{
  "schema": "qfs.upstream-code-audit.v1",
  "upstream": {
    "repository": "moonshotai/Kimi-K3",
    "revision": "f831ab66814297da540d832a5235f8e904f29d06"
  },
  "executed_during_audit": false,
  "weights_downloaded": false,
  "files": [
    {
      "file": "configuration_kimi_k3.py",
      "sha256": "735eb9ebe593e17d231e08e1df7f7be9b5ee0e079f511aa201f9572077b416ae",
      "bytes": 11343,
      "imports": [
        {
          "line": 1,
          "statement": "from typing import Optional"
        },
        {
          "line": 3,
          "statement": "from transformers.configuration_utils import PretrainedConfig"
        }
      ],
      "effect_related_calls": [],
      "top_level_nodes": []
    },
    {
      "file": "encoding_k3.py",
      "sha256": "49ff03305fdc4be26867972788d36150b67f8a9e852e62bb7959d87482223676",
      "bytes": 26304,
      "imports": [
        {
          "line": 9,
          "statement": "from __future__ import annotations"
        },
        {
          "line": 11,
          "statement": "import json"
        },
        {
          "line": 12,
          "statement": "from dataclasses import dataclass"
        },
        {
          "line": 13,
          "statement": "from typing import Any, Iterable, Optional"
        }
      ],
      "effect_related_calls": [
        {
          "line": 498,
          "call": "_open_tag"
        },
        {
          "line": 526,
          "call": "_open_tag"
        },
        {
          "line": 577,
          "call": "_open_tag"
        },
        {
          "line": 519,
          "call": "_open_tag"
        },
        {
          "line": 532,
          "call": "_open_tag"
        },
        {
          "line": 626,
          "call": "_internal_system_message"
        },
        {
          "line": 706,
          "call": "_internal_system_message"
        },
        {
          "line": 725,
          "call": "_internal_system_message"
        },
        {
          "line": 746,
          "call": "_open_tag"
        },
        {
          "line": 747,
          "call": "_open_tag"
        },
        {
          "line": 536,
          "call": "_open_tag"
        },
        {
          "line": 648,
          "call": "_open_tag"
        },
        {
          "line": 714,
          "call": "_internal_system_message"
        },
        {
          "line": 735,
          "call": "_internal_system_message"
        },
        {
          "line": 541,
          "call": "_open_tag"
        },
        {
          "line": 547,
          "call": "_open_tag"
        },
        {
          "line": 658,
          "call": "_open_tag"
        },
        {
          "line": 680,
          "call": "_open_tag"
        },
        {
          "line": 694,
          "call": "_open_tag"
        }
      ],
      "top_level_nodes": [
        {
          "line": 1,
          "code": "'Kimi K3 XTML encoding helpers.\\n\\nThis module keeps chat rendering in Python.\\nCallers that need token IDs should consume ``EncodeSegment`` objects directly:\\nstructural markers may be encoded as tiktoken special tokens, while user/tool\\ntext and attribute values are encoded as ordinary text.\\n'"
        },
        {
          "line": 15,
          "code": "OPEN_TOKEN = '<|open|>'"
        },
        {
          "line": 16,
          "code": "CLOSE_TOKEN = '<|close|>'"
        },
        {
          "line": 17,
          "code": "SEP_TOKEN = '<|sep|>'"
        },
        {
          "line": 18,
          "code": "END_OF_MSG_TOKEN = '<|end_of_msg|>'"
        },
        {
          "line": 19,
          "code": "IMAGE_PLACEHOLDER = '<|kimi_image_placeholder|>'"
        },
        {
          "line": 21,
          "code": "_VALID_THINKING_EFFORTS = {'low', 'high', 'max'}"
        },
        {
          "line": 193,
          "code": "XtmlArgument = tuple[str, str, str]"
        }
      ]
    },
    {
      "file": "kimi_k3_processor.py",
      "sha256": "ec9f7e86d2ab0eee07a8e7e7c037046e77ac3c25a710ad1298ec13be3b585b54",
      "bytes": 7660,
      "imports": [
        {
          "line": 7,
          "statement": "from transformers.feature_extraction_utils import BatchFeature"
        },
        {
          "line": 8,
          "statement": "from transformers.processing_utils import ProcessorMixin"
        },
        {
          "line": 9,
          "statement": "from transformers.utils import logging"
        },
        {
          "line": 11,
          "statement": "from .media_utils import ensure_media_type"
        }
      ],
      "effect_related_calls": [],
      "top_level_nodes": [
        {
          "line": 1,
          "code": "\"Kimi-K3 processor: wraps vision processor + tokenizer into a single interface.\\n\\nChat rendering (including XTML tool-result ordering) is handled by the\\ntokenizer's Python encoder; this processor adds multimodal media preprocessing.\\n\""
        },
        {
          "line": 13,
          "code": "logger = logging.get_logger(__name__)"
        }
      ]
    },
    {
      "file": "kimi_k3_vision_processing.py",
      "sha256": "d122b30bfd3a51a6f05d4bfcfda1e657827322b1353f7caefeebc2835d7736b5",
      "bytes": 6686,
      "imports": [
        {
          "line": 4,
          "statement": "import json"
        },
        {
          "line": 5,
          "statement": "from typing import Any, Dict, Optional, Union"
        },
        {
          "line": 7,
          "statement": "import numpy as np"
        },
        {
          "line": 8,
          "statement": "import torch"
        },
        {
          "line": 9,
          "statement": "from PIL import Image"
        },
        {
          "line": 10,
          "statement": "from transformers.image_processing_utils import BaseImageProcessor, BatchFeature"
        },
        {
          "line": 12,
          "statement": "from transformers.utils import TensorType"
        },
        {
          "line": 14,
          "statement": "from .media_utils import MediaInput, TransparentBgConfig, _to_tensor, ensure_media_type, image_to_np, navit_patchify, navit_resize_image, normalize"
        }
      ],
      "effect_related_calls": [
        {
          "line": 170,
          "call": "config_dict.copy"
        }
      ],
      "top_level_nodes": [
        {
          "line": 1,
          "code": "'Image processor class for Kimi-K3.\\n'"
        }
      ]
    },
    {
      "file": "media_utils.py",
      "sha256": "78403540328f9847d6b7ebc5c44eb2e6a752863de0afb7d0710728bb161dc60d",
      "bytes": 13844,
      "imports": [
        {
          "line": 1,
          "statement": "import base64"
        },
        {
          "line": 2,
          "statement": "import functools"
        },
        {
          "line": 3,
          "statement": "import io"
        },
        {
          "line": 4,
          "statement": "import math"
        },
        {
          "line": 5,
          "statement": "from dataclasses import dataclass"
        },
        {
          "line": 6,
          "statement": "from typing import Literal, TypedDict"
        },
        {
          "line": 8,
          "statement": "import numpy as np"
        },
        {
          "line": 9,
          "statement": "from PIL import Image"
        },
        {
          "line": 361,
          "statement": "import torch"
        }
      ],
      "effect_related_calls": [
        {
          "line": 182,
          "call": "Image.open"
        },
        {
          "line": 184,
          "call": "Image.open"
        },
        {
          "line": 186,
          "call": "Image.open"
        }
      ],
      "top_level_nodes": [
        {
          "line": 17,
          "code": "MediaInput = ImageInput"
        }
      ]
    },
    {
      "file": "modeling_kimi_k3.py",
      "sha256": "b9171c96726eda55234c92ac8dfae7e24c512fda68968ae8f2c3782b42665ea2",
      "bytes": 53444,
      "imports": [
        {
          "line": 25,
          "statement": "import math"
        },
        {
          "line": 26,
          "statement": "from collections.abc import Sequence"
        },
        {
          "line": 27,
          "statement": "from copy import deepcopy"
        },
        {
          "line": 28,
          "statement": "from typing import Optional"
        },
        {
          "line": 30,
          "statement": "import numpy as np"
        },
        {
          "line": 31,
          "statement": "import torch"
        },
        {
          "line": 32,
          "statement": "import torch.nn as nn"
        },
        {
          "line": 33,
          "statement": "import torch.nn.functional as F"
        },
        {
          "line": 34,
          "statement": "from transformers import activations"
        },
        {
          "line": 42,
          "statement": "from transformers.activations import PytorchGELUTanh"
        },
        {
          "line": 43,
          "statement": "from transformers.configuration_utils import PretrainedConfig"
        },
        {
          "line": 44,
          "statement": "from transformers.modeling_utils import PreTrainedModel"
        },
        {
          "line": 45,
          "statement": "from transformers.models.llava.modeling_llava import LlavaCausalLMOutputWithPast"
        },
        {
          "line": 47,
          "statement": "from transformers.utils import is_flash_attn_2_available"
        },
        {
          "line": 49,
          "statement": "from .configuration_kimi_k3 import KimiK3Config"
        },
        {
          "line": 50,
          "statement": "from .modeling_kimi_linear import KimiLinearForCausalLM"
        },
        {
          "line": 37,
          "statement": "from transformers.activations import PytorchGELUTanh"
        },
        {
          "line": 54,
          "statement": "from flash_attn import flash_attn_varlen_func"
        },
        {
          "line": 39,
          "statement": "from transformers.activations import GELUTanh"
        }
      ],
      "effect_related_calls": [
        {
          "line": 658,
          "call": "deepcopy"
        }
      ],
      "top_level_nodes": [
        {
          "line": 36,
          "code": "try:\n    from transformers.activations import PytorchGELUTanh\nexcept ImportError:\n    from transformers.activations import GELUTanh\n    activations.PytorchGELUTanh = GELUTanh\n    PytorchGELUTanh = GELUTanh"
        },
        {
          "line": 53,
          "code": "if is_flash_attn_2_available():\n    from flash_attn import flash_attn_varlen_func\nelse:\n    flash_attn_varlen_func = None"
        },
        {
          "line": 136,
          "code": "VL_VISION_ATTENTION_FUNCTIONS = {'flash_attention_2': multihead_attention, 'eager': eager_attention}"
        }
      ]
    },
    {
      "file": "modeling_kimi_linear.py",
      "sha256": "9e3564c70ac21854ce5a090cc946c5dc76b70d1050ef50840449181a20fff44a",
      "bytes": 51506,
      "imports": [
        {
          "line": 24,
          "statement": "import math"
        },
        {
          "line": 25,
          "statement": "from collections.abc import Callable"
        },
        {
          "line": 26,
          "statement": "from typing import Any"
        },
        {
          "line": 28,
          "statement": "import torch"
        },
        {
          "line": 29,
          "statement": "import torch.nn.functional as F"
        },
        {
          "line": 30,
          "statement": "import transformers"
        },
        {
          "line": 31,
          "statement": "from einops import rearrange"
        },
        {
          "line": 32,
          "statement": "from packaging import version"
        },
        {
          "line": 33,
          "statement": "from torch import nn"
        },
        {
          "line": 34,
          "statement": "from transformers.activations import ACT2FN"
        },
        {
          "line": 35,
          "statement": "from transformers.cache_utils import Cache"
        },
        {
          "line": 36,
          "statement": "from transformers.generation import GenerationMixin"
        },
        {
          "line": 37,
          "statement": "from transformers.masking_utils import create_causal_mask"
        },
        {
          "line": 38,
          "statement": "from transformers.modeling_flash_attention_utils import FlashAttentionKwargs"
        },
        {
          "line": 39,
          "statement": "from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast"
        },
        {
          "line": 40,
          "statement": "from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel"
        },
        {
          "line": 41,
          "statement": "from transformers.processing_utils import Unpack"
        },
        {
          "line": 42,
          "statement": "from transformers.pytorch_utils import ALL_LAYERNORM_LAYERS"
        },
        {
          "line": 43,
          "statement": "from transformers.utils import TransformersKwargs, auto_docstring, can_return_tuple, logging"
        },
        {
          "line": 44,
          "statement": "from transformers.utils.generic import OutputRecorder, check_model_inputs"
        },
        {
          "line": 55,
          "statement": "from .configuration_kimi_k3 import KimiLinearConfig"
        },
        {
          "line": 47,
          "statement": "from fla.modules import FusedRMSNormGated, ShortConvolution"
        },
        {
          "line": 48,
          "statement": "from fla.ops.kda import chunk_kda, fused_recurrent_kda"
        },
        {
          "line": 50,
          "statement": "from fla.ops.utils.index import prepare_cu_seqlens_from_mask, prepare_lens_from_mask"
        },
        {
          "line": 51,
          "statement": "from fla.utils import tensor_cache"
        },
        {
          "line": 699,
          "statement": "import torch.nn.init as init"
        }
      ],
      "effect_related_calls": [],
      "top_level_nodes": [
        {
          "line": 46,
          "code": "try:\n    from fla.modules import FusedRMSNormGated, ShortConvolution\n    from fla.ops.kda import chunk_kda, fused_recurrent_kda\n    from fla.ops.utils.index import prepare_cu_seqlens_from_mask, prepare_lens_from_mask\n    from fla.utils import tensor_cache\nexcept ImportError:\n    raise ImportError('Plese run `pip install -U fla-core`')"
        },
        {
          "line": 57,
          "code": "assert version.parse(transformers.__version__) >= version.parse('4.56.0'), 'Please upgrade transformers to >= 4.56.0'"
        },
        {
          "line": 60,
          "code": "logger = logging.get_logger(__name__)"
        },
        {
          "line": 85,
          "code": "ACT2FN['situ'] = SituAndMul"
        },
        {
          "line": 239,
          "code": "ALL_LAYERNORM_LAYERS.append(KimiRMSNorm)"
        }
      ]
    },
    {
      "file": "tokenization_kimi.py",
      "sha256": "f28ea66e2d862a2a5814970b2ce40c2f7d8296ff09aed90a7e7def689b906944",
      "bytes": 16145,
      "imports": [
        {
          "line": 1,
          "statement": "import os"
        },
        {
          "line": 2,
          "statement": "from logging import getLogger"
        },
        {
          "line": 3,
          "statement": "from pathlib import Path"
        },
        {
          "line": 4,
          "statement": "from shutil import copyfile"
        },
        {
          "line": 5,
          "statement": "from typing import Dict, Iterator, List, Optional, Tuple, Union, cast"
        },
        {
          "line": 7,
          "statement": "import tiktoken"
        },
        {
          "line": 8,
          "statement": "from tiktoken.load import load_tiktoken_bpe"
        },
        {
          "line": 9,
          "statement": "from tokenizers import AddedToken"
        },
        {
          "line": 10,
          "statement": "from transformers.convert_slow_tokenizer import bytes_to_unicode"
        },
        {
          "line": 11,
          "statement": "from transformers.tokenization_utils import PreTrainedTokenizer"
        },
        {
          "line": 14,
          "statement": "from .encoding_k3 import build_chat_segments, is_batched_conversation"
        },
        {
          "line": 16,
          "statement": "from encoding_k3 import build_chat_segments, is_batched_conversation"
        }
      ],
      "effect_related_calls": [
        {
          "line": 99,
          "call": "load_tiktoken_bpe"
        },
        {
          "line": 353,
          "call": "copyfile"
        }
      ],
      "top_level_nodes": [
        {
          "line": 13,
          "code": "try:\n    from .encoding_k3 import build_chat_segments, is_batched_conversation\nexcept ImportError:\n    from encoding_k3 import build_chat_segments, is_batched_conversation"
        },
        {
          "line": 18,
          "code": "logger = getLogger(__name__)"
        },
        {
          "line": 19,
          "code": "VOCAB_FILES_NAMES = {'vocab_file': 'tiktoken.model'}"
        }
      ]
    }
  ],
  "license": "Kimi K3 License for Moonshot code with Apache-2.0 adapted portions; FLA reference adaptations MIT. All three license texts retained.",
  "import_time_effects": [
    "Imports Python/NumPy/Torch/Transformers; upstream K3 hard imports FLA and imports/registry-mutates ACT2FN and ALL_LAYERNORM_LAYERS.",
    "Upstream processor/tokenizer closure is not used by fixture: byte tokenizer is authored independently. tiktoken load_tiktoken_bpe can read local or remote vocabulary at tokenizer construction. media_utils reads image filenames/data on explicit invocation, not import.",
    "No subprocess, shell, eval/exec, socket, network request, or model-weight load found in upstream module import-time statements. Decorators and dependency imports execute library initialization. This is source review, not sandboxing."
  ],
  "runtime_scope": "Four-file patched full Kimi K3 wrapper closure plus installed torch/numpy/transformers. CPU reference math replaces FLA imports entirely; no lazy or eager GPU dependency imports from text runtime. Vision keeps conditional flash-attn import and executes original eager or true Torch SDPA attention on CPU, according to the actual loaded backend.",
  "network_policy": "No upstream model weights fetched. Custom capture explicitly permits pinned code metadata/download from the selected private model repo; local model weights. Not entirely offline capture. Comparison independent of code and GPU.",
  "patch_provenance": "runtime-patches.json; cpu_reference.py; reference/pin.json",
  "review_limitations": "Static source review is not a security guarantee. Dependency implementations belong to the pinned environment. Main must run all runtime checks before release.",
  "reference_sources": {
    "LICENSE": {
      "sha256": "41a83c8187efc1e3ccc21909e806a9e52338e69448554d9754706c3d1cd610e7",
      "bytes": 1096
    },
    "convolution.py": {
      "sha256": "2ab60fc508097457a2ad01dd1f6602ed7e1b47d8a2468949944402b7d5451b8a",
      "bytes": 1194
    },
    "fused_norm_gate.py": {
      "sha256": "109de719d6f9b4e692e3fd504eda4a6e897156c8dc9ac464a6f94193fa94d677",
      "bytes": 40532
    },
    "gate.py": {
      "sha256": "7ea46d62149b8e28e0827e070fee327576de783cb981e3b855182f2b8198c16e",
      "bytes": 15758
    },
    "l2norm.py": {
      "sha256": "5160cc4ff716c9975c655b81fd0e1bae700501dff3b4d6b045680a93185a1e99",
      "bytes": 7792
    },
    "naive.py": {
      "sha256": "60a32285d4b67068ff633b48bbe8ab31028066d24f00d27e12199a88fc73f016",
      "bytes": 6339
    },
    "pin.json": {
      "sha256": "b99a0f4427f07729fc1bde070e48a1814c395f413a8103ff6828cb36bd94ca47",
      "bytes": 111
    },
    "short_conv.py": {
      "sha256": "e2ee319f11b0421545ea5eaa233f12f91e6904ce3dba83cee04a4b25110e6afe",
      "bytes": 10389
    }
  },
  "patched_runtime_sources": {
    "__init__.py": {
      "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
      "bytes": 0
    },
    "configuration_kimi_k3.py": {
      "sha256": "735eb9ebe593e17d231e08e1df7f7be9b5ee0e079f511aa201f9572077b416ae",
      "bytes": 11343
    },
    "cpu_reference.py": {
      "sha256": "f5ab2904ea43b434bcc17fb61a104f6c2975e5f52a39cffbbcfa94b358350e4e",
      "bytes": 5976
    },
    "modeling_kimi_k3.py": {
      "sha256": "344de6dfed495bd6e66e1d16abac12ba6cfd222e872b15ad73003f38462f35af",
      "bytes": 54765
    },
    "modeling_kimi_linear.py": {
      "sha256": "94956f39cd28bf211acf70a3e38f469db5d24637b76a4f58b0b096527bb70df5",
      "bytes": 50914
    }
  },
  "generation_compatibility": "Root directly inherits GenerationMixin; keeps upstream prepare_inputs_for_generation slicing; migrates cached image mask and beam cache access to KimiDynamicCache.",
  "dtype_initialization": "Untouched FP32 initialization -> temporary FP32 checkpoint -> actual BF16 loader selecting FP32-retained parameters -> per-tensor original.to(target_dtype) equality -> final exact mixed-dtype save/reload. No pre-rounding of sensitive parameters.",
  "vision_backend": "Original eager/FlashAttention2 dispatch plus true Torch SDPA packed-clip path. Backend-specific floating-point equivalence is not claimed."
}