manifest.json

{
  "model_id": "zai-org/GLM-OCR",
  "dtype": "float16",
  "opset": 18,
  "static": true,
  "external_data": true,
  "image_size": {
    "width": 840,
    "height": 840
  },
  "max_seq_len": 2048,
  "image_token_id": 59280,
  "eos_token_ids": [
    59246,
    59253
  ],
  "export_prompt": "Recognize the text in the image and output in Markdown format. Preserve the original layout (headings/paragraphs/tables/formulas). Do not fabricate content that does not exist in the image.",
  "hidden_size": 1536,
  "t_img": 900,
  "export_devices": {
    "vision": "cuda",
    "text": "cuda"
  },
  "graphs": {
    "vision": "fp16/glm_ocr_vision.onnx",
    "embed": "fp16/glm_ocr_embed.onnx",
    "rope": "fp16/glm_ocr_rope_document.onnx",
    "decode_prefill_kv": "fp16/glm_ocr_decode_prefill_kv.onnx",
    "decode_step_kv": "fp16/glm_ocr_decode_step_kv.onnx",
    "vision_quant": "quant/glm_ocr_vision_quant.onnx"
  },
  "kv_cache": {
    "num_layers": 16,
    "num_key_value_heads": 8,
    "head_dim": 96,
    "max_cache_len": 2048
  },
  "default_profile": "document",
  "prompt_profiles": {
    "document": {
      "prompt": "Recognize the text in the image and output in Markdown format. Preserve the original layout (headings/paragraphs/tables/formulas). Do not fabricate content that does not exist in the image.",
      "rope": "fp16/glm_ocr_rope_document.onnx"
    },
    "text": {
      "prompt": "Text Recognition:",
      "rope": "fp16/glm_ocr_rope_text.onnx"
    },
    "table": {
      "prompt": "Table Recognition:",
      "rope": "fp16/glm_ocr_rope_table.onnx"
    },
    "formula": {
      "prompt": "Formula Recognition:",
      "rope": "fp16/glm_ocr_rope_formula.onnx"
    }
  },
  "notes": [
    "Vision wrapper handles packed [T,D] outputs by unsqueezing to [1,T,D].",
    "Rope graphs are prompt-profile specific constants generated from get_rope_index (mRoPE).",
    "Do splice in JS: replace contiguous image_token_id block of length t_img with image_embeds.",
    "Decode outputs logits for last token only.",
    "custom_w8: quantized MatMul/Gemm weights for graph 'decode_prefill_kv'.",
    "dual-vision artifact: graphs.vision=fp16 and graphs.vision_quant=quantized."
  ],
  "vision": {
    "onnx": "artifact_glm_ocr_web_split/fp16/glm_ocr_vision.onnx",
    "sha256": "efa009d7c358e4ebf83515b80a96fd828a05e40bd86be190dfabd9fc0f29bc73",
    "bytes": 715251,
    "data": "artifact_glm_ocr_web_split/fp16/glm_ocr_vision.onnx.data",
    "data_sha256": "8691969b71702daabc378e031b66c62d13d054a7cfed17e2729ea707079d2b45",
    "data_bytes": 14571213824
  },
  "embed": {
    "onnx": "artifact_glm_ocr_web_split/fp16/glm_ocr_embed.onnx",
    "sha256": "700497f7747ce77b2b34f519b662d2bee68f44e638de9391301a74011fd2bf20",
    "bytes": 1791,
    "data": "artifact_glm_ocr_web_split/fp16/glm_ocr_embed.onnx.data",
    "data_sha256": "44c2c5e0a2a8d65605f06897249a03fb1d11051aee155ba598e1d1a302ababd0",
    "data_bytes": 364904448
  },
  "rope": {
    "onnx": "artifact_glm_ocr_web_split/fp16/glm_ocr_rope_document.onnx",
    "sha256": "57bbac5220e5828a5ea0bf901974bc44afa190e3e7d9927bceca3ca8a63c4dc3",
    "bytes": 4170,
    "data": "artifact_glm_ocr_web_split/fp16/glm_ocr_rope_document.onnx.data",
    "data_sha256": "cb38103e3aceeb1adeb7104611bd035656dd8433b3c03eb96ac2b5e5df9b55b4",
    "data_bytes": 98304
  },
  "decode": {
    "onnx": "artifact_glm_ocr_web_split/fp16/glm_ocr_decode.onnx",
    "sha256": "6d536a5b1e671e2229bb88a050c747ad6fdc9db3292927af4552c33928a46001",
    "bytes": 2429565,
    "data": "artifact_glm_ocr_web_split/fp16/glm_ocr_decode.onnx.data",
    "data_sha256": "6847fe2bfe59df72975856dc61730b2ecb8a0f79455a898375fc4f7c2c7cb2da",
    "data_bytes": 2328431616
  },
  "decode_prefill_kv": {
    "onnx": "artifact_glm_ocr_web_split/fp16/glm_ocr_decode_prefill_kv.onnx",
    "sha256": "27974f2680f37205f1a51244dbe7fc65e0eefbcf926a3fc9e7a34a5095d4a755",
    "bytes": 2547577,
    "data": "artifact_glm_ocr_web_split/fp16/glm_ocr_decode_prefill_kv.onnx.data",
    "data_sha256": "6847fe2bfe59df72975856dc61730b2ecb8a0f79455a898375fc4f7c2c7cb2da",
    "data_bytes": 2328431616
  },
  "decode_step_kv": {
    "onnx": "artifact_glm_ocr_web_split/fp16/glm_ocr_decode_step_kv.onnx",
    "sha256": "cea69c66eee2ca5e2e1e62185fd50bae0b431735027d107d070a4d8e02f4a318",
    "bytes": 4162536,
    "data": "artifact_glm_ocr_web_split/fp16/glm_ocr_decode_step_kv.onnx.data",
    "data_sha256": "b3a254982459a269a566bdcaa574f9edb4c77fb470c0745e32a36e6ecc1ca927",
    "data_bytes": 2328349696
  },
  "rope_document": {
    "onnx": "artifact_glm_ocr_web_split/fp16/glm_ocr_rope_document.onnx",
    "sha256": "57bbac5220e5828a5ea0bf901974bc44afa190e3e7d9927bceca3ca8a63c4dc3",
    "bytes": 4170,
    "data": "artifact_glm_ocr_web_split/fp16/glm_ocr_rope_document.onnx.data",
    "data_sha256": "cb38103e3aceeb1adeb7104611bd035656dd8433b3c03eb96ac2b5e5df9b55b4",
    "data_bytes": 98304
  },
  "rope_text": {
    "onnx": "artifact_glm_ocr_web_split/fp16/glm_ocr_rope_text.onnx",
    "sha256": "b68daba410b29a635de18b135212569db1c28add28fbe7c5b63dcd0c7b3d0b63",
    "bytes": 4166,
    "data": "artifact_glm_ocr_web_split/fp16/glm_ocr_rope_text.onnx.data",
    "data_sha256": "4147cb50328e4bb79a289496b2d0dea536c02898fc62591fad72da0b636a33ee",
    "data_bytes": 98304
  },
  "rope_table": {
    "onnx": "artifact_glm_ocr_web_split/fp16/glm_ocr_rope_table.onnx",
    "sha256": "88faf6507c9baa52f4069b4a6469c5ab692038a538ef3d9df52ad2408648ba22",
    "bytes": 4167,
    "data": "artifact_glm_ocr_web_split/fp16/glm_ocr_rope_table.onnx.data",
    "data_sha256": "4147cb50328e4bb79a289496b2d0dea536c02898fc62591fad72da0b636a33ee",
    "data_bytes": 98304
  },
  "rope_formula": {
    "onnx": "artifact_glm_ocr_web_split/fp16/glm_ocr_rope_formula.onnx",
    "sha256": "de0c135e8dd941ec1a0246384d20c20984eb2bbed8fd08b43830ba40c081f232",
    "bytes": 4169,
    "data": "artifact_glm_ocr_web_split/fp16/glm_ocr_rope_formula.onnx.data",
    "data_sha256": "2929c9761dd6510fa6734a2ded0bd07d7b8f2705072a0542e76b7ccda9c0f713",
    "data_bytes": 98304
  }
}