d1-omni-600M on Core AI: fp16 decision graphs (L64-L4096), vision and audio graphs, Swift host, reference fixture

914184e verified about 1 month ago

20.4 kB

{
 "schema": "d1-omni-coreai/1",
 "model": {
  "source": "LiquidAI/d1-omni-600M",
  "revision": "414f8d6438174f5b2133a9c21a478fc42625e308",
  "license": "LFM Open License v1.0",
  "conversion": "coreai-model-zoo conversion/d1_omni (re-authored in plain PyTorch, coreai-torch 0.4.1; the graphs' debug locations stripped with coreai-torch strip_debug_info, conversion/d1_omni/strip_ship.py)"
 },
 "runtime": "Core AI (macOS 27 / iOS 27)",
 "precision": "fp16 (weights and compute; the exceptions are listed in each graph's metadata.json)",
 "graphs": {
  "decision": {
   "buckets": [
    64,
    128,
    256,
    512,
    1024,
    2048,
    4096
   ],
   "bucket_rule": "the smallest L that holds P + len(ids) (host.py bucket_for(positions, ALL_BUCKETS))",
   "inputs": {
    "input_ids": {
     "dtype": "int32",
     "shape": [
      1,
      "L"
     ],
     "values": "0 on [0, P), the row's ids on [P, P+n), 0 (<|pad|>) after"
    },
    "prefix_embeds": {
     "dtype": "float32",
     "shape": [
      1,
      "L",
      1024
     ],
     "values": "the image / audio prefix on [0, P), 0.0 elsewhere (finite)"
    },
    "pad_mask": {
     "dtype": "float32",
     "shape": [
      1,
      "L"
     ],
     "values": "1.0 on [0, P+n), 0.0 after"
    },
    "prefix_mask": {
     "dtype": "float32",
     "shape": [
      1,
      "L"
     ],
     "values": "1.0 on [0, P), 0.0 after"
    },
    "keep_right": {
     "dtype": "float32",
     "shape": [
      1,
      "L"
     ],
     "values": "0.0 at P-1 when P > 0, 1.0 elsewhere"
    },
    "qtype_onehot": {
     "dtype": "float32",
     "shape": [
      1,
      3
     ],
     "order": [
      "choice",
      "score",
      "noul"
     ]
    }
   },
   "outputs": {
    "scores": {
     "dtype": "float32",
     "shape": [
      1,
      "L"
     ],
     "read": "at P + markers (the <|mask|> positions)"
    }
   },
   "max_length": 16384,
   "image_text_length": 896,
   "audio_text_length": 15360,
   "min_text_positions": 64,
   "prefix_hidden": 1024,
   "folders": {
    "macos": {
     "decide-fp16-L64": "macos/decide-fp16-L64",
     "decide-fp16-L128": "macos/decide-fp16-L128",
     "decide-fp16-L256": "macos/decide-fp16-L256",
     "decide-fp16-L512": "macos/decide-fp16-L512",
     "decide-fp16-L1024": "macos/decide-fp16-L1024",
     "decide-fp16-L2048": "macos/decide-fp16-L2048",
     "decide-fp16-L4096": "macos/decide-fp16-L4096"
    },
    "ios": {
     "decide-fp16-L64": "ios/decide-fp16-L64",
     "decide-fp16-L128": "ios/decide-fp16-L128",
     "decide-fp16-L256": "ios/decide-fp16-L256",
     "decide-fp16-L512": "ios/decide-fp16-L512",
     "decide-fp16-L1024": "ios/decide-fp16-L1024",
     "decide-fp16-L2048": "ios/decide-fp16-L2048",
     "decide-fp16-L4096": "ios/decide-fp16-L4096"
    },
    "ios-h19p": {
     "decide-fp16-L64": "ios-h19p/decide-fp16-L64",
     "decide-fp16-L128": "ios-h19p/decide-fp16-L128",
     "decide-fp16-L256": "ios-h19p/decide-fp16-L256",
     "decide-fp16-L512": "ios-h19p/decide-fp16-L512",
     "decide-fp16-L1024": "ios-h19p/decide-fp16-L1024",
     "decide-fp16-L2048": "ios-h19p/decide-fp16-L2048"
    }
   }
  },
  "vision": {
   "role": "d1-omni image prefix: one crop per call",
   "inputs": {
    "pixel_values": {
     "dtype": "float32",
     "shape": [
      1,
      1024,
      768
     ],
     "values": "the crop's 16x16 patches row-major over (ph, pw), each [py][px][c], (x - 127.5) / 127.5; 0.0 past ph*pw"
    },
    "pos_embed": {
     "dtype": "float32",
     "shape": [
      1,
      1024,
      768
     ],
     "values": "position_table resized to (ph, pw), bilinear + antialias (host.position_embeddings / position_embeddings_numpy), row-major; 0.0 past ph*pw"
    },
    "patch_mask": {
     "dtype": "float32",
     "shape": [
      1,
      1024
     ],
     "values": "1.0 on the ph*pw patches, 0.0 after"
    },
    "unshuffle_index": {
     "dtype": "int32",
     "shape": [
      256,
      4
     ],
     "values": "token (i, j) = row i * (pw/2) + j: patches 2i*pw+2j, 2i*pw+2j+1, (2i+1)*pw+2j, (2i+1)*pw+2j+1; 0 past (ph/2)(pw/2) (host.unshuffle_index)"
    }
   },
   "outputs": {
    "prefix": {
     "dtype": "float32",
     "shape": [
      1,
      256,
      1024
     ],
     "read": "rows [0, (ph/2)(pw/2)): the crop's prefix embeddings"
    }
   },
   "max_patches": 1024,
   "max_tokens": 256,
   "patch": 16,
   "crops": "vision.preprocess(): an image whose rounded area exceeds 2 x 512^2 is cut into the 512 px tiles of the closest grid (2 to 10 tiles, row-major) plus a thumbnail, else the thumbnail alone; each crop resized by torchvision (bilinear, antialias) from the RGB image (host.layout / image_crops)",
   "prefix": "each crop's first (ph/2)(pw/2) output rows, crops in order, images in request order; this is the decision graph's prefix_embeds [0, P)",
   "position_table": {
    "file": "position_table.f32",
    "shape": [
     16,
     16,
     768
    ],
    "dtype": "float32",
    "byte_order": "little",
    "layout": "row-major (row, column, channel)",
    "resize": "per crop to (ph, pw): bilinear, align_corners false, antialias (torch's float32 kernel: width pass then height pass, fused multiply-add per tap; host.position_embeddings_numpy)",
    "sha256": "69b3c175e58af01f99eb4aa566e2c238d2e8b24c9a0d505fbcb6d5e3a095bf6f"
   },
   "precision_detail": {
    "weights": "fp16",
    "compute": "fp16 (LayerNorm, the attention softmax and the projector's gelu in fp32)"
   },
   "folders": {
    "macos": {
     "vision-fp16": "macos/vision-fp16"
    },
    "ios": {
     "vision-fp16": "ios/vision-fp16"
    },
    "ios-h19p": {
     "vision-fp16": "ios-h19p/vision-fp16"
    }
   }
  },
  "audio": {
   "buckets_s": [
    5,
    10,
    20,
    30
   ],
   "per_bucket": {
    "5s": {
     "bucket": {
      "seconds": 5,
      "F": 501,
      "F2": 251,
      "F4": 126,
      "T": 63,
      "holds": "clips of n samples with 1 + n // 160 <= 501 after waveform()'s cut and pad (host.audio_bucket_for: the smallest of [5, 10, 20, 30] s that holds it)",
      "max_prefix_rows": 63
     },
     "inputs": {
      "mel": {
       "dtype": "float32",
       "shape": [
        1,
        128,
        501
       ],
       "values": "the clip's normalised log-mel (mel_host: waveform() + MelFrontend), columns t >= frames 0.0, zero-padded to F"
      },
      "mask_f": {
       "dtype": "float32",
       "shape": [
        1,
        501
       ],
       "values": "1.0 on t < l0 = frames (n // 160)"
      },
      "mask_f2": {
       "dtype": "float32",
       "shape": [
        1,
        251
       ],
       "values": "1.0 on t < l1 = (l0 - 1) // 2 + 1"
      },
      "mask_f4": {
       "dtype": "float32",
       "shape": [
        1,
        126
       ],
       "values": "1.0 on t < l2 = (l1 - 1) // 2 + 1"
      },
      "mask_t": {
       "dtype": "float32",
       "shape": [
        1,
        63
       ],
       "values": "1.0 on t < P = l3 = (l2 - 1) // 2 + 1"
      }
     },
     "outputs": {
      "prefix": {
       "dtype": "float32",
       "shape": [
        1,
        63,
        1024
       ],
       "read": "rows [0, P): the clip's prefix embeddings"
      }
     }
    },
    "10s": {
     "bucket": {
      "seconds": 10,
      "F": 1001,
      "F2": 501,
      "F4": 251,
      "T": 126,
      "holds": "clips of n samples with 1 + n // 160 <= 1001 after waveform()'s cut and pad (host.audio_bucket_for: the smallest of [5, 10, 20, 30] s that holds it)",
      "max_prefix_rows": 125
     },
     "inputs": {
      "mel": {
       "dtype": "float32",
       "shape": [
        1,
        128,
        1001
       ],
       "values": "the clip's normalised log-mel (mel_host: waveform() + MelFrontend), columns t >= frames 0.0, zero-padded to F"
      },
      "mask_f": {
       "dtype": "float32",
       "shape": [
        1,
        1001
       ],
       "values": "1.0 on t < l0 = frames (n // 160)"
      },
      "mask_f2": {
       "dtype": "float32",
       "shape": [
        1,
        501
       ],
       "values": "1.0 on t < l1 = (l0 - 1) // 2 + 1"
      },
      "mask_f4": {
       "dtype": "float32",
       "shape": [
        1,
        251
       ],
       "values": "1.0 on t < l2 = (l1 - 1) // 2 + 1"
      },
      "mask_t": {
       "dtype": "float32",
       "shape": [
        1,
        126
       ],
       "values": "1.0 on t < P = l3 = (l2 - 1) // 2 + 1"
      }
     },
     "outputs": {
      "prefix": {
       "dtype": "float32",
       "shape": [
        1,
        126,
        1024
       ],
       "read": "rows [0, P): the clip's prefix embeddings"
      }
     }
    },
    "20s": {
     "bucket": {
      "seconds": 20,
      "F": 2001,
      "F2": 1001,
      "F4": 501,
      "T": 251,
      "holds": "clips of n samples with 1 + n // 160 <= 2001 after waveform()'s cut and pad (host.audio_bucket_for: the smallest of [5, 10, 20, 30] s that holds it)",
      "max_prefix_rows": 250
     },
     "inputs": {
      "mel": {
       "dtype": "float32",
       "shape": [
        1,
        128,
        2001
       ],
       "values": "the clip's normalised log-mel (mel_host: waveform() + MelFrontend), columns t >= frames 0.0, zero-padded to F"
      },
      "mask_f": {
       "dtype": "float32",
       "shape": [
        1,
        2001
       ],
       "values": "1.0 on t < l0 = frames (n // 160)"
      },
      "mask_f2": {
       "dtype": "float32",
       "shape": [
        1,
        1001
       ],
       "values": "1.0 on t < l1 = (l0 - 1) // 2 + 1"
      },
      "mask_f4": {
       "dtype": "float32",
       "shape": [
        1,
        501
       ],
       "values": "1.0 on t < l2 = (l1 - 1) // 2 + 1"
      },
      "mask_t": {
       "dtype": "float32",
       "shape": [
        1,
        251
       ],
       "values": "1.0 on t < P = l3 = (l2 - 1) // 2 + 1"
      }
     },
     "outputs": {
      "prefix": {
       "dtype": "float32",
       "shape": [
        1,
        251,
        1024
       ],
       "read": "rows [0, P): the clip's prefix embeddings"
      }
     }
    },
    "30s": {
     "bucket": {
      "seconds": 30,
      "F": 3001,
      "F2": 1501,
      "F4": 751,
      "T": 376,
      "holds": "clips of n samples with 1 + n // 160 <= 3001 after waveform()'s cut and pad (host.audio_bucket_for: the smallest of [5, 10, 20, 30] s that holds it)",
      "max_prefix_rows": 375
     },
     "inputs": {
      "mel": {
       "dtype": "float32",
       "shape": [
        1,
        128,
        3001
       ],
       "values": "the clip's normalised log-mel (mel_host: waveform() + MelFrontend), columns t >= frames 0.0, zero-padded to F"
      },
      "mask_f": {
       "dtype": "float32",
       "shape": [
        1,
        3001
       ],
       "values": "1.0 on t < l0 = frames (n // 160)"
      },
      "mask_f2": {
       "dtype": "float32",
       "shape": [
        1,
        1501
       ],
       "values": "1.0 on t < l1 = (l0 - 1) // 2 + 1"
      },
      "mask_f4": {
       "dtype": "float32",
       "shape": [
        1,
        751
       ],
       "values": "1.0 on t < l2 = (l1 - 1) // 2 + 1"
      },
      "mask_t": {
       "dtype": "float32",
       "shape": [
        1,
        376
       ],
       "values": "1.0 on t < P = l3 = (l2 - 1) // 2 + 1"
      }
     },
     "outputs": {
      "prefix": {
       "dtype": "float32",
       "shape": [
        1,
        376,
        1024
       ],
       "read": "rows [0, P): the clip's prefix embeddings"
      }
     }
    }
   },
   "role": "d1-omni audio prefix: one clip per call",
   "waveform": "16 kHz mono; int16 / 32768 or float32 as is; cut to 480,000 samples (30 s); zero-padded to 8,000 (0.5 s); no resampling (the publisher's waveform())",
   "mel": {
    "spec": "mel_host.py (module docstring): preemphasis 0.97; center=True constant-padded STFT, n_fft 512, hop 160, Hann(400, periodic=False) centred in 512; power (sqrt(re^2+im^2))^2; Slaney mel 128 (mel_filters_128x257_f32.bin); log(x + 2^-24); per mel row over the frames t < n // 160: (x - mean) / (std(ddof 1) + 1e-5); 0.0 after",
    "python": "mel_host.mel_torch (the publisher's code, bit for bit)",
    "swift_form": "mel_host.mel_numpy (the same steps in NumPy, float64)"
   },
   "filterbank": {
    "file": "mel_filters_128x257_f32.bin",
    "shape": [
     128,
     257
    ],
    "dtype": "float32",
    "byte_order": "little",
    "layout": "row-major (mel row, FFT bin)",
    "made_by": "the publisher's slaney_filterbank() (librosa's Slaney mel, 0 to 8 kHz)",
    "sha256": "bce5ec5f194a5913f6508cee5a85512e7bad2352db8fc28f5c6ff75af8b09137"
   },
   "prefix": "the graph's first P rows (P = host.audio_prefix_length(n)); this is the decision graph's prefix_embeds [0, P) with max_len = min(15360, 16384 - P) and state None -> {}",
   "precision_detail": {
    "weights": "fp16 (every Linear / Conv weight and bias, pos_bias_u / v)",
    "compute": "fp16 (LayerNorm, the attention softmax, the BatchNorm affine, both GELUs and the output in fp32)",
    "constants_fp32": "the relative-position table [1, 2T-1, 512] and the folded BatchNorm scale / shift"
   },
   "folders": {
    "macos": {
     "audio-fp16-5s": "macos/audio-fp16-5s",
     "audio-fp16-10s": "macos/audio-fp16-10s",
     "audio-fp16-20s": "macos/audio-fp16-20s",
     "audio-fp16-30s": "macos/audio-fp16-30s"
    },
    "ios": {
     "audio-fp16-5s": "ios/audio-fp16-5s",
     "audio-fp16-10s": "ios/audio-fp16-10s",
     "audio-fp16-20s": "ios/audio-fp16-20s",
     "audio-fp16-30s": "ios/audio-fp16-30s"
    },
    "ios-h19p": {
     "audio-fp16-5s": "ios-h19p/audio-fp16-5s",
     "audio-fp16-10s": "ios-h19p/audio-fp16-10s",
     "audio-fp16-20s": "ios-h19p/audio-fp16-20s",
     "audio-fp16-30s": "ios-h19p/audio-fp16-30s"
    }
   }
  }
 },
 "token_ids": {
  "<|pad|>": 0,
  "<|startoftext|>": 1,
  "<|im_end|>": 7,
  "<|mask|>": 16,
  "<|reserved_7|>": 17,
  "<|reserved_8|>": 18,
  "<|reserved_9|>": 19,
  "<|reserved_10|>": 20,
  "<|reserved_11|>": 21
 },
 "delimiters": {
  "state": "<|reserved_7|>",
  "q": "<|reserved_8|>",
  "opt": "<|reserved_9|>",
  "opt_end": "<|reserved_10|>",
  "decide": "<|reserved_11|>",
  "marker": "<|mask|>"
 },
 "temperatures": {
  "choice:11+": 1.372515082359314,
  "choice:2": 1.7465145587921143,
  "choice:3-5": 1.3998981714248657,
  "choice:6-10": 1.1751071214675903,
  "noul:2": 1.6663223505020142,
  "score:3-5": 1.7301132678985596,
  "score:6-10": 1.0,
  "choice": 1.0,
  "score": 1.0,
  "noul": 1.0
 },
 "temperature_rule": "text rows: marker logits / temperatures[f'{type}:{2|3-5|6-10|11+}' by option count] if present else temperatures[type]; image and audio rows: no temperature",
 "noul_order": [
  "false",
  "true"
 ],
 "noul_reported": [
  "yes",
  "no"
 ],
 "tokenizer": {
  "folder": "tokenizer/",
  "encode": "tok(escape(s), add_special_tokens=False): no <|startoftext|> from the post-processor; escape() rewrites every <|name|> in caller text as <¦name¦>"
 },
 "host": {
  "python": "coreai-model-zoo conversion/d1_omni/host.py (request_rows, graph_inputs, probabilities_from_logits, image_crops_inputs, audio_inputs)",
  "swift": "coreai-model-zoo apps/D1Omni (library D1Omni, CLI d1omni; a copy in host/Package.swift and host/Sources): text, image and audio requests; the JPEG rule is in host/README.md"
 },
 "platforms": {
  "macos": "the .aimodel; the Mac specializes it at load (the Swift host's JIT and AOT gave the same marker logits bit for bit and the same ms)",
  "ios": "the same .aimodel; every phone specializes it on its first load",
  "ios-h19p": "iPhone 18 Pro (h19p) GPU AOT of the same bundles except decide-fp16-L4096 (its compile exceeds the phone's per-process memory limit at the first call: a phone runs ios/decide-fp16-L4096); another phone generation refuses them at load and takes ios/"
 },
 "staged": {
  "decide-fp16-L64": {
   "bundle": "d1_omni_decide_fp16_L64.aimodel",
   "bundle_bytes": 761631060,
   "main_hash": "e9e43ae73782c138dc785bc7fb14d484319926e95c19d08506b3c83bec357f74",
   "aimodelc_h19p": "d1_omni_decide_fp16_L64.h19p.aimodelc",
   "aimodelc_h19p_bytes": 761747067,
   "aimodelc_h19p_main_hash": "eff8f9efff9098fbaebd4ec088d08430f9b5b1b040d1d82a2c0edce25fe7b88f"
  },
  "decide-fp16-L128": {
   "bundle": "d1_omni_decide_fp16_L128.aimodel",
   "bundle_bytes": 761647561,
   "main_hash": "aa6af50cfa06c729f4a618ec0cb6d7a85515ed7f32159e81e5e49991417bf79f",
   "aimodelc_h19p": "d1_omni_decide_fp16_L128.h19p.aimodelc",
   "aimodelc_h19p_bytes": 761763726,
   "aimodelc_h19p_main_hash": "fe544a478d29e743b5e615a18b4d4f98925ba89905dda872efe0e341fe68fd26"
  },
  "decide-fp16-L256": {
   "bundle": "d1_omni_decide_fp16_L256.aimodel",
   "bundle_bytes": 761680330,
   "main_hash": "eb4b4f9dd21f8454b49d7433b31e18b2a00c39215f5049c7137ff53306ee943b",
   "aimodelc_h19p": "d1_omni_decide_fp16_L256.h19p.aimodelc",
   "aimodelc_h19p_bytes": 761796331,
   "aimodelc_h19p_main_hash": "7ba3b11ff7d253b1f97ecca8c177d18f953b9e844ff9eabda6b28454b71de937"
  },
  "decide-fp16-L512": {
   "bundle": "d1_omni_decide_fp16_L512.aimodel",
   "bundle_bytes": 761745868,
   "main_hash": "4f73c0b4742272e6e8c57a0368ca14521ae7845b3e5d469e697aa5e4027e48b7",
   "aimodelc_h19p": "d1_omni_decide_fp16_L512.h19p.aimodelc",
   "aimodelc_h19p_bytes": 761861994,
   "aimodelc_h19p_main_hash": "1d54cd013075829da188bc5be1ed7ee1f3f071371aa391f2bc61738c9f37558e"
  },
  "decide-fp16-L1024": {
   "bundle": "d1_omni_decide_fp16_L1024.aimodel",
   "bundle_bytes": 761876787,
   "main_hash": "92bf9613ca22138709d741196796fcde9302f01b66cfade2e5c9b6237ef35c73",
   "aimodelc_h19p": "d1_omni_decide_fp16_L1024.h19p.aimodelc",
   "aimodelc_h19p_bytes": 761993099,
   "aimodelc_h19p_main_hash": "c21863c40b02ef7c0e3e7d87297016864d209ab306bc16bc551c7841202b1cd7"
  },
  "decide-fp16-L2048": {
   "bundle": "d1_omni_decide_fp16_L2048.aimodel",
   "bundle_bytes": 762139098,
   "main_hash": "51b6cbe1b3c13267c544d7f596bdc265e7214b68d6386a03dd1f31c45d1a4ee5",
   "aimodelc_h19p": "d1_omni_decide_fp16_L2048.h19p.aimodelc",
   "aimodelc_h19p_bytes": 762255228,
   "aimodelc_h19p_main_hash": "117d089252f05d23fdcdbaebcead8da2fc4bf1f69da18f6f873ee000c487332d"
  },
  "decide-fp16-L4096": {
   "bundle": "d1_omni_decide_fp16_L4096.aimodel",
   "bundle_bytes": 762676555,
   "main_hash": "4c38676fe90011b05fc79e6c5cca04fe2eb0f9b72c92e71aec64e1d3ffa664a5"
  },
  "vision-fp16": {
   "bundle": "d1_omni_vision_fp16.aimodel",
   "bundle_bytes": 188218374,
   "main_hash": "d65cf00787bab8bd5ee390cc8820ab600ba44cfe5a72ee43574a0ffd3d3d148c",
   "aimodelc_h19p": "d1_omni_vision_fp16.h19p.aimodelc",
   "aimodelc_h19p_bytes": 188303983,
   "aimodelc_h19p_main_hash": "fc3e1278f91c97cdf9419aa154b5041e7f6fb82525d245c4135aab972d83bef1"
  },
  "audio-fp16-5s": {
   "bundle": "d1_omni_audio_fp16_5s.aimodel",
   "bundle_bytes": 224920212,
   "main_hash": "b3240c83a181d87adda6fa97fe52661ab1fa6d82b6c62b651d5481cfe5138685",
   "aimodelc_h19p": "d1_omni_audio_fp16_5s.h19p.aimodelc",
   "aimodelc_h19p_bytes": 225226318,
   "aimodelc_h19p_main_hash": "09a5141b87f4abc2be1169bbda2528ad492cfb9681ad9bc8761d237c9ad7839c"
  },
  "audio-fp16-10s": {
   "bundle": "d1_omni_audio_fp16_10s.aimodel",
   "bundle_bytes": 225049263,
   "main_hash": "0efcce362b71f05bb90eb11eb22b087521e990135f2bfdd2efd72a6b213eaa01",
   "aimodelc_h19p": "d1_omni_audio_fp16_10s.h19p.aimodelc",
   "aimodelc_h19p_bytes": 225355344,
   "aimodelc_h19p_main_hash": "0a66cbd973f93c56bcc086df520399d97a05c0857d8aab73c3e58e59209a12ae"
  },
  "audio-fp16-20s": {
   "bundle": "d1_omni_audio_fp16_20s.aimodel",
   "bundle_bytes": 225305277,
   "main_hash": "9de9ec76dc8c6abb22688ce7e842cb056cee0ca15ffe6cb3a66ecdfe0dab515e",
   "aimodelc_h19p": "d1_omni_audio_fp16_20s.h19p.aimodelc",
   "aimodelc_h19p_bytes": 225611159,
   "aimodelc_h19p_main_hash": "0039312e4d292a6c00f18480a8dc90c66743762546123ffa842f166e1d382e5c"
  },
  "audio-fp16-30s": {
   "bundle": "d1_omni_audio_fp16_30s.aimodel",
   "bundle_bytes": 225561282,
   "main_hash": "6bb3911322160cd29ec93e72c5d78e7c9ab07b3a1073552fe2fda053bee61c65",
   "aimodelc_h19p": "d1_omni_audio_fp16_30s.h19p.aimodelc",
   "aimodelc_h19p_bytes": 225867256,
   "aimodelc_h19p_main_hash": "1354d0680107fc4b551fe93c35ab5806dc6c882bbe5c32f3f295f1b6d1e798ab"
  }
 },
 "hash_note": "main_hash = the .aimodel's main.hash, sha256 of its main.mlirb (the name of the compile-cache entry a JIT load creates); aimodelc_h19p_main_hash = the AOT's main.hash"
}

	{
	"schema": "d1-omni-coreai/1",
	"model": {
	"source": "LiquidAI/d1-omni-600M",
	"revision": "414f8d6438174f5b2133a9c21a478fc42625e308",
	"license": "LFM Open License v1.0",
	"conversion": "coreai-model-zoo conversion/d1_omni (re-authored in plain PyTorch, coreai-torch 0.4.1; the graphs' debug locations stripped with coreai-torch strip_debug_info, conversion/d1_omni/strip_ship.py)"
	},
	"runtime": "Core AI (macOS 27 / iOS 27)",
	"precision": "fp16 (weights and compute; the exceptions are listed in each graph's metadata.json)",
	"graphs": {
	"decision": {
	"buckets": [
	64,
	128,
	256,
	512,
	1024,
	2048,
	4096
	],
	"bucket_rule": "the smallest L that holds P + len(ids) (host.py bucket_for(positions, ALL_BUCKETS))",
	"inputs": {
	"input_ids": {
	"dtype": "int32",
	"shape": [
	1,
	"L"
	],
	"values": "0 on [0, P), the row's ids on [P, P+n), 0 (<\|pad\|>) after"
	},
	"prefix_embeds": {
	"dtype": "float32",
	"shape": [
	1,
	"L",
	1024
	],
	"values": "the image / audio prefix on [0, P), 0.0 elsewhere (finite)"
	},
	"pad_mask": {
	"dtype": "float32",
	"shape": [
	1,
	"L"
	],
	"values": "1.0 on [0, P+n), 0.0 after"
	},
	"prefix_mask": {
	"dtype": "float32",
	"shape": [
	1,
	"L"
	],
	"values": "1.0 on [0, P), 0.0 after"
	},
	"keep_right": {
	"dtype": "float32",
	"shape": [
	1,
	"L"
	],
	"values": "0.0 at P-1 when P > 0, 1.0 elsewhere"
	},
	"qtype_onehot": {
	"dtype": "float32",
	"shape": [
	1,
	3
	],
	"order": [
	"choice",
	"score",
	"noul"
	]
	}
	},
	"outputs": {
	"scores": {
	"dtype": "float32",
	"shape": [
	1,
	"L"
	],
	"read": "at P + markers (the <\|mask\|> positions)"
	}
	},
	"max_length": 16384,
	"image_text_length": 896,
	"audio_text_length": 15360,
	"min_text_positions": 64,
	"prefix_hidden": 1024,
	"folders": {
	"macos": {
	"decide-fp16-L64": "macos/decide-fp16-L64",
	"decide-fp16-L128": "macos/decide-fp16-L128",
	"decide-fp16-L256": "macos/decide-fp16-L256",
	"decide-fp16-L512": "macos/decide-fp16-L512",
	"decide-fp16-L1024": "macos/decide-fp16-L1024",
	"decide-fp16-L2048": "macos/decide-fp16-L2048",
	"decide-fp16-L4096": "macos/decide-fp16-L4096"
	},
	"ios": {
	"decide-fp16-L64": "ios/decide-fp16-L64",
	"decide-fp16-L128": "ios/decide-fp16-L128",
	"decide-fp16-L256": "ios/decide-fp16-L256",
	"decide-fp16-L512": "ios/decide-fp16-L512",
	"decide-fp16-L1024": "ios/decide-fp16-L1024",
	"decide-fp16-L2048": "ios/decide-fp16-L2048",
	"decide-fp16-L4096": "ios/decide-fp16-L4096"
	},
	"ios-h19p": {
	"decide-fp16-L64": "ios-h19p/decide-fp16-L64",
	"decide-fp16-L128": "ios-h19p/decide-fp16-L128",
	"decide-fp16-L256": "ios-h19p/decide-fp16-L256",
	"decide-fp16-L512": "ios-h19p/decide-fp16-L512",
	"decide-fp16-L1024": "ios-h19p/decide-fp16-L1024",
	"decide-fp16-L2048": "ios-h19p/decide-fp16-L2048"
	}
	}
	},
	"vision": {
	"role": "d1-omni image prefix: one crop per call",
	"inputs": {
	"pixel_values": {
	"dtype": "float32",
	"shape": [
	1,
	1024,
	768
	],
	"values": "the crop's 16x16 patches row-major over (ph, pw), each [py][px][c], (x - 127.5) / 127.5; 0.0 past ph*pw"
	},
	"pos_embed": {
	"dtype": "float32",
	"shape": [
	1,
	1024,
	768
	],
	"values": "position_table resized to (ph, pw), bilinear + antialias (host.position_embeddings / position_embeddings_numpy), row-major; 0.0 past ph*pw"
	},
	"patch_mask": {
	"dtype": "float32",
	"shape": [
	1,
	1024
	],
	"values": "1.0 on the ph*pw patches, 0.0 after"
	},
	"unshuffle_index": {
	"dtype": "int32",
	"shape": [
	256,
	4
	],
	"values": "token (i, j) = row i * (pw/2) + j: patches 2ipw+2j, 2ipw+2j+1, (2i+1)pw+2j, (2i+1)pw+2j+1; 0 past (ph/2)(pw/2) (host.unshuffle_index)"
	}
	},
	"outputs": {
	"prefix": {
	"dtype": "float32",
	"shape": [
	1,
	256,
	1024
	],
	"read": "rows [0, (ph/2)(pw/2)): the crop's prefix embeddings"
	}
	},
	"max_patches": 1024,
	"max_tokens": 256,
	"patch": 16,
	"crops": "vision.preprocess(): an image whose rounded area exceeds 2 x 512^2 is cut into the 512 px tiles of the closest grid (2 to 10 tiles, row-major) plus a thumbnail, else the thumbnail alone; each crop resized by torchvision (bilinear, antialias) from the RGB image (host.layout / image_crops)",
	"prefix": "each crop's first (ph/2)(pw/2) output rows, crops in order, images in request order; this is the decision graph's prefix_embeds [0, P)",
	"position_table": {
	"file": "position_table.f32",
	"shape": [
	16,
	16,
	768
	],
	"dtype": "float32",
	"byte_order": "little",
	"layout": "row-major (row, column, channel)",
	"resize": "per crop to (ph, pw): bilinear, align_corners false, antialias (torch's float32 kernel: width pass then height pass, fused multiply-add per tap; host.position_embeddings_numpy)",
	"sha256": "69b3c175e58af01f99eb4aa566e2c238d2e8b24c9a0d505fbcb6d5e3a095bf6f"
	},
	"precision_detail": {
	"weights": "fp16",
	"compute": "fp16 (LayerNorm, the attention softmax and the projector's gelu in fp32)"
	},
	"folders": {
	"macos": {
	"vision-fp16": "macos/vision-fp16"
	},
	"ios": {
	"vision-fp16": "ios/vision-fp16"
	},
	"ios-h19p": {
	"vision-fp16": "ios-h19p/vision-fp16"
	}
	}
	},
	"audio": {
	"buckets_s": [
	5,
	10,
	20,
	30
	],
	"per_bucket": {
	"5s": {
	"bucket": {
	"seconds": 5,
	"F": 501,
	"F2": 251,
	"F4": 126,
	"T": 63,
	"holds": "clips of n samples with 1 + n // 160 <= 501 after waveform()'s cut and pad (host.audio_bucket_for: the smallest of [5, 10, 20, 30] s that holds it)",
	"max_prefix_rows": 63
	},
	"inputs": {
	"mel": {
	"dtype": "float32",
	"shape": [
	1,
	128,
	501
	],
	"values": "the clip's normalised log-mel (mel_host: waveform() + MelFrontend), columns t >= frames 0.0, zero-padded to F"
	},
	"mask_f": {
	"dtype": "float32",
	"shape": [
	1,
	501
	],
	"values": "1.0 on t < l0 = frames (n // 160)"
	},
	"mask_f2": {
	"dtype": "float32",
	"shape": [
	1,
	251
	],
	"values": "1.0 on t < l1 = (l0 - 1) // 2 + 1"
	},
	"mask_f4": {
	"dtype": "float32",
	"shape": [
	1,
	126
	],
	"values": "1.0 on t < l2 = (l1 - 1) // 2 + 1"
	},
	"mask_t": {
	"dtype": "float32",
	"shape": [
	1,
	63
	],
	"values": "1.0 on t < P = l3 = (l2 - 1) // 2 + 1"
	}
	},
	"outputs": {
	"prefix": {
	"dtype": "float32",
	"shape": [
	1,
	63,
	1024
	],
	"read": "rows [0, P): the clip's prefix embeddings"
	}
	}
	},
	"10s": {
	"bucket": {
	"seconds": 10,
	"F": 1001,
	"F2": 501,
	"F4": 251,
	"T": 126,
	"holds": "clips of n samples with 1 + n // 160 <= 1001 after waveform()'s cut and pad (host.audio_bucket_for: the smallest of [5, 10, 20, 30] s that holds it)",
	"max_prefix_rows": 125
	},
	"inputs": {
	"mel": {
	"dtype": "float32",
	"shape": [
	1,
	128,
	1001
	],
	"values": "the clip's normalised log-mel (mel_host: waveform() + MelFrontend), columns t >= frames 0.0, zero-padded to F"
	},
	"mask_f": {
	"dtype": "float32",
	"shape": [
	1,
	1001
	],
	"values": "1.0 on t < l0 = frames (n // 160)"
	},
	"mask_f2": {
	"dtype": "float32",
	"shape": [
	1,
	501
	],
	"values": "1.0 on t < l1 = (l0 - 1) // 2 + 1"
	},
	"mask_f4": {
	"dtype": "float32",
	"shape": [
	1,
	251
	],
	"values": "1.0 on t < l2 = (l1 - 1) // 2 + 1"
	},
	"mask_t": {
	"dtype": "float32",
	"shape": [
	1,
	126
	],
	"values": "1.0 on t < P = l3 = (l2 - 1) // 2 + 1"
	}
	},
	"outputs": {
	"prefix": {
	"dtype": "float32",
	"shape": [
	1,
	126,
	1024
	],
	"read": "rows [0, P): the clip's prefix embeddings"
	}
	}
	},
	"20s": {
	"bucket": {
	"seconds": 20,
	"F": 2001,
	"F2": 1001,
	"F4": 501,
	"T": 251,
	"holds": "clips of n samples with 1 + n // 160 <= 2001 after waveform()'s cut and pad (host.audio_bucket_for: the smallest of [5, 10, 20, 30] s that holds it)",
	"max_prefix_rows": 250
	},
	"inputs": {
	"mel": {
	"dtype": "float32",
	"shape": [
	1,
	128,
	2001
	],
	"values": "the clip's normalised log-mel (mel_host: waveform() + MelFrontend), columns t >= frames 0.0, zero-padded to F"
	},
	"mask_f": {
	"dtype": "float32",
	"shape": [
	1,
	2001
	],
	"values": "1.0 on t < l0 = frames (n // 160)"
	},
	"mask_f2": {
	"dtype": "float32",
	"shape": [
	1,
	1001
	],
	"values": "1.0 on t < l1 = (l0 - 1) // 2 + 1"
	},
	"mask_f4": {
	"dtype": "float32",
	"shape": [
	1,
	501
	],
	"values": "1.0 on t < l2 = (l1 - 1) // 2 + 1"
	},
	"mask_t": {
	"dtype": "float32",
	"shape": [
	1,
	251
	],
	"values": "1.0 on t < P = l3 = (l2 - 1) // 2 + 1"
	}
	},
	"outputs": {
	"prefix": {
	"dtype": "float32",
	"shape": [
	1,
	251,
	1024
	],
	"read": "rows [0, P): the clip's prefix embeddings"
	}
	}
	},
	"30s": {
	"bucket": {
	"seconds": 30,
	"F": 3001,
	"F2": 1501,
	"F4": 751,
	"T": 376,
	"holds": "clips of n samples with 1 + n // 160 <= 3001 after waveform()'s cut and pad (host.audio_bucket_for: the smallest of [5, 10, 20, 30] s that holds it)",
	"max_prefix_rows": 375
	},
	"inputs": {
	"mel": {
	"dtype": "float32",
	"shape": [
	1,
	128,
	3001
	],
	"values": "the clip's normalised log-mel (mel_host: waveform() + MelFrontend), columns t >= frames 0.0, zero-padded to F"
	},
	"mask_f": {
	"dtype": "float32",
	"shape": [
	1,
	3001
	],
	"values": "1.0 on t < l0 = frames (n // 160)"
	},
	"mask_f2": {
	"dtype": "float32",
	"shape": [
	1,
	1501
	],
	"values": "1.0 on t < l1 = (l0 - 1) // 2 + 1"
	},
	"mask_f4": {
	"dtype": "float32",
	"shape": [
	1,
	751
	],
	"values": "1.0 on t < l2 = (l1 - 1) // 2 + 1"
	},
	"mask_t": {
	"dtype": "float32",
	"shape": [
	1,
	376
	],
	"values": "1.0 on t < P = l3 = (l2 - 1) // 2 + 1"
	}
	},
	"outputs": {
	"prefix": {
	"dtype": "float32",
	"shape": [
	1,
	376,
	1024
	],
	"read": "rows [0, P): the clip's prefix embeddings"
	}
	}
	}
	},
	"role": "d1-omni audio prefix: one clip per call",
	"waveform": "16 kHz mono; int16 / 32768 or float32 as is; cut to 480,000 samples (30 s); zero-padded to 8,000 (0.5 s); no resampling (the publisher's waveform())",
	"mel": {
	"spec": "mel_host.py (module docstring): preemphasis 0.97; center=True constant-padded STFT, n_fft 512, hop 160, Hann(400, periodic=False) centred in 512; power (sqrt(re^2+im^2))^2; Slaney mel 128 (mel_filters_128x257_f32.bin); log(x + 2^-24); per mel row over the frames t < n // 160: (x - mean) / (std(ddof 1) + 1e-5); 0.0 after",
	"python": "mel_host.mel_torch (the publisher's code, bit for bit)",
	"swift_form": "mel_host.mel_numpy (the same steps in NumPy, float64)"
	},
	"filterbank": {
	"file": "mel_filters_128x257_f32.bin",
	"shape": [
	128,
	257
	],
	"dtype": "float32",
	"byte_order": "little",
	"layout": "row-major (mel row, FFT bin)",
	"made_by": "the publisher's slaney_filterbank() (librosa's Slaney mel, 0 to 8 kHz)",
	"sha256": "bce5ec5f194a5913f6508cee5a85512e7bad2352db8fc28f5c6ff75af8b09137"
	},
	"prefix": "the graph's first P rows (P = host.audio_prefix_length(n)); this is the decision graph's prefix_embeds [0, P) with max_len = min(15360, 16384 - P) and state None -> {}",
	"precision_detail": {
	"weights": "fp16 (every Linear / Conv weight and bias, pos_bias_u / v)",
	"compute": "fp16 (LayerNorm, the attention softmax, the BatchNorm affine, both GELUs and the output in fp32)",
	"constants_fp32": "the relative-position table [1, 2T-1, 512] and the folded BatchNorm scale / shift"
	},
	"folders": {
	"macos": {
	"audio-fp16-5s": "macos/audio-fp16-5s",
	"audio-fp16-10s": "macos/audio-fp16-10s",
	"audio-fp16-20s": "macos/audio-fp16-20s",
	"audio-fp16-30s": "macos/audio-fp16-30s"
	},
	"ios": {
	"audio-fp16-5s": "ios/audio-fp16-5s",
	"audio-fp16-10s": "ios/audio-fp16-10s",
	"audio-fp16-20s": "ios/audio-fp16-20s",
	"audio-fp16-30s": "ios/audio-fp16-30s"
	},
	"ios-h19p": {
	"audio-fp16-5s": "ios-h19p/audio-fp16-5s",
	"audio-fp16-10s": "ios-h19p/audio-fp16-10s",
	"audio-fp16-20s": "ios-h19p/audio-fp16-20s",
	"audio-fp16-30s": "ios-h19p/audio-fp16-30s"
	}
	}
	}
	},
	"token_ids": {
	"<\|pad\|>": 0,
	"<\|startoftext\|>": 1,
	"<\|im_end\|>": 7,
	"<\|mask\|>": 16,
	"<\|reserved_7\|>": 17,
	"<\|reserved_8\|>": 18,
	"<\|reserved_9\|>": 19,
	"<\|reserved_10\|>": 20,
	"<\|reserved_11\|>": 21
	},
	"delimiters": {
	"state": "<\|reserved_7\|>",
	"q": "<\|reserved_8\|>",
	"opt": "<\|reserved_9\|>",
	"opt_end": "<\|reserved_10\|>",
	"decide": "<\|reserved_11\|>",
	"marker": "<\|mask\|>"
	},
	"temperatures": {
	"choice:11+": 1.372515082359314,
	"choice:2": 1.7465145587921143,
	"choice:3-5": 1.3998981714248657,
	"choice:6-10": 1.1751071214675903,
	"noul:2": 1.6663223505020142,
	"score:3-5": 1.7301132678985596,
	"score:6-10": 1.0,
	"choice": 1.0,
	"score": 1.0,
	"noul": 1.0
	},
	"temperature_rule": "text rows: marker logits / temperatures[f'{type}:{2\|3-5\|6-10\|11+}' by option count] if present else temperatures[type]; image and audio rows: no temperature",
	"noul_order": [
	"false",
	"true"
	],
	"noul_reported": [
	"yes",
	"no"
	],
	"tokenizer": {
	"folder": "tokenizer/",
	"encode": "tok(escape(s), add_special_tokens=False): no <\|startoftext\|> from the post-processor; escape() rewrites every <\|name\|> in caller text as <¦name¦>"
	},
	"host": {
	"python": "coreai-model-zoo conversion/d1_omni/host.py (request_rows, graph_inputs, probabilities_from_logits, image_crops_inputs, audio_inputs)",
	"swift": "coreai-model-zoo apps/D1Omni (library D1Omni, CLI d1omni; a copy in host/Package.swift and host/Sources): text, image and audio requests; the JPEG rule is in host/README.md"
	},
	"platforms": {
	"macos": "the .aimodel; the Mac specializes it at load (the Swift host's JIT and AOT gave the same marker logits bit for bit and the same ms)",
	"ios": "the same .aimodel; every phone specializes it on its first load",
	"ios-h19p": "iPhone 18 Pro (h19p) GPU AOT of the same bundles except decide-fp16-L4096 (its compile exceeds the phone's per-process memory limit at the first call: a phone runs ios/decide-fp16-L4096); another phone generation refuses them at load and takes ios/"
	},
	"staged": {
	"decide-fp16-L64": {
	"bundle": "d1_omni_decide_fp16_L64.aimodel",
	"bundle_bytes": 761631060,
	"main_hash": "e9e43ae73782c138dc785bc7fb14d484319926e95c19d08506b3c83bec357f74",
	"aimodelc_h19p": "d1_omni_decide_fp16_L64.h19p.aimodelc",
	"aimodelc_h19p_bytes": 761747067,
	"aimodelc_h19p_main_hash": "eff8f9efff9098fbaebd4ec088d08430f9b5b1b040d1d82a2c0edce25fe7b88f"
	},
	"decide-fp16-L128": {
	"bundle": "d1_omni_decide_fp16_L128.aimodel",
	"bundle_bytes": 761647561,
	"main_hash": "aa6af50cfa06c729f4a618ec0cb6d7a85515ed7f32159e81e5e49991417bf79f",
	"aimodelc_h19p": "d1_omni_decide_fp16_L128.h19p.aimodelc",
	"aimodelc_h19p_bytes": 761763726,
	"aimodelc_h19p_main_hash": "fe544a478d29e743b5e615a18b4d4f98925ba89905dda872efe0e341fe68fd26"
	},
	"decide-fp16-L256": {
	"bundle": "d1_omni_decide_fp16_L256.aimodel",
	"bundle_bytes": 761680330,
	"main_hash": "eb4b4f9dd21f8454b49d7433b31e18b2a00c39215f5049c7137ff53306ee943b",
	"aimodelc_h19p": "d1_omni_decide_fp16_L256.h19p.aimodelc",
	"aimodelc_h19p_bytes": 761796331,
	"aimodelc_h19p_main_hash": "7ba3b11ff7d253b1f97ecca8c177d18f953b9e844ff9eabda6b28454b71de937"
	},
	"decide-fp16-L512": {
	"bundle": "d1_omni_decide_fp16_L512.aimodel",
	"bundle_bytes": 761745868,
	"main_hash": "4f73c0b4742272e6e8c57a0368ca14521ae7845b3e5d469e697aa5e4027e48b7",
	"aimodelc_h19p": "d1_omni_decide_fp16_L512.h19p.aimodelc",
	"aimodelc_h19p_bytes": 761861994,
	"aimodelc_h19p_main_hash": "1d54cd013075829da188bc5be1ed7ee1f3f071371aa391f2bc61738c9f37558e"
	},
	"decide-fp16-L1024": {
	"bundle": "d1_omni_decide_fp16_L1024.aimodel",
	"bundle_bytes": 761876787,
	"main_hash": "92bf9613ca22138709d741196796fcde9302f01b66cfade2e5c9b6237ef35c73",
	"aimodelc_h19p": "d1_omni_decide_fp16_L1024.h19p.aimodelc",
	"aimodelc_h19p_bytes": 761993099,
	"aimodelc_h19p_main_hash": "c21863c40b02ef7c0e3e7d87297016864d209ab306bc16bc551c7841202b1cd7"
	},
	"decide-fp16-L2048": {
	"bundle": "d1_omni_decide_fp16_L2048.aimodel",
	"bundle_bytes": 762139098,
	"main_hash": "51b6cbe1b3c13267c544d7f596bdc265e7214b68d6386a03dd1f31c45d1a4ee5",
	"aimodelc_h19p": "d1_omni_decide_fp16_L2048.h19p.aimodelc",
	"aimodelc_h19p_bytes": 762255228,
	"aimodelc_h19p_main_hash": "117d089252f05d23fdcdbaebcead8da2fc4bf1f69da18f6f873ee000c487332d"
	},
	"decide-fp16-L4096": {
	"bundle": "d1_omni_decide_fp16_L4096.aimodel",
	"bundle_bytes": 762676555,
	"main_hash": "4c38676fe90011b05fc79e6c5cca04fe2eb0f9b72c92e71aec64e1d3ffa664a5"
	},
	"vision-fp16": {
	"bundle": "d1_omni_vision_fp16.aimodel",
	"bundle_bytes": 188218374,
	"main_hash": "d65cf00787bab8bd5ee390cc8820ab600ba44cfe5a72ee43574a0ffd3d3d148c",
	"aimodelc_h19p": "d1_omni_vision_fp16.h19p.aimodelc",
	"aimodelc_h19p_bytes": 188303983,
	"aimodelc_h19p_main_hash": "fc3e1278f91c97cdf9419aa154b5041e7f6fb82525d245c4135aab972d83bef1"
	},
	"audio-fp16-5s": {
	"bundle": "d1_omni_audio_fp16_5s.aimodel",
	"bundle_bytes": 224920212,
	"main_hash": "b3240c83a181d87adda6fa97fe52661ab1fa6d82b6c62b651d5481cfe5138685",
	"aimodelc_h19p": "d1_omni_audio_fp16_5s.h19p.aimodelc",
	"aimodelc_h19p_bytes": 225226318,
	"aimodelc_h19p_main_hash": "09a5141b87f4abc2be1169bbda2528ad492cfb9681ad9bc8761d237c9ad7839c"
	},
	"audio-fp16-10s": {
	"bundle": "d1_omni_audio_fp16_10s.aimodel",
	"bundle_bytes": 225049263,
	"main_hash": "0efcce362b71f05bb90eb11eb22b087521e990135f2bfdd2efd72a6b213eaa01",
	"aimodelc_h19p": "d1_omni_audio_fp16_10s.h19p.aimodelc",
	"aimodelc_h19p_bytes": 225355344,
	"aimodelc_h19p_main_hash": "0a66cbd973f93c56bcc086df520399d97a05c0857d8aab73c3e58e59209a12ae"
	},
	"audio-fp16-20s": {
	"bundle": "d1_omni_audio_fp16_20s.aimodel",
	"bundle_bytes": 225305277,
	"main_hash": "9de9ec76dc8c6abb22688ce7e842cb056cee0ca15ffe6cb3a66ecdfe0dab515e",
	"aimodelc_h19p": "d1_omni_audio_fp16_20s.h19p.aimodelc",
	"aimodelc_h19p_bytes": 225611159,
	"aimodelc_h19p_main_hash": "0039312e4d292a6c00f18480a8dc90c66743762546123ffa842f166e1d382e5c"
	},
	"audio-fp16-30s": {
	"bundle": "d1_omni_audio_fp16_30s.aimodel",
	"bundle_bytes": 225561282,
	"main_hash": "6bb3911322160cd29ec93e72c5d78e7c9ab07b3a1073552fe2fda053bee61c65",
	"aimodelc_h19p": "d1_omni_audio_fp16_30s.h19p.aimodelc",
	"aimodelc_h19p_bytes": 225867256,
	"aimodelc_h19p_main_hash": "1354d0680107fc4b551fe93c35ab5806dc6c882bbe5c32f3f295f1b6d1e798ab"
	}
	},
	"hash_note": "main_hash = the .aimodel's main.hash, sha256 of its main.mlirb (the name of the compile-cache entry a JIT load creates); aimodelc_h19p_main_hash = the AOT's main.hash"
	}