputs": [
        "input_features",
        "attention_mask"
      ],
      "outputs": [
        "bpe_logits",
        "hidden_4",
        "hidden_8",
        "hidden_12",
        "hidden_last"
      ],
      "notes": "Fixed-shape app bucket: pad input_features to 800 frames. Trim bpe_logits to ceil(valid_frames / bpe_pooling_window) before CTC collapse."
    },
    "projector": {
      "file": "projector.onnx",
      "inputs": [
        "multilayer_features"
      ],
      "outputs": [
        "audio_embeds"
      ]
    },
    "embed_tokens": {
      "file": "embed_tokens.onnx",
      "inputs": [
        "token_ids"
      ],
      "outputs": [
        "text_embeds"
      ]
    },
    "editor": {
      "file": "editor.onnx",
      "inputs": [
        "inputs_embeds",
        "position_ids"
      ],
      "outputs": [
        "token_ids"
      ],
      "notes": "Non-causal bidirectional editor. Host code should pass one flattened sample sequence at a time."
    }
  },
  "host_pipeline": [
    "Extract Granite log-mel features with the copied preprocessor config and pad/truncate to 800 encoder frames.",
    "Run encoder.onnx.",
    "Trim padded BPE logits using the valid frame count, argmax, unique-consecutive collapse, and remove blank_token_id.",
    "Concatenate hidden_4, hidden_8, hidden_12, and hidden_last, then run projector.onnx.",
    "Trim audio embeddings using floor(valid_frames / projector.downsample_rate).",
    "Insert blank edit slots around the encoder CTC token IDs.",
    "Run embed_tokens.onnx for the slotted token IDs.",
    "Concatenate audio embeddings and text embeddings, create monotonic position_ids, then run editor.onnx.",
    "Select the text segment logits, argmax, unique-consecutive collapse, remove blank_token_id, and decode with tokenizer.json."
  ],
  "validation": {
    "onnx_checker": "run scripts/granite_nar_export.py validate --model-dir <onnx-bundle>",
    "ort_cpu_session_load": "run scripts/granite_nar_export.py validate --model-dir <onnx-bundle>"
  },
  "notes": [
    "Fallback is controlled by ONNX Runtime session creation in Rust, not encoded inside the ONNX files.",
    "INT4 weight-only MatMulNBits bundle with an editor that returns token_ids instead of full vocabulary logits.",
    "Portable bundle: rank-5 attention MatMuls are flattened to rank 3, fixed 800-frame shape chains are baked as constants, and GLU Split nodes are replaced with Slice pairs so the full encoder runs correctly on DirectML.",
    "DirectML is preferred on compatible Windows GPUs; ONNX Runtime CPU remains the portable fallback."
  ],
  "variant": "int4-argmax-dml-static",
  "encoder_dml_safe": true
}