Upload folder using huggingface_hub

README.md

@@ -0,0 +1,59 @@
+---
+language: en
+library_name: mlx
+pipeline_tag: image-text-to-text
+tags:
+- mlx
+license: other
+license_name: kimi-k3
+license_link: https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE
+base_model: moonshotai/Kimi-K3
+base_model_relation: quantized
+---
+
+# kernelpool/Kimi-K3-2bit-UVMAX
+
+Mixed-precision (UVMAX) quantization of [moonshotai/Kimi-K3](https://huggingface.co/moonshotai/Kimi-K3).
+
+## What is UVMAX?
+
+UVMAX is a mixed-precision scheme: bit widths are assigned per tensor class
+from measured round-trip quantization error, rather than uniformly.
+
+| Tensor class | Bits | Parameters | Size | Share |
+|---|---|---|---|---|
+| Expert FFNs (routed, latent space) | 2 (gs 128) | 2.72 T | 713.2 GiB | 93.8% |
+| Shared experts, MoE latent projections, dense MLP | 8 (gs 64) | 17.5 B | 17.4 GiB | 2.3% |
+| Attention (KDA + MLA, all projections) | 6 (gs 64) | 36 B | 27.4 GiB | 3.6% |
+| Embeddings, `lm_head` | 4 (gs 64) | 2.4 B | 1.2 GiB | 0.2% |
+| MoE routers | 8 (gs 64) | 0.6 B | 0.6 GiB | 0.1% |
+| Vision tower + projector (unquantized bf16) | — | 0.4 B | 0.8 GiB | 0.1% |
+| Norms, AttnRes projections, gate params (unquantized) | — | — | 0.1 GiB | <0.1% |
+
+## Use with mlx
+
+This model requires Kimi K3 support from mlx-lm PR #TBD, which has not yet
+been merged. Until it is included in an mlx-lm release, install mlx-lm from
+the PR branch:
+
+```bash
+pip install git+https://github.com/ml-explore/mlx-lm.git@refs/pull/TBD/head
+pip install tiktoken
+```
+
+```python
+from mlx_lm import load, generate
+
+model, tokenizer = load(
+    "kernelpool/Kimi-K3-2bit-UVMAX",
+    tokenizer_config={"trust_remote_code": True},
+    trust_remote_code=True,
+)
+
+prompt = "hello"
+
+messages = [{"role": "user", "content": prompt}]
+prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
+
+response = generate(model, tokenizer, prompt=prompt, verbose=True)
+```

added_tokens.json

@@ -0,0 +1,9 @@
+{
+  "<|end_header_id|>": 163844,
+  "<|im_assistant|>": 163842,
+  "<|im_end|>": 163840,
+  "<|im_middle|>": 163846,
+  "<|im_system|>": 163845,
+  "<|im_user|>": 163841,
+  "<|start_header_id|>": 163843
+}

configuration_kimi_k3.py

@@ -0,0 +1,285 @@
+from typing import Optional
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class KimiLinearConfig(PretrainedConfig):
+    model_type = "kimi_linear"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        model_type="kimi_linear",
+        vocab_size=163840,
+        hidden_size=4096,
+        head_dim=None,
+        intermediate_size=11008,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=None,
+        hidden_act="silu",
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        tie_word_embeddings=False,
+        moe_intermediate_size: Optional[int] = None,
+        moe_renormalize: bool = True,
+        moe_router_activation_func: str = "sigmoid",
+        num_experts: Optional[int] = None,
+        num_experts_per_token: Optional[int] = None,
+        num_shared_experts: int = 0,
+        routed_scaling_factor: float = 1.0,
+        first_k_dense_replace: int = 0,
+        moe_layer_freq: int = 1,
+        use_grouped_topk: bool = True,
+        num_expert_group: int = 1,
+        topk_group: int = 1,
+        q_lora_rank: Optional[int] = None,
+        kv_lora_rank: Optional[int] = None,
+        qk_nope_head_dim: Optional[int] = None,
+        qk_rope_head_dim: Optional[int] = None,
+        v_head_dim: Optional[int] = None,
+        mla_use_nope: Optional[bool] = False,
+        mla_use_output_gate: Optional[bool] = False,
+        num_nextn_predict_layers: int = 0,
+        linear_attn_config: Optional[dict] = None,
+        attn_res_block_size: Optional[int] = None,
+        latent_moe_use_norm: bool = False,
+        activation_situ_beta: Optional[float] = None,
+        activation_situ_linear_beta: Optional[float] = None,
+        max_position_embeddings: int = 4096,
+        routed_expert_hidden_size: Optional[int] = None,
+        topk_method: str = "noaux_tc",
+        **kwargs,
+    ):
+        self.model_type = model_type
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.head_dim = (
+            head_dim if head_dim is not None else hidden_size // num_attention_heads
+        )
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+
+        self.q_lora_rank = q_lora_rank
+        self.kv_lora_rank = kv_lora_rank
+        self.qk_nope_head_dim = qk_nope_head_dim
+        self.qk_rope_head_dim = qk_rope_head_dim
+        self.v_head_dim = v_head_dim
+        self.mla_use_nope = mla_use_nope
+        self.mla_use_output_gate = mla_use_output_gate
+        # moe config
+        self.num_experts = num_experts
+        self.num_experts_per_token = num_experts_per_token
+        self.moe_renormalize = moe_renormalize
+        self.num_shared_experts = num_shared_experts
+        self.routed_scaling_factor = routed_scaling_factor
+        self.moe_router_activation_func = moe_router_activation_func
+        assert self.moe_router_activation_func in ("softmax", "sigmoid")
+        self.moe_intermediate_size = moe_intermediate_size
+        self.first_k_dense_replace = first_k_dense_replace
+        self.moe_layer_freq = moe_layer_freq
+        self.use_grouped_topk = use_grouped_topk
+        self.num_expert_group = num_expert_group
+        self.topk_group = topk_group
+        self.num_nextn_predict_layers = num_nextn_predict_layers
+
+        self.attn_res_block_size = attn_res_block_size
+        self.latent_moe_use_norm = latent_moe_use_norm
+        self.activation_situ_beta = activation_situ_beta
+        self.activation_situ_linear_beta = activation_situ_linear_beta
+        self.max_position_embeddings = max_position_embeddings
+        self.routed_expert_hidden_size = routed_expert_hidden_size
+        self.topk_method = topk_method
+
+        if linear_attn_config is not None:
+            assert linear_attn_config["kda_layers"] is not None
+            assert linear_attn_config["full_attn_layers"] is not None
+        self.linear_attn_config = linear_attn_config
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+    @property
+    def is_mla(self):
+        return (
+            self.q_lora_rank is not None
+            or self.kv_lora_rank is not None
+            or self.qk_nope_head_dim is not None
+            or self.qk_rope_head_dim is not None
+            or self.v_head_dim is not None
+            or self.mla_use_nope is True
+        )
+
+    @property
+    def is_moe(self):
+        return self.num_experts is not None
+
+    @property
+    def is_linear_attn(self) -> bool:
+        return not (
+            self.linear_attn_config is None
+            or (
+                isinstance(self.linear_attn_config, dict)
+                and self.linear_attn_config["kda_layers"] is not None
+                and len(self.linear_attn_config["kda_layers"]) == 0
+            )
+        )
+
+    def is_kda_layer(self, layer_idx: int):
+        return (
+            self.linear_attn_config is not None
+            and (layer_idx + 1) in self.linear_attn_config["kda_layers"]
+        )
+
+
+class KimiK3VisionConfig(PretrainedConfig):
+
+    def __init__(
+            self,
+            patch_size: int = 14,
+            init_pos_emb_height: int = 64,
+            init_pos_emb_width: int = 64,
+            init_pos_emb_time: int = 4,
+            pos_emb_type: str = 'divided_fixed',
+            vt_num_attention_heads: int = 12,
+            vt_num_hidden_layers: int = 27,
+            vt_hidden_size: int = 1024,
+            vt_intermediate_size: int = 4096,
+            merge_kernel_size: tuple = (2, 2),
+            merge_type: str = 'sd2_tpool',
+            _attn_implementation: str = 'flash_attention_2',
+            # MM Projector parameters
+            mm_projector_type: str = 'patchmergerv2',
+            mm_hidden_size: int | None = None,
+            projector_hidden_act: str = "gelu",
+            projector_ln_eps: float = 1e-5,
+            # vision tower parameters
+            qkv_hidden_size: int = 1536,
+            norm_type: str = 'rmsnorm',
+            attn_bias: bool = False,
+            patch_embed_proj_bias: bool = False,
+            mlp_type: str = 'mlp2',
+            linear_bias: bool = False,
+            activation_func: str = 'gelu_pytorch_tanh',
+            pos_emb_interpolation_mode: str = 'bilinear',
+            # Other parameters
+            ignore_index: int = -100,
+            media_placeholder_token_id: int = 163605,
+            pad_token_id: int = 0,
+            text_hidden_size=7168,
+            **kwargs):
+
+        self.patch_size = patch_size
+        self.init_pos_emb_height = init_pos_emb_height
+        self.init_pos_emb_width = init_pos_emb_width
+        self.init_pos_emb_time = init_pos_emb_time
+        self.pos_emb_type = pos_emb_type
+        self.vt_num_attention_heads = vt_num_attention_heads
+        self.vt_num_hidden_layers = vt_num_hidden_layers
+        self.vt_hidden_size = vt_hidden_size
+        self.vt_intermediate_size = vt_intermediate_size
+        self.merge_kernel_size = merge_kernel_size
+        self.merge_type = merge_type
+        self._attn_implementation = _attn_implementation
+
+        # MM Projector config
+        self.mm_projector_type = mm_projector_type
+        self.mm_hidden_size = mm_hidden_size if mm_hidden_size is not None else vt_hidden_size
+        self.projector_hidden_act = projector_hidden_act
+        self.projector_ln_eps = projector_ln_eps
+        self.text_hidden_size = text_hidden_size
+
+        # vision tower parameters
+        self.qkv_hidden_size = qkv_hidden_size
+        self.norm_type = norm_type
+        self.attn_bias = attn_bias
+        self.patch_embed_proj_bias = patch_embed_proj_bias
+        self.mlp_type = mlp_type
+        self.linear_bias = linear_bias
+        self.activation_func = activation_func
+        self.pos_emb_interpolation_mode = pos_emb_interpolation_mode
+
+        super().__init__(**kwargs)
+
+
+class KimiK3Config(PretrainedConfig):
+    """Kimi-K3 model configuration.
+
+    Args:
+        text_config (dict | KimiLinearConfig): Configuration for the text model.
+        
+        Vision Tower Parameters (from MoonViT3dConfig):
+            patch_size (int): Patch size for vision tower.
+            init_pos_emb_height (int): Initial position embedding height.
+            init_pos_emb_width (int): Initial position embedding width.
+            init_pos_emb_time (int): Initial position embedding time dimension.
+            pos_emb_type (str): Type of position embedding.
+            vt_num_attention_heads (int): Number of attention heads in vision tower.
+            vt_num_hidden_layers (int): Number of hidden layers in vision tower.
+            vt_hidden_size (int): Hidden size of vision tower.
+            vt_intermediate_size (int): Intermediate size in vision tower FFN.
+            merge_kernel_size (tuple): Kernel size for patch merging.
+            merge_type (str): Type of merge operation.
+            _attn_implementation (str): Attention implementation type.
+        
+        MM Projector Parameters (from MultiModalProjectorConfig):
+            mm_projector_type (str): Type of multimodal projector.
+            mm_hidden_size (int): Hidden size from vision tower (should match vt_hidden_size).
+            projector_hidden_act (str): Activation function for projector.
+            projector_ln_eps (float): Layer norm epsilon for projector.
+        
+        Other Parameters:
+            ignore_index (int): The ignore index for the loss function.
+            media_placeholder_token_id (int): The token ID to use for media placeholders.
+            pad_token_id (int): The token ID to use for padding.
+    """
+
+    model_type = "kimi_k3"
+
+    def __init__(
+        self,
+        text_config: dict | KimiLinearConfig = None,
+        vision_config: dict | KimiK3VisionConfig = None,
+        # Other parameters
+        ignore_index: int = -100,
+        media_placeholder_token_id: int = 163605,
+        pad_token_id: int = 0,
+        **kwargs,
+    ):
+        if isinstance(text_config, dict):
+            text_config = KimiLinearConfig(**text_config)
+        if isinstance(vision_config, dict):
+            vision_config = KimiK3VisionConfig(**vision_config)
+        self.text_config = text_config
+        self.vision_config = vision_config
+        # Other config
+        self.ignore_index = ignore_index
+        self.media_placeholder_token_id = media_placeholder_token_id
+        if getattr(self.text_config, "quantization_config", None) is not None:
+            self.quantization_config = self.text_config.quantization_config
+
+        super().__init__(pad_token_id=pad_token_id, **kwargs)

encoding_k3.py

@@ -0,0 +1,647 @@
+"""Kimi K3 XTML encoding helpers.
+
+This module keeps chat rendering in Python.
+Callers that need token IDs should consume ``EncodeSegment`` objects directly:
+structural markers may be encoded as tiktoken special tokens, while user/tool
+text and attribute values are encoded as ordinary text.
+"""
+
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass
+from typing import Any, Iterable, Optional
+
+OPEN_TOKEN = "<|open|>"
+CLOSE_TOKEN = "<|close|>"
+SEP_TOKEN = "<|sep|>"
+END_OF_MSG_TOKEN = "<|end_of_msg|>"
+IMAGE_PLACEHOLDER = "<|kimi_image_placeholder|>"
+
+_VALID_THINKING_EFFORTS = {"low", "high", "max"}
+
+
+@dataclass(frozen=True)
+class EncodeSegment:
+    text: str
+    allow_special: bool = False
+
+
+class _ImagePromptState:
+    def __init__(self, image_prompts: Optional[list[str]] = None):
+        self.image_prompts = image_prompts
+        self.index = 0
+
+    def next_prompt(self) -> str:
+        if self.image_prompts is None:
+            return IMAGE_PLACEHOLDER
+        if self.index >= len(self.image_prompts):
+            raise ValueError("More image placeholders than image prompts.")
+        prompt = self.image_prompts[self.index]
+        self.index += 1
+        return prompt
+
+    def assert_consumed(self) -> None:
+        if self.image_prompts is None:
+            return
+        if self.index != len(self.image_prompts):
+            raise ValueError(
+                f"image prompt count {len(self.image_prompts)} != "
+                f"consumed placeholder count {self.index}"
+            )
+
+
+def _segment(text: Any, *, allow_special: bool = False) -> list[EncodeSegment]:
+    text = str(text)
+    if not text:
+        return []
+    return [EncodeSegment(text, allow_special=allow_special)]
+
+
+def _control(text: str) -> list[EncodeSegment]:
+    return _segment(text, allow_special=True)
+
+
+def _text(text: Any) -> list[EncodeSegment]:
+    return _segment(text, allow_special=False)
+
+
+def _append_text(
+    segments: list[EncodeSegment],
+    text: Any,
+    image_state: _ImagePromptState,
+) -> None:
+    text = str(text)
+    if text == "":
+        return
+    if image_state.image_prompts is None or IMAGE_PLACEHOLDER not in text:
+        segments.extend(_text(text))
+        return
+
+    parts = text.split(IMAGE_PLACEHOLDER)
+    for i, part in enumerate(parts):
+        segments.extend(_text(part))
+        if i < len(parts) - 1:
+            segments.extend(_segment(image_state.next_prompt(),
+                                     allow_special=True))
+
+
+def _escape_attr_value(value: Any) -> str:
+    return str(value).replace("&", "&amp;").replace('"', "&quot;")
+
+
+def _attr(key: str, value: Any) -> list[EncodeSegment]:
+    return (
+        _text(f" {key}")
+        + _text('="')
+        + _text(_escape_attr_value(value))
+        + _text('"')
+    )
+
+
+def _open_tag(tag: str, attrs: Iterable[tuple[str, Any]] = ()) -> list[EncodeSegment]:
+    segments: list[EncodeSegment] = []
+    segments.extend(_control(OPEN_TOKEN))
+    segments.extend(_text(tag))
+    for key, value in attrs:
+        segments.extend(_attr(key, value))
+    segments.extend(_control(SEP_TOKEN))
+    return segments
+
+def _close_tag(tag: str) -> list[EncodeSegment]:
+    segments: list[EncodeSegment] = []
+    segments.extend(_control(CLOSE_TOKEN))
+    segments.extend(_text(tag))
+    segments.extend(_control(SEP_TOKEN))
+    return segments
+
+
+def _end_of_msg() -> list[EncodeSegment]:
+    return _control(END_OF_MSG_TOKEN)
+
+
+def _json_compact(value: Any) -> str:
+    return json.dumps(value, ensure_ascii=False, separators=(",", ":"))
+
+
+def _is_mapping(value: Any) -> bool:
+    return isinstance(value, dict)
+
+
+def _xtml_type(value: Any) -> str:
+    if isinstance(value, bool):
+        return "boolean"
+    if value is None:
+        return "null"
+    if isinstance(value, (int, float)) and not isinstance(value, bool):
+        return "number"
+    if isinstance(value, str):
+        return "string"
+    if _is_mapping(value):
+        return "object"
+    return "array"
+
+
+def _xtml_value(value: Any) -> str:
+    if isinstance(value, str):
+        return value
+    return json.dumps(value, ensure_ascii=False)
+
+
+def _get_value(obj: Any, key: str, default: Any = None) -> Any:
+    if isinstance(obj, dict):
+        return obj.get(key, default)
+    return getattr(obj, key, default)
+
+
+def extract_response_schema(response_format: Any) -> Any:
+    if response_format is None:
+        return None
+
+    json_schema = _get_value(response_format, "json_schema")
+    if json_schema is None:
+        return None
+
+    if isinstance(json_schema, dict):
+        return json_schema.get(
+            "schema",
+            json_schema.get("json_schema", json_schema),
+        )
+
+    schema = _get_value(json_schema, "schema")
+    if schema is not None:
+        return schema
+
+    schema = _get_value(json_schema, "json_schema")
+    if schema is not None:
+        return schema
+
+    return json_schema
+
+
+def deep_sort_dict(obj: Any) -> Any:
+    if isinstance(obj, dict):
+        return {k: deep_sort_dict(v) for k, v in sorted(obj.items())}
+    if isinstance(obj, list):
+        return [deep_sort_dict(item) for item in obj]
+    return obj
+
+
+def normalize_tool_arguments(arguments: Any) -> tuple[dict[str, Any], Optional[str]]:
+    if arguments is None:
+        return {}, None
+    if isinstance(arguments, dict):
+        return arguments, None
+    if isinstance(arguments, str):
+        if not arguments.strip():
+            return {}, None
+        try:
+            parsed = json.loads(arguments)
+        except json.JSONDecodeError:
+            return {}, arguments
+        if not isinstance(parsed, dict):
+            raise ValueError("Kimi K3 tool call arguments must be a JSON object.")
+        return parsed, None
+    raise TypeError(
+        "Kimi K3 tool call arguments must be a dict or a JSON object string."
+    )
+
+
+def normalize_message(message: Any) -> Any:
+    if not isinstance(message, dict):
+        return message
+
+    normalized = dict(message)
+
+    tools = normalized.get("tools")
+    if tools is not None:
+        normalized["tools"] = deep_sort_dict(tools)
+
+    tool_calls = normalized.get("tool_calls")
+    if not tool_calls:
+        return normalized
+
+    normalized_calls = []
+    for tool_call in tool_calls:
+        if not isinstance(tool_call, dict):
+            normalized_calls.append(tool_call)
+            continue
+
+        tc = dict(tool_call)
+        function = tc.get("function")
+        if isinstance(function, dict):
+            fn = dict(function)
+            arguments, json_block = normalize_tool_arguments(fn.get("arguments"))
+            fn["arguments"] = arguments
+            if json_block is None:
+                fn.pop("_xtml_json_block", None)
+            else:
+                fn["_xtml_json_block"] = json_block
+            tc["function"] = fn
+        else:
+            arguments, json_block = normalize_tool_arguments(tc.get("arguments"))
+            tc["arguments"] = arguments
+            if json_block is None:
+                tc.pop("_xtml_json_block", None)
+            else:
+                tc["_xtml_json_block"] = json_block
+        normalized_calls.append(tc)
+
+    normalized["tool_calls"] = normalized_calls
+    return normalized
+
+
+def normalize_conversation(conversation: Any) -> Any:
+    if not isinstance(conversation, list):
+        return conversation
+
+    def normalize_messages(messages: list[Any]) -> list[Any]:
+        return [normalize_message(message) for message in messages]
+
+    if conversation and isinstance(conversation[0], list):
+        return [normalize_messages(messages) for messages in conversation]
+    return normalize_messages(conversation)
+
+
+def _tool_call_id_index(tool_calls: Any) -> dict:
+    """Map assistant ``tool_calls[].id`` to ``(1-based position, function name)``.
+
+    The position mirrors the chat template's enumeration over ``tool_calls``
+    (every entry advances the position, even an id-less one). Duplicate ids keep
+    their first occurrence.
+    """
+    index: dict = {}
+    if not isinstance(tool_calls, list):
+        return index
+    for position, tool_call in enumerate(tool_calls, start=1):
+        if not isinstance(tool_call, dict):
+            continue
+        call_id = tool_call.get("id")
+        if call_id is None:
+            continue
+        key = str(call_id)
+        if key in index:
+            continue
+        function = tool_call.get("function")
+        name = (
+            function.get("name") if isinstance(function, dict) else tool_call.get("name")
+        )
+        index[key] = (position, name)
+    return index
+
+
+def normalize_xtml_tool_result_messages(messages: list[Any]) -> list[Any]:
+    """Re-sort K3 XTML tool results into assistant ``tool_calls`` order.
+
+    Serving frameworks generally deliver tool results already in call order. A
+    direct Transformers caller, however, may pass OpenAI-style tool messages in any
+    order, so each run of consecutive tool messages is matched against the most
+    recent preceding assistant ``tool_calls`` by opaque ``tool_call_id`` ==
+    ``tool_calls[].id`` (K3 drops the ``func:index`` format requirement) and
+    sorted by the matched 1-based position. The matched call is authoritative,
+    so each matched message's ``tool`` is set to that call's function name --
+    this keeps an explicit (and possibly stale) ``tool``/``name`` from drifting
+    out of sync with the reordered position. ``index`` is still derived from the
+    rendered position by the chat template. A run that cannot be fully matched is
+    left untouched. Re-running is idempotent.
+
+    This function is side-effect free: matched tool messages are shallow-copied
+    before their ``tool``/``name`` is rewritten, and every other message is
+    appended to the output as-is. The input list and its message objects are
+    never mutated.
+    """
+    if not isinstance(messages, list):
+        return messages
+
+    output: list[Any] = []
+    current_index: dict = {}
+    i = 0
+    n = len(messages)
+
+    while i < n:
+        message = messages[i]
+
+        if isinstance(message, dict) and message.get("role") == "assistant":
+            tool_calls = message.get("tool_calls")
+            current_index = _tool_call_id_index(tool_calls) if tool_calls else {}
+            output.append(message)
+            i += 1
+            continue
+
+        if not isinstance(message, dict) or message.get("role") != "tool":
+            output.append(message)
+            i += 1
+            continue
+
+        run: list[tuple] = []  # (position, original_offset, message, name)
+        unresolved = False
+        offset = 0
+        while (
+            i < n and isinstance(messages[i], dict) and messages[i].get("role") == "tool"
+        ):
+            tool_message = messages[i]
+            call_id = tool_message.get("tool_call_id", tool_message.get("id"))
+            matched = current_index.get(str(call_id)) if call_id is not None else None
+            if matched is None:
+                unresolved = True
+                run.append((None, offset, tool_message, None))
+            else:
+                position, name = matched
+                run.append((position, offset, tool_message, name))
+            offset += 1
+            i += 1
+
+        if unresolved:
+            output.extend(item[2] for item in run)
+        else:
+            run.sort(key=lambda item: (item[0], item[1]))
+            for _, _, tool_message, name in run:
+                if name is None:
+                    output.append(tool_message)
+                    continue
+                # The id-matched call is authoritative: align tool (and any
+                # explicit name) so the rendered XTML tool attribute cannot
+                # disagree with the reordered position. Copy first so the
+                # caller's message object is never mutated.
+                resolved = dict(tool_message)
+                resolved["tool"] = name
+                if "name" in resolved:
+                    resolved["name"] = name
+                output.append(resolved)
+
+    return output
+
+
+def is_batched_conversation(conversation: Any) -> bool:
+    return (
+        isinstance(conversation, list)
+        and bool(conversation)
+        and isinstance(conversation[0], list)
+    )
+
+
+def _render_content_segments(
+    content: Any,
+    image_state: _ImagePromptState,
+) -> list[EncodeSegment]:
+    segments: list[EncodeSegment] = []
+    if isinstance(content, str):
+        _append_text(segments, content, image_state)
+    elif content is not None:
+        for part in content:
+            if part["type"] in ["image", "image_url"]:
+                segments.extend(
+                    _segment(image_state.next_prompt(), allow_special=True))
+            else:
+                _append_text(segments, part["text"], image_state)
+    return segments
+
+
+def _internal_system_message(message_type: str, body: str) -> list[EncodeSegment]:
+    segments: list[EncodeSegment] = []
+    segments.extend(_open_tag("message", [("role", "system"), ("type", message_type)]))
+    segments.extend(_text(body.strip()))
+    segments.extend(_close_tag("message"))
+    segments.extend(_end_of_msg())
+    return segments
+
+
+def _render_assistant_segments(
+    message: dict[str, Any],
+    image_state: _ImagePromptState,
+    thinking: bool = True,
+) -> list[EncodeSegment]:
+    segments: list[EncodeSegment] = []
+    # The <think> channel is structural: in thinking mode every assistant
+    # message carries the open/close tags even when there is no reasoning
+    # content to fill in. In non-thinking mode the channel is dropped
+    # entirely.
+    if thinking:
+        reasoning_content = message.get("reasoning_content") or message.get(
+            "reasoning"
+        )
+        segments.extend(_open_tag("think"))
+        if reasoning_content is not None and str(reasoning_content).strip():
+            _append_text(segments, reasoning_content, image_state)
+        segments.extend(_close_tag("think"))
+
+    segments.extend(_open_tag("response"))
+    segments.extend(_render_content_segments(message.get("content"), image_state))
+    segments.extend(_close_tag("response"))
+
+    tool_calls = message.get("tool_calls")
+    if tool_calls:
+        segments.extend(_open_tag("tools"))
+        for index, tool_call in enumerate(tool_calls, start=1):
+            fn = tool_call.get("function", tool_call)
+            segments.extend(
+                _open_tag("call", [("tool", fn["name"]), ("index", index)])
+            )
+            args = fn.get("arguments", {})
+            json_block = fn.get("_xtml_json_block")
+            if json_block is not None:
+                segments.extend(_open_tag("json", [("type", "object")]))
+                _append_text(segments, json_block, image_state)
+                segments.extend(_close_tag("json"))
+            elif _is_mapping(args):
+                for key, value in args.items():
+                    segments.extend(
+                        _open_tag(
+                            "argument",
+                            [("key", key), ("type", _xtml_type(value))],
+                        )
+                    )
+                    _append_text(segments, _xtml_value(value), image_state)
+                    segments.extend(_close_tag("argument"))
+            segments.extend(_close_tag("call"))
+        segments.extend(_close_tag("tools"))
+
+    return segments
+
+
+def _render_tool_declare(tools: Any, *, dynamic: bool = False) -> list[EncodeSegment]:
+    if dynamic:
+        body = (
+            "## New Tools Available\n"
+            "The system dynamically extends the toolset via lazy-loading.\n"
+            "You have access to all existing and extended tools.\n"
+            "Here are the specs for the extended tools.\n\n"
+            "```json\n"
+            f"{_json_compact(tools)}\n"
+            "```"
+        )
+    else:
+        body = (
+            "# Tools\n"
+            "Here are the available tools, described in JSONSchema.\n\n"
+            "```json\n"
+            f"{_json_compact(tools)}\n"
+            "```"
+        )
+    segments: list[EncodeSegment] = []
+    segments.extend(_open_tag("message", [("role", "system"), ("type", "tool-declare")]))
+    segments.extend(_text(body))
+    segments.extend(_close_tag("message"))
+    segments.extend(_end_of_msg())
+    return segments
+
+
+def build_chat_segments(
+    messages: list[Any],
+    tools: Optional[list[dict]] = None,
+    *,
+    add_generation_prompt: bool = True,
+    thinking: bool = True,
+    image_prompts: Optional[list[str]] = None,
+    **kwargs: Any,
+) -> list[EncodeSegment]:
+    # Re-sort tool results by tool_call_id at the lowest layer so every caller
+    # (processor or direct tokenizer) gets correctly ordered XTML. The helper is
+    # side-effect free, so the caller's message objects are left untouched.
+    messages = normalize_xtml_tool_result_messages(messages)
+    messages = normalize_conversation(messages)
+    tools = deep_sort_dict(tools)
+
+    kwargs = dict(kwargs)
+    response_format = kwargs.get("response_format")
+    if "response_schema" not in kwargs:
+        response_schema = extract_response_schema(response_format)
+        if response_schema is not None:
+            kwargs["response_schema"] = response_schema
+    if kwargs.get("response_schema") is not None:
+        kwargs["response_schema"] = deep_sort_dict(kwargs["response_schema"])
+
+    image_state = _ImagePromptState(image_prompts)
+    segments: list[EncodeSegment] = []
+
+    tool_calls = None
+    tool_index = 0
+
+    if tools:
+        segments.extend(_render_tool_declare(tools))
+
+    thinking_effort = kwargs.get("thinking_effort")
+    if thinking and thinking_effort is not None:
+        assert thinking_effort in _VALID_THINKING_EFFORTS, (
+            f"Unsupported thinking_effort={thinking_effort!r}; "
+            f"supported values are {sorted(_VALID_THINKING_EFFORTS)}."
+        )
+    if thinking and thinking_effort in _VALID_THINKING_EFFORTS:
+        segments.extend(
+            _internal_system_message(
+                "thinking-effort",
+                "`thinking_effort` guides on how much to think in your "
+                "thinking channel (not including the response channel), "
+                "supported values include `low`, `medium`, `high`, and `max`.\n"
+                f"Now the system is invoked with `thinking_effort={thinking_effort}`.",
+            )
+        )
+
+    for message_index, message in enumerate(messages):
+        if not isinstance(message, dict):
+            continue
+
+        role = message["role"]
+        if role == "user":
+            attrs = [("role", "user")]
+            if message.get("name"):
+                attrs.append(("name", message["name"]))
+            segments.extend(_open_tag("message", attrs))
+            segments.extend(_render_content_segments(message.get("content"), image_state))
+            segments.extend(_close_tag("message"))
+            segments.extend(_end_of_msg())
+        elif role == "system" and message.get("tools"):
+            segments.extend(_render_tool_declare(message["tools"], dynamic=True))
+        elif role == "system":
+            attrs = [("role", "system")]
+            if message.get("name"):
+                attrs.append(("name", message["name"]))
+            segments.extend(_open_tag("message", attrs))
+            segments.extend(_render_content_segments(message.get("content"), image_state))
+            segments.extend(_close_tag("message"))
+            segments.extend(_end_of_msg())
+        elif role == "tool":
+            tool_index += 1
+            tool_name = message.get("tool", message.get("name"))
+            if (
+                tool_name is None
+                and tool_calls is not None
+                and tool_index <= len(tool_calls)
+            ):
+                tc = tool_calls[tool_index - 1]
+                fn = tc.get("function", tc)
+                tool_name = fn["name"]
+            if tool_name is None:
+                raise ValueError(
+                    "Kimi K3 tool messages need a resolvable tool name: "
+                    "carry `tool`/`name`, or match a preceding assistant "
+                    "tool_call by order."
+                )
+            segments.extend(
+                _open_tag(
+                    "message",
+                    [("role", "tool"), ("tool", tool_name), ("index", tool_index)],
+                )
+            )
+            segments.extend(_render_content_segments(message.get("content"), image_state))
+            segments.extend(_close_tag("message"))
+            segments.extend(_end_of_msg())
+        elif role == "assistant":
+            tool_calls = message.get("tool_calls")
+            tool_index = 0
+            attrs = [("role", "assistant")]
+            if message.get("name"):
+                attrs.append(("name", message["name"]))
+            segments.extend(_open_tag("message", attrs))
+            segments.extend(_render_assistant_segments(message, image_state, thinking))
+            segments.extend(_close_tag("message"))
+            segments.extend(_end_of_msg())
+
+    tool_choice = kwargs.get("tool_choice")
+    if tool_choice == "required":
+        segments.extend(
+            _internal_system_message(
+                "tool-choice",
+                "The system is invoked with `tool_choice=required`.\n"
+                "You MUST call tools in the next message.",
+            )
+        )
+    elif tool_choice == "none":
+        segments.extend(
+            _internal_system_message(
+                "tool-choice",
+                "The system is invoked with `tool_choice=none`.\n"
+                "You MUST NOT call any tools in the next message.",
+            )
+        )
+
+    rf = kwargs.get("response_format")
+    rf_type = _get_value(rf, "type", rf) if isinstance(rf, dict) else rf
+    if rf_type == "json_object":
+        segments.extend(
+            _internal_system_message(
+                "response-format",
+                "The system is invoked with `response_format=json_object`.\n"
+                "Your response must be raw JSON data without markdown code "
+                "blocks (```json) or any additional formatting.",
+            )
+        )
+    elif rf_type == "json_schema":
+        schema = _json_compact(kwargs.get("response_schema"))
+        segments.extend(
+            _internal_system_message(
+                "response-format",
+                "The system is invoked with `response_format=json_schema`.\n"
+                "Your response must be raw JSON data without markdown code "
+                "blocks (```json) or any additional formatting.\n"
+                "The JSON data must match the following schema:\n"
+                f"```json\n{schema}\n```",
+            )
+        )
+
+    if add_generation_prompt:
+        segments.extend(_open_tag("message", [("role", "assistant")]))
+        segments.extend(_open_tag("think" if thinking else "response"))
+
+    image_state.assert_consumed()
+    return segments

generation_config.json

@@ -0,0 +1,4 @@
+{
+  "max_length": 1048576,
+  "eos_token_id": 163586
+}

kimi_k3_processor.py

@@ -0,0 +1,187 @@
+"""Kimi-K3 processor: wraps vision processor + tokenizer into a single interface.
+
+Chat rendering (including XTML tool-result ordering) is handled by the
+tokenizer's Python encoder; this processor adds multimodal media preprocessing.
+"""
+
+from transformers.feature_extraction_utils import BatchFeature
+from transformers.processing_utils import ProcessorMixin
+from transformers.utils import logging
+
+from .media_utils import ensure_media_type
+
+logger = logging.get_logger(__name__)
+
+# ── KimiK3Processor ───────────────────────────────────────────────────
+
+
+class KimiK3Processor(ProcessorMixin):
+    r"""
+    Constructs a KimiK3 processor which wraps a KimiK3 image processor
+    and a tokenizer into a single processor.
+
+    [`KimiK3Processor`] offers all the functionalities of
+    [`KimiK3VisionProcessor`] and [`TikTokenTokenizer`].
+
+    Args:
+        image_processor ([`KimiK3VisionProcessor`], *optional*):
+            The image processor is a required input.
+        tokenizer ([`TikTokenTokenizer`], *optional*):
+            The tokenizer is a required input.
+        chat_template (`str`, *optional*): Kept for ProcessorMixin
+            compatibility. Kimi K3 chat encoding is implemented in Python by
+            the tokenizer.
+    """
+
+    attributes = ["image_processor", "tokenizer"]
+    valid_kwargs = ["chat_template"]
+    image_processor_class = "AutoImageProcessor"
+    tokenizer_class = "AutoTokenizer"
+
+    def __init__(
+        self,
+        image_processor=None,
+        tokenizer=None,
+        chat_template=None,
+        **kwargs,
+    ):
+        super().__init__(image_processor,
+                         tokenizer,
+                         chat_template=chat_template)
+        self.media_processor = image_processor
+        self.image_placeholder = "<|kimi_image_placeholder|>"
+
+    # ── Media preprocessing ────────────────────────────────────────────
+
+    def update_raw_text(self, text: str, image_prompts: list[str]) -> str:
+        # Replace image placeholders
+        image_count = text.count(self.image_placeholder)
+        if image_count > 0:
+            assert image_count == len(image_prompts), (
+                f"image placeholder count {image_count} != "
+                f"image_prompts count {len(image_prompts)}")
+            text_parts = text.split(self.image_placeholder)
+            assert len(text_parts) == len(image_prompts) + 1
+            text = "".join([
+                text_parts[i] + image_prompts[i]
+                for i in range(len(image_prompts))
+            ])
+            text += text_parts[-1]
+
+        return text
+
+    def preprocess_medias(self,
+                          medias: list[dict]) -> tuple[list[dict], list[str]]:
+        """Process media items and generate corresponding prompts.
+
+        Returns:
+            A tuple of (updated_medias, image_prompts).
+        """
+        updated_medias = []
+        image_prompts = []
+        for media in medias:
+            if media['type'] == 'image':
+                updated_medias.append(media)
+                img = ensure_media_type(
+                    media,
+                    transparent_bg_config=self.media_processor.
+                    _transparent_bg_config,
+                    transparent_bg_fill_stage=self.media_processor.
+                    _transparent_bg_fill_stage,
+                )['image']
+                w, h = img.size
+                image_prompts.append(
+                    self.media_processor.make_image_prompt(w, h))
+            else:
+                raise ValueError(f"unsupported media type: {media['type']}")
+        return updated_medias, image_prompts
+
+    # ── Main entry points ──────────────────────────────────────────────
+
+    def __call__(self,
+                 messages: list[dict] = None,
+                 medias: list[dict] = None,
+                 text: str = None,
+                 return_tensors: str = "pt",
+                 **kwargs) -> BatchFeature:
+        """
+        Process multimodal inputs for Kimi-K3 model.
+
+        Args:
+            messages: List of message dicts with 'role' and 'content' fields.
+                     If provided, medias and text will be extracted automatically.
+            medias: Pre-extracted list of media dicts.
+            text: Pre-formatted text string.
+            return_tensors: Format of returned tensors. Default: 'pt'.
+            **kwargs: Additional arguments passed to apply_chat_template.
+
+        Returns:
+            BatchFeature with fields: input_ids, attention_mask,
+            pixel_values, grid_thws.
+        """
+        if messages is None and (medias is None or text is None):
+            raise ValueError(
+                "Provide either 'messages' or both 'medias' and 'text'")
+
+        if medias is not None and text is not None:
+            updated_medias, image_prompts = (self.preprocess_medias(medias))
+            preprocessed = self.media_processor.preprocess(
+                updated_medias, return_tensors=return_tensors)
+            text = self.update_raw_text(text, image_prompts)
+            text_inputs = self.tokenizer(text, return_tensors=return_tensors)
+            return BatchFeature(data={**text_inputs, **preprocessed.data})
+
+        if medias is None:
+            medias = self._extract_medias_from_messages(messages)
+        updated_medias, image_prompts = (self.preprocess_medias(medias))
+        preprocessed = self.media_processor.preprocess(
+            updated_medias, return_tensors=return_tensors)
+
+        if text is None:
+            text_inputs = self.tokenizer.apply_chat_template(
+                messages,
+                tokenize=True,
+                return_tensors=return_tensors,
+                return_dict=True,
+                image_prompts=image_prompts,
+                **kwargs)
+            return BatchFeature(data={**text_inputs, **preprocessed.data})
+
+        text = self.update_raw_text(text, image_prompts)
+        text_inputs = self.tokenizer(text, return_tensors=return_tensors)
+        return BatchFeature(data={**text_inputs, **preprocessed.data})
+
+    @staticmethod
+    def _extract_medias_from_messages(messages: list[dict]) -> list[dict]:
+        """Extract media items from messages in a single pass."""
+        medias = []
+        for msg in messages:
+            if msg['role'] != 'user' or not msg.get('content'):
+                continue
+
+            for content_part in msg['content']:
+                if not isinstance(content_part, dict):
+                    continue
+
+                content_type = content_part.get('type')
+                if content_type in ['image_url', 'image']:
+                    image_data = content_part.get(content_type)
+                    assert image_data is not None, f"image data is missing for content part: {content_part}"
+                    medias.append({
+                        'type': 'image',
+                        'image': image_data,
+                    })
+        return medias
+
+    def apply_chat_template(self, messages, **kwargs):
+        return self.tokenizer.apply_chat_template(messages, **kwargs)
+
+    def batch_decode(self, *args, **kwargs):
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    def decode(self, *args, **kwargs):
+        return self.tokenizer.decode(*args, **kwargs)
+
+    @property
+    def model_input_names(self):
+        return ['input_ids', 'attention_mask', 'pixel_values', 'grid_thws']

kimi_k3_vision_processing.py

@@ -0,0 +1,179 @@
+"""Image processor class for Kimi-K3.
+"""
+
+import json
+from typing import Any, Dict, Optional, Union
+
+import numpy as np
+import torch
+from PIL import Image
+from transformers.image_processing_utils import (BaseImageProcessor,
+                                                 BatchFeature)
+from transformers.utils import TensorType
+
+from .media_utils import (MediaInput, TransparentBgConfig, _to_tensor,
+                          ensure_media_type, image_to_np, navit_patchify,
+                          navit_resize_image, normalize)
+
+
+class KimiK3VisionProcessor(BaseImageProcessor):
+    model_type = "kimi_k3"
+
+    def __init__(
+        self,
+        media_proc_cfg: dict,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.media_proc_cfg = media_proc_cfg
+
+    @property
+    def _transparent_bg_config(self) -> Optional[TransparentBgConfig]:
+        cfg = self.media_proc_cfg.get("transparent_bg_config")
+        if cfg is None:
+            return None
+        if isinstance(cfg, TransparentBgConfig):
+            return cfg
+        return TransparentBgConfig(**cfg)
+
+    @property
+    def _transparent_bg_fill_stage(self) -> str:
+        return self.media_proc_cfg.get("transparent_bg_fill_stage",
+                                       "before_resize")
+
+    def media_tokens_calculator(self, media: MediaInput):
+        media = ensure_media_type(
+            media,
+            transparent_bg_config=self._transparent_bg_config,
+            transparent_bg_fill_stage=self._transparent_bg_fill_stage,
+        )
+        ret = self.get_resize_config(media)
+        return ret['num_tokens']
+
+    @classmethod
+    def make_image_prompt(cls, width: int, height: int) -> str:
+        """Build the K3 image placeholder with resolution info."""
+        return (f"<|media_begin|>image {width}x{height}"
+                f"<|media_content|><|media_pad|><|media_end|>")
+
+    def get_resize_config(self, media_input: MediaInput) -> dict:
+        if media_input['type'] == 'image':
+            w, h = media_input['image'].size
+            ret = navit_resize_image(
+                w, h, self.media_proc_cfg['patch_size'],
+                self.media_proc_cfg['merge_kernel_size'],
+                self.media_proc_cfg['in_patch_limit'],
+                self.media_proc_cfg['patch_limit_on_one_side'],
+                self.media_proc_cfg['fixed_output_tokens'])
+            return ret
+        else:
+            raise ValueError("Unsupported type: {}".format(
+                media_input['type']))
+
+    def resize_image(self, image: Image.Image, new_width: int, new_height: int,
+                     pad_width: int, pad_height: int) -> np.ndarray:
+        image_np = image_to_np(
+            image,
+            (new_width, new_height),
+            "resize",
+            transparent_bg_config=self._transparent_bg_config,
+            transparent_bg_fill_stage=self._transparent_bg_fill_stage,
+        )
+        image_np = np.pad(
+            image_np,
+            ((0, pad_height), (0, pad_width), (0, 0)),
+            mode="constant",
+            constant_values=0,
+        )
+        return image_np
+
+    def preprocess(
+        self,
+        medias: list[MediaInput],
+        return_tensors: Optional[Union[str, TensorType]] = None,
+    ) -> BatchFeature:
+        """
+        Preprocess a atom vision input (images) into model-ready tensors.
+
+        Args:
+            medias: List of MediaInput.
+            return_tensors: Desired output format ('pt', 'np', 'tf', or None).
+
+        Returns:
+            BatchFeature containing 'pixel_values' and 'grid_thws' tensors.
+        """
+        if not isinstance(medias, list):
+            medias = [medias]
+        if medias:
+            pixel_values = []
+            for item in medias:
+                item = ensure_media_type(
+                    item,
+                    transparent_bg_config=self._transparent_bg_config,
+                    transparent_bg_fill_stage=self._transparent_bg_fill_stage,
+                )
+                resize_config = self.get_resize_config(item)
+                new_width, new_height, pad_width, pad_height = resize_config[
+                    'new_width'], resize_config['new_height'], resize_config[
+                        'pad_width'], resize_config['pad_height']
+                if item['type'] == 'image':
+                    image = item['image']
+                    image_np = self.resize_image(image, new_width, new_height,
+                                                 pad_width, pad_height)
+                    pixel_values.append(np.expand_dims(image_np, axis=0))
+                else:
+                    raise ValueError("Unsupported type: {}".format(
+                        item['type']))
+            normalized_pixel_values = []
+            image_std_inv = 1.0 / np.array(self.media_proc_cfg['image_std'])
+            image_mean = np.array(self.media_proc_cfg['image_mean'])
+            for pixels in pixel_values:
+                pixels = normalize(pixels, image_mean, image_std_inv)
+                pixels_and_thw = navit_patchify(
+                    pixels,
+                    self.media_proc_cfg['patch_size'],
+                )
+                normalized_pixel_values.append(pixels_and_thw)
+
+            pixel_values = torch.cat([
+                _to_tensor(pixel_value['pixel_values'])
+                for pixel_value in normalized_pixel_values
+            ])
+            grid_thws = torch.cat([
+                _to_tensor(pixel_value['grid_thw'],
+                           dtype=torch.int64).unsqueeze(0)
+                for pixel_value in normalized_pixel_values
+            ])
+
+            data = {
+                'pixel_values': pixel_values,
+                'grid_thws': grid_thws,
+            }
+
+        else:
+            data = {}
+
+        return BatchFeature(data=data, tensor_type=return_tensors)
+
+    def __repr__(self):
+        return f"KimiK3VisionProcessor(media_proc_cfg={self.media_proc_cfg})"
+
+    def to_dict(self) -> Dict[str, Any]:
+        output = super().to_dict()
+        output["media_proc_cfg"] = self.media_proc_cfg
+        if "media_processor" in output:
+            del output["media_processor"]
+        return output
+
+    @classmethod
+    def from_dict(cls, config_dict: Dict[str, Any], **kwargs):
+        config = config_dict.copy()
+        media_proc_cfg = config.pop("media_proc_cfg", {})
+        return cls(media_proc_cfg=media_proc_cfg, **config, **kwargs)
+
+    def to_json_string(self):
+        dictionary = self.to_dict()
+        for key, value in dictionary.items():
+            if hasattr(value, 'tolist'):
+                dictionary[key] = value.tolist()
+        return json.dumps(dictionary, indent=2, sort_keys=True) + "\n"

media_utils.py

@@ -0,0 +1,376 @@
+import base64
+import functools
+import io
+import math
+from dataclasses import dataclass
+from typing import Literal, TypedDict
+
+import numpy as np
+from PIL import Image
+
+
+class ImageInput(TypedDict):
+    type: Literal['image']
+    image: Image.Image
+
+
+MediaInput = ImageInput
+
+
+@dataclass
+class TransparentBgConfig:
+    """The config of the transparent background."""
+
+    pattern: Literal["white", "black", "gray", "chessboard"] = "black"
+    """The pattern of the transparent background."""
+
+    chessboard_square_size: int = 16
+    """The size of the squares in the chessboard background."""
+
+    chessboard_square_on_top_left: bool = True
+    """Whether to start the chessboard with a white square on the top left."""
+
+    chessboard_white_value: int = 255
+    """The value of the white pixels in the background."""
+
+    chessboard_gray_value: int = 200
+    """The value of the gray pixels in the background."""
+
+
+@functools.lru_cache(maxsize=256)
+def _create_chessboard_background(
+    height: int,
+    width: int,
+    square_size: int,
+    square_on_top_left: bool,
+    white_value: int,
+    gray_value: int,
+) -> np.ndarray:
+    """Create a chessboard background."""
+    bg = np.ones((height, width, 3), dtype=np.uint8) * white_value
+    for y in range(0, height, square_size):
+        for x in range(0, width, square_size):
+            if (y // square_size + x // square_size) % 2 == (
+                    1 if square_on_top_left else 0):
+                bg[y:y + square_size, x:x + square_size] = gray_value
+    return bg
+
+
+def fill_transparent_bg_with(
+    image: Image.Image,
+    transparent_bg_config: TransparentBgConfig | None = None,
+) -> Image.Image:
+    """Composite a (possibly) transparent image onto a configured background.
+
+    When ``transparent_bg_config`` is ``None``, the image is simply converted
+    to RGB (preserving the historical behavior). Otherwise the alpha channel
+    is alpha-composited over a background generated according to the config.
+    """
+    if transparent_bg_config is None:
+        return image.convert("RGB")
+
+    if image.mode == "RGB":
+        return image
+
+    has_alpha = "A" in image.getbands() or "transparency" in image.info
+    if not has_alpha:
+        return image.convert("RGB")
+
+    img = np.array(image.convert("RGBA"))
+    height, width = img.shape[:2]
+    bg_pattern = transparent_bg_config.pattern
+    if bg_pattern == "white":
+        bg = np.full((height, width, 3), 255, dtype=np.uint8)
+    elif bg_pattern == "black":
+        bg = np.zeros((height, width, 3), dtype=np.uint8)
+    elif bg_pattern == "gray":
+        bg = np.full((height, width, 3), 128, dtype=np.uint8)
+    elif bg_pattern == "chessboard":
+        bg = _create_chessboard_background(
+            height,
+            width,
+            transparent_bg_config.chessboard_square_size,
+            transparent_bg_config.chessboard_square_on_top_left,
+            transparent_bg_config.chessboard_white_value,
+            transparent_bg_config.chessboard_gray_value,
+        )
+    else:
+        raise ValueError(f"Invalid background pattern: {bg_pattern}")
+
+    alpha = img[:, :, 3]
+    img_rgb = img[:, :, :3]
+    alpha_normalized = alpha.astype(np.float32) / 255.0
+    alpha_3d = np.stack([alpha_normalized] * 3, axis=2)
+    result = alpha_3d * img_rgb + (1 - alpha_3d) * bg
+    result = result.astype(np.uint8)
+    return Image.fromarray(result)
+
+
+def navit_resize_image(
+    width: int,
+    height: int,
+    patch_size: int,
+    merge_kernel_size: int,
+    in_patch_limit: int,
+    patch_limit_on_one_side: int,
+    fixed_output_tokens: int | None,
+):
+    # Apply the patch limits.
+    s1 = math.sqrt(
+        in_patch_limit /
+        (max(1.0, width // patch_size) * max(1.0, height // patch_size)))
+    s2 = patch_limit_on_one_side * patch_size / width
+    s3 = patch_limit_on_one_side * patch_size / height
+    scale = min(1.0, s1, s2, s3)
+    new_w, new_h = max(1, int(width * scale)), max(1, int(height * scale))
+    new_w = min(new_w, patch_limit_on_one_side * patch_size)
+    new_h = min(new_h, patch_limit_on_one_side * patch_size)
+
+    # Calculate the padding to make the height and width divisible by the merge kernel size and patch size.
+    factor = merge_kernel_size * patch_size
+
+    pad_height = (factor - new_h % factor) % factor
+    pad_width = (factor - new_w % factor) % factor
+
+    if fixed_output_tokens is not None:
+        num_tokens = fixed_output_tokens
+    else:
+        # Calculate new dimensions after padding and patching
+        token_height = (new_h + pad_height) // factor
+        token_width = (new_w + pad_width) // factor
+
+        assert token_height * merge_kernel_size <= patch_limit_on_one_side, (
+            f"token_height {token_height} * merge_kernel_size {merge_kernel_size} > patch_limit_on_one_side {patch_limit_on_one_side}"
+        )
+        assert token_width * merge_kernel_size <= patch_limit_on_one_side, (
+            f"token_width {token_width} * merge_kernel_size {merge_kernel_size} > patch_limit_on_one_side {patch_limit_on_one_side}"
+        )
+
+        num_tokens = token_height * token_width
+    return {
+        "num_tokens": num_tokens,
+        "new_width": new_w,
+        "new_height": new_h,
+        "pad_width": pad_width,
+        "pad_height": pad_height,
+        "sampled_nframes": 1,
+    }
+
+
+def _to_pil(
+    data: str | bytes | Image.Image,
+    transparent_bg_config: TransparentBgConfig | None = None,
+    to_rgb: bool = True,
+) -> Image.Image:
+    """Load an image and (optionally) composite its transparent background.
+
+    Args:
+        data: A PIL Image, a base64 ``data:`` URL, a file path, or raw bytes.
+        transparent_bg_config: The config used to fill the transparent
+            background. ``None`` keeps the historical behavior of converting
+            to RGB without compositing.
+        to_rgb: If ``False`` the image is returned as-is (the
+            ``transparent_bg_config`` is ignored). The caller is then
+            expected to call :func:`fill_transparent_bg_with` later — e.g.
+            after a resize.
+    """
+    if isinstance(data, Image.Image):
+        image = data
+    elif isinstance(data, str):
+        if data.startswith("data:"):
+            raw_base64 = data.split(",")[1]
+            image = Image.open(io.BytesIO(base64.b64decode(raw_base64)))
+        else:
+            image = Image.open(data)
+    elif isinstance(data, bytes):
+        image = Image.open(io.BytesIO(data))
+    else:
+        raise ValueError(f"Unsupported data type: {type(data)}")
+
+    if not to_rgb:
+        return image
+
+    return fill_transparent_bg_with(image, transparent_bg_config)
+
+
+def ensure_media_type(
+    media: MediaInput,
+    transparent_bg_config: TransparentBgConfig | None = None,
+    transparent_bg_fill_stage: Literal["before_resize",
+                                       "after_resize"] = "before_resize",
+) -> MediaInput:
+    if media['type'] == 'image':
+        media['image'] = _to_pil(
+            media['image'],
+            transparent_bg_config=transparent_bg_config,
+            to_rgb=transparent_bg_fill_stage == "before_resize",
+        )
+        return media
+    else:
+        raise ValueError(f"Unsupported media type: {media['type']}")
+
+
+def image_to_np(
+    image: Image.Image,
+    resize_to: tuple[int, int] | None = None,
+    mode: str = "resize",
+    raise_error_for_ill_resize: bool = True,
+    transparent_bg_config: TransparentBgConfig | None = None,
+    transparent_bg_fill_stage: Literal["before_resize",
+                                       "after_resize"] = "before_resize",
+) -> np.ndarray:
+    """Convert an image to a numpy array.
+
+    Args:
+        content: The image to convert.
+        resize_to: The size to resize the image to.
+        mode: The mode to resize the image to.
+        raise_error_for_ill_resize: Whether to raise an error for ill-sized resize.
+        transparent_bg_config: The config of the transparent background. Only
+            used when ``transparent_bg_fill_stage == "after_resize"`` (the
+            caller is responsible for filling before resize otherwise).
+        transparent_bg_fill_stage: When to composite the transparent
+            background — before or after the resize step.
+
+    Returns:
+        A numpy array.
+    """
+    assert isinstance(image, Image.Image), "image must be a PIL Image"
+    if resize_to is not None:
+        if mode == "resize":
+            image = image.resize(resize_to, resample=Image.Resampling.BICUBIC)
+            if transparent_bg_fill_stage == "after_resize":
+                image = fill_transparent_bg_with(image, transparent_bg_config)
+
+        elif mode == "rescale_and_pad_to_center":
+            scale = min(resize_to[0] / image.width,
+                        resize_to[1] / image.height, 1.0)
+            new_width = round(image.width * scale)
+            new_height = round(image.height * scale)
+            if new_width == 0 or new_height == 0:
+                if raise_error_for_ill_resize:
+                    raise ValueError(
+                        f"Invalid resize to: {resize_to}, from image size: {image.size}"
+                    )
+                else:
+                    return np.zeros((resize_to[1], resize_to[0], 3),
+                                    dtype=np.uint8)
+
+            image = image.resize((new_width, new_height),
+                                 resample=Image.Resampling.BICUBIC)
+            if transparent_bg_fill_stage == "after_resize":
+                image = fill_transparent_bg_with(image, transparent_bg_config)
+            padding_left = (resize_to[0] - new_width) // 2
+            padding_right = resize_to[0] - new_width - padding_left
+            padding_top = (resize_to[1] - new_height) // 2
+            padding_bottom = resize_to[1] - new_height - padding_top
+            image = np.asarray(image)
+            image = np.pad(
+                image,
+                ((padding_top, padding_bottom), (padding_left, padding_right),
+                 (0, 0)),
+                mode="constant",
+                constant_values=0,
+            )
+            assert image.shape == (resize_to[1], resize_to[0], 3)
+
+        elif mode == "rescale_and_pad_to_rightbottom":
+            scale = min(resize_to[0] / image.width,
+                        resize_to[1] / image.height, 1.0)
+            new_width = round(image.width * scale)
+            new_height = round(image.height * scale)
+            if new_width == 0 or new_height == 0:
+                if raise_error_for_ill_resize:
+                    raise ValueError(
+                        f"Invalid resize to: {resize_to}, from image size: {image.size}"
+                    )
+                else:
+                    return np.zeros((resize_to[1], resize_to[0], 3),
+                                    dtype=np.uint8)
+
+            image = image.resize((new_width, new_height),
+                                 resample=Image.Resampling.BICUBIC)
+            if transparent_bg_fill_stage == "after_resize":
+                image = fill_transparent_bg_with(image, transparent_bg_config)
+            padding_right = resize_to[0] - new_width
+            padding_bottom = resize_to[1] - new_height
+            image = np.asarray(image)
+            image = np.pad(
+                image,
+                ((0, padding_bottom), (0, padding_right), (0, 0)),
+                mode="constant",
+                constant_values=0,
+            )
+            assert image.shape == (resize_to[1], resize_to[0], 3)
+
+        else:
+            raise ValueError(f"Invalid mode: {mode}")
+
+    if isinstance(image, Image.Image):
+        return np.asarray(image)
+    else:
+        return image
+
+
+def navit_patchify(pixel_values: np.ndarray,
+                   patch_size: int) -> dict[str, np.ndarray]:
+    """Reshape the pixel values to a navit shape.
+
+    Args:
+        pixel_values: np.ndarray, shape (t, h, w, c)
+        patch_size: int
+
+    Returns:
+        dict[str, np.ndarray]
+        - patches: np.ndarray, shape (t * h//patch_size * w//patch_size, c, patch_size, patch_size)
+        - grid_thw: np.ndarray, (t, h//patch_size, w//patch_size)
+    """
+    T, H, W, C = pixel_values.shape
+    assert C == 3, "pixel_values must have 3 channels"
+
+    patches = pixel_values.reshape(T, H // patch_size, patch_size,
+                                   W // patch_size, patch_size, C)
+    # (T, H//patch_size, W//patch_size, C, patch_size, patch_size)
+    patches = patches.transpose(0, 1, 3, 5, 2, 4)
+    patches = patches.reshape(-1, C, patch_size, patch_size)
+    grid_thw = np.array([T, H // patch_size, W // patch_size])
+    return {"pixel_values": patches, "grid_thw": grid_thw}
+
+
+def normalize(x: np.ndarray,
+              mean,
+              std_inv,
+              pixels_dtype: np.dtype = np.float32) -> np.ndarray:
+    """Normalize the image.
+
+    Args:
+        x: The image to normalize. The shape is (..., 3). The dtype is uint8. The range is [0, 255].
+        mean: The mean of the image.
+        std_inv: The inverse of the std of the image.
+        pixels_dtype: The dtype of the image.
+    Returns:
+        The normalized image. The shape is (..., 3). The dtype is determined by the pixels_dtype.
+    """
+    x = (x / 255.0).astype(pixels_dtype)
+    x -= mean
+    x *= std_inv
+    return x
+
+
+def _to_tensor(data, **kwargs):
+    import torch
+
+    if isinstance(data, np.ndarray):
+        return torch.from_numpy(data).to(**kwargs)
+    elif isinstance(data, torch.Tensor):
+        return data.to(**kwargs)
+    elif isinstance(data, list):
+        return [_to_tensor(item, **kwargs) for item in data]
+    elif isinstance(data, tuple):
+        return tuple(_to_tensor(item, **kwargs) for item in data)
+    elif isinstance(data, dict):
+        return {k: _to_tensor(v, **kwargs) for k, v in data.items()}
+    elif data is None:
+        return None
+    else:
+        raise ValueError(f"Unsupported data type: {type(data)}")

model-00001-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d1a60ccb44d2c70085717a0a0100fe2a950460ae95810a53e5a4324f057f9bd9
+size 4936231310

model-00002-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:64d80ecd05bd201eae321c4586090340d42ade4b9442668b02e063afb8e95d95
+size 5240783442

model-00003-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:726818a99125faa1d487a620f2c305a73510d05b6520a1ac73412a34706a00f5
+size 3645759823

model-00004-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:93fb29164cc8cbbc280a737a9f48e0f96f6ff01a476d9c952e59e088a9ee19e7
+size 5240783440

model-00005-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bfcbf3605e7342e25b905352424e15b10edf3f84712da07c59f48305d9670c09
+size 3473372860

model-00006-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0ee31b98a9f4cfe1613546d5765a1035a1df5153e6fb1ab193bb62b63cc28215
+size 5240783442

model-00007-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dddba61cee4c9d21983a860c8307e1c0b7a1e4fb3ec99c20397b64ca7052a3f9
+size 3645759841

model-00008-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48c711e0a5e1f564ec80a3f4c04223a5854ac6748098bb1242fb3c1612b143cf
+size 5240783442

model-00009-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc8a5cce236a8e74d99e1db3a7001183ab75d1fcc25254432b29294e55e6426c
+size 3645759791

model-00010-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bbab4be97bff48cac50ef176f410fb00e67dc6a6856bb2b7c791a12d270aaf09
+size 5240783442

model-00011-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f3531c5d8094cd2f4f47a08fd6fb9a7d5b0120b0cbb7d39dbd5261875c3e757
+size 3645759843

model-00012-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:04f0a9182d8ece493a78d0f24f2ba76c92bd63c5773e4a64f3c489894aee30db
+size 5240783442

model-00013-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2dd1fce85d0092395b468b1ea9317814b8a16ea484e1052e58178e121017d4d5
+size 3473372798

model-00014-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b5cfbeba263044db8170129b3f81b489e23541f6fcc89e59e4f7cec5dd82b1b7
+size 5240783442

model-00015-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ddc5381e8674b0da7d51109b48e7407a9ebc2b6192421a0739cd579ac43ea865
+size 3645759831

model-00016-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f912afd3d01adf32a7e7732a6b89ca22b63eccd9584b93115bae5bb180ff0109
+size 5240783442

model-00017-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:68fe5808be766e7b87bfb11ac8da91d1941b76e5996aaed7b6df2a17e5f65c01
+size 3645759841

model-00018-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5b4a892622abfed3c92bab16461de495cfc5d51cbbf395a3cd35980b4b204245
+size 5240783442

model-00019-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf62390fbc5fae3a9073ca8d50b35380721374f912da372c8ce0fde92caab870
+size 3645759871

model-00020-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:555bd270de301a64e523758785b192c16a65d8d26e4d294592b755e2f18c36f9
+size 5240783444

model-00021-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9a599037e73adcadaac5cff3719c8be06264b7c6aecf41a0c532a2e3b834c634
+size 3473372912

model-00022-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:12c245aec45b5047995b29c3bb8eb62e80a6654340cd5b52edb031999b5c4b4f
+size 5240783446

model-00023-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf701db169d3522a2b40d5af86ed9c1c6208c8607697d7b0032467e66467852e
+size 3645759894

model-00024-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0d67ce61bc274ffa289a3f07136bfb751a49747ed2c6cf47dbd624c407c5398
+size 5240783446

model-00025-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96ef21f857bba47bddef11522183a58ab8309f236ed711dd7d71b58eb749626b
+size 3645759820

model-00026-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a1b0f53d75af4a74e24c5181cc92822a49a9166c1333d49d2725393d81d7510
+size 5240783446

model-00027-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b7da5b618e2f5efc06033eedfef58943e1bf5b4271698b590f2c3129631321d7
+size 3645759816

model-00028-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e4466b8b7b5beabfd21af28a387e0cf6b194cecd845dcdab9e323cfe8ae15802
+size 5240783442

model-00029-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:76335537d6c7bc6d5e3c8ced7e7bdbee6a66bca0a241e3816f521bdfb402f85e
+size 3473372910

model-00030-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:221364c03a3f92a55ad44180269993b6e423fd99403ce1d35d812c56d89203c7
+size 5240783446

model-00031-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:91eeff8ee44d5e6389b7b9fe730b1f880ae53e17b9d2d087d295460f44be07d5
+size 3645759808

model-00032-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1bddbb57aedc66efc1d63b79bd2d0c203f8086ee6215862ce63e5b5046d6e30f
+size 5240783446

model-00033-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:911bc1259e5672b7dc4d565f1ef5c9b9229019fa0044334113850727396a43cb
+size 3645759898

model-00034-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e59bd8c1ffecdf3effba49609389b6cce2246b54464ee3f1ae481a046be16c6c
+size 5240783446

model-00035-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f5b86f3b55bff7f005c2911e92bebe70a7e4c0cac3a076758f45f0131529bf5a
+size 3645759894

model-00036-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e58c766c07e49851dab73a896fe530dc8dabac60d73b2ff637c7c7495cb85f9
+size 5240783446

model-00037-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:13b79e949543ad129f22db48c59e68edd1b4260142051bac165f196d499e1a47
+size 3473372912

model-00038-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ef490d649a1e97523688ba123b3caa589c883a9df2614f325fde96cd25cdd360
+size 5240783446

model-00039-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5d40e9e62be243c7bddf5c37b5d4b379a490a5b88b1cc850dff7d554bf46394f
+size 3645759814

model-00040-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1daa5d84bc3d5b568509ac86c477c8f02b45bc7d7e4e305b0b034ca3288bacf9
+size 5240783446

model-00041-of-00185.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7aeac53f439fb92aee6a380def5513bee59653549b42a83baf4ee19fc7a1ffdd
+size 3645759896