Add files using upload-large-folder tool

README.md

@@ -0,0 +1,95 @@
+---
+license: mit
+base_model:
+- inference-optimization/Kimi-K3-0.40B
+library_name: compressed-tensors
+tags:
+- quantized
+- mxfp4
+---
+
+# Kimi-K3-0.40B-MXFP4
+
+This is an MXFP4-quantized version of [inference-optimization/Kimi-K3-0.40B](https://huggingface.co/inference-optimization/Kimi-K3-0.40B), a tiny model derived from [moonshotai/Kimi-K3](https://huggingface.co/moonshotai/Kimi-K3). Created for testing and development.
+
+## Model Details
+
+- **Base Model**: inference-optimization/Kimi-K3-0.40B
+- **Architecture**: kimi_k3
+- **Total Parameters**: 0.40B
+- **Activated Parameters**: ~0.22B (MoE: 2 of 8 experts active per token, plus 1 shared expert)
+- **Quantization**: W4A16 MXFP4 (`mxfp4-pack-quantized`), group size 32
+
+## Quantization Config
+
+Matches the quantization scheme used in [moonshotai/Kimi-K3](https://huggingface.co/moonshotai/Kimi-K3/blob/main/config.json):
+
+| Field | Value |
+|---|---|
+| Format | `mxfp4-pack-quantized` |
+| Weights | 4-bit float, group_size=32, minmax observer |
+| Scale dtype | `torch.uint8` |
+| Activations | unquantized (W4A16) |
+| Ignored layers | `self_attn`, `shared_experts`, `lm_head`, `vision_tower` |
+
+## Usage
+
+```python
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from compressed_tensors.offload import dispatch_model
+
+model = AutoModelForCausalLM.from_pretrained(
+    "inference-optimization/Kimi-K3-0.40B-MXFP4",
+    trust_remote_code=True,
+    dtype=torch.bfloat16,
+)
+tokenizer = AutoTokenizer.from_pretrained(
+    "inference-optimization/Kimi-K3-0.40B-MXFP4",
+    trust_remote_code=True,
+)
+dispatch_model(model)
+
+sample = tokenizer("Hello my name is", return_tensors="pt")
+sample = {k: v.to(model.device) for k, v in sample.items()}
+output = model.generate(
+    **sample,
+    max_new_tokens=100,
+    eos_token_id=tokenizer.eos_token_id,
+    pad_token_id=tokenizer.pad_token_id,
+)
+print(tokenizer.decode(output[0].tolist(), skip_special_tokens=True))
+```
+
+## Creation Process
+
+Quantized using [llm-compressor](https://github.com/vllm-project/llm-compressor):
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from llmcompressor import oneshot
+from llmcompressor.modifiers.quantization import QuantizationModifier
+
+MODEL_ID = "inference-optimization/Kimi-K3-0.40B"
+model = AutoModelForCausalLM.from_pretrained(MODEL_ID, trust_remote_code=True)
+tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
+
+recipe = QuantizationModifier(
+    targets="Linear",
+    scheme="MXFP4A16",
+    ignore=[
+        "re:.*self_attn.*",
+        "re:.*shared_experts.*",
+        "re:.*lm_head.*",
+        "re:.*vision_tower.*",
+    ],
+)
+oneshot(model=model, recipe=recipe)
+model.save_pretrained(SAVE_DIR, save_compressed=True)
+tokenizer.save_pretrained(SAVE_DIR)
+```
+
+## Notes
+
+- `trust_remote_code=True` is required to load the custom modeling files.
+- Load with `dtype=torch.bfloat16` to match the decompressed weight dtype.

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
+}

config.json

@@ -0,0 +1,153 @@
+{
+  "architectures": [
+    "KimiK3ForConditionalGeneration"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_kimi_k3.KimiK3Config",
+    "AutoModel": "modeling_kimi_k3.KimiK3ForConditionalGeneration",
+    "AutoModelForCausalLM": "modeling_kimi_k3.KimiK3ForConditionalGeneration"
+  },
+  "dtype": "float32",
+  "ignore_index": -100,
+  "media_placeholder_token_id": 163605,
+  "model_type": "kimi_k3",
+  "pad_token_id": 0,
+  "text_config": {
+    "_name_or_path": "",
+    "activation_situ_beta": 4.0,
+    "activation_situ_linear_beta": 25.0,
+    "architectures": null,
+    "attn_res_block_size": 4,
+    "auto_map": {
+      "AutoConfig": "configuration_kimi_k3.KimiLinearConfig",
+      "AutoModel": "modeling_kimi_linear.KimiLinearModel",
+      "AutoModelForCausalLM": "modeling_kimi_linear.KimiLinearForCausalLM"
+    },
+    "bos_token_id": 1,
+    "chunk_size_feed_forward": 0,
+    "dtype": null,
+    "eos_token_id": 2,
+    "first_k_dense_replace": 1,
+    "head_dim": 74,
+    "hidden_act": "situ",
+    "hidden_size": 1024,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "initializer_range": 0.02,
+    "intermediate_size": 2048,
+    "is_encoder_decoder": false,
+    "kv_lora_rank": 128,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "latent_moe_use_norm": true,
+    "linear_attn_config": {
+      "full_attn_layers": [
+        4,
+        8
+      ],
+      "head_dim": 32,
+      "kda_layers": [
+        1,
+        2,
+        3,
+        5,
+        6,
+        7
+      ],
+      "num_heads": 8,
+      "short_conv_kernel_size": 4,
+      "use_full_rank_gate": true
+    },
+    "max_position_embeddings": 4096,
+    "mla_use_nope": true,
+    "mla_use_output_gate": true,
+    "model_type": "kimi_linear",
+    "moe_intermediate_size": 256,
+    "moe_layer_freq": 1,
+    "moe_renormalize": true,
+    "moe_router_activation_func": "sigmoid",
+    "num_attention_heads": 8,
+    "num_expert_group": 1,
+    "num_experts": 8,
+    "num_experts_per_token": 2,
+    "num_hidden_layers": 8,
+    "num_key_value_heads": 8,
+    "num_nextn_predict_layers": 0,
+    "num_shared_experts": 1,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "pad_token_id": 0,
+    "problem_type": null,
+    "q_lora_rank": 256,
+    "qk_nope_head_dim": 64,
+    "qk_rope_head_dim": 32,
+    "return_dict": true,
+    "rms_norm_eps": 1e-05,
+    "rope_parameters": {
+      "rope_theta": 10000.0,
+      "rope_type": "default"
+    },
+    "rope_theta": 10000.0,
+    "routed_expert_hidden_size": 512,
+    "routed_scaling_factor": 1.0,
+    "tie_word_embeddings": false,
+    "topk_group": 1,
+    "topk_method": "noaux_tc",
+    "use_cache": true,
+    "use_grouped_topk": true,
+    "v_head_dim": 64,
+    "vocab_size": 163840
+  },
+  "transformers_version": "5.15.0.dev0",
+  "vision_config": {
+    "_name_or_path": "",
+    "activation_func": "gelu_pytorch_tanh",
+    "architectures": null,
+    "attn_bias": false,
+    "chunk_size_feed_forward": 0,
+    "dtype": null,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "init_pos_emb_height": 64,
+    "init_pos_emb_time": 4,
+    "init_pos_emb_width": 64,
+    "is_encoder_decoder": false,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "linear_bias": false,
+    "merge_kernel_size": [
+      2,
+      2
+    ],
+    "merge_type": "sd2_tpool",
+    "mlp_type": "mlp2",
+    "mm_hidden_size": 256,
+    "mm_projector_type": "patchmergerv2",
+    "model_type": "",
+    "norm_type": "rmsnorm",
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "patch_embed_proj_bias": false,
+    "patch_size": 14,
+    "pos_emb_interpolation_mode": "bilinear",
+    "pos_emb_type": "divided_fixed",
+    "problem_type": null,
+    "projector_hidden_act": "gelu",
+    "projector_ln_eps": 1e-05,
+    "qkv_hidden_size": 1536,
+    "return_dict": true,
+    "text_hidden_size": 1024,
+    "vt_hidden_size": 256,
+    "vt_intermediate_size": 512,
+    "vt_num_attention_heads": 4,
+    "vt_num_hidden_layers": 2
+  }
+}

configuration_kimi_k3.py

@@ -0,0 +1,286 @@
+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 +1,651 @@
+"""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 +1,10 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "output_attentions": false,
+  "output_hidden_states": false,
+  "pad_token_id": 0,
+  "transformers_version": "5.15.0.dev0",
+  "use_cache": false
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a39b7ed2769ee2f9891a19cef6f1ca7986a3f295c09759b5b3135285e5d7a678
+size 1582323896

modeling_kimi_k3.py

@@ -0,0 +1,1355 @@
+# coding=utf-8
+# Copyright 2025-2026 The Moonshot AI Team and HuggingFace Inc. team. All rights reserved.
+#
+# The code is based on llava (llava/modeling_llava.py), but modified for Kimi-K3.
+#
+# Licensing Information:
+# - Code derived from llava (llava/modeling_llava.py) is licensed under the Apache License, Version 2.0.
+# - Other parts of the code are licensed under the Kimi K3 License (see the LICENSE file in this repository).
+#
+# Apache License, Version 2.0:
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+# NOTE: Reference implementation for model architecture; see the model card for production deployment.
+import math
+from collections.abc import Sequence
+from copy import deepcopy
+from typing import Optional
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import activations
+
+try:
+    from transformers.activations import PytorchGELUTanh
+except ImportError:
+    from transformers.activations import GELUTanh
+
+    activations.PytorchGELUTanh = GELUTanh
+    PytorchGELUTanh = GELUTanh
+from transformers.activations import PytorchGELUTanh
+from transformers.configuration_utils import PretrainedConfig
+from transformers.modeling_utils import PreTrainedModel
+from transformers.models.llava.modeling_llava import LlavaCausalLMOutputWithPast
+from transformers.utils import is_flash_attn_2_available
+
+from .configuration_kimi_k3 import KimiK3Config
+from .modeling_kimi_linear import KimiLinearForCausalLM
+
+# Flash attention imports
+if is_flash_attn_2_available():
+    from flash_attn import flash_attn_varlen_func
+else:
+    flash_attn_varlen_func = None
+
+
+def multihead_attention(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    q_cu_seqlens: torch.Tensor | None = None,
+    k_cu_seqlens: torch.Tensor | None = None,
+    max_seqlen_q: int | None = None,
+    max_seqlen_k: int | None = None,
+    deterministic: bool = False,
+):
+    """Multi-head attention using flash attention 2.
+
+    Args:
+        q, k, v: tensor of shape (batch_size, seqlen, num_heads, head_dim),
+            or (tot_seqlens, num_heads, head_dim) if packing.
+        q_cu_seqlens (torch.Tensor): cumulative sequence lengths of q.
+            The first element should be 0 and the last element should be q.shape[0].
+        k_cu_seqlens (torch.Tensor): cumulative sequence lengths of k.
+            The first element should be 0 and the last element should be k.shape[0].
+
+    Returns:
+        output: shape (batch_size, seqlen, dim) or (tot_seqlens, dim) if packing,
+            where dim = num_heads * head_dim
+    """
+    attn_out = flash_attn_varlen_func(
+        q,
+        k,
+        v,
+        q_cu_seqlens,
+        k_cu_seqlens,
+        max_seqlen_q,
+        max_seqlen_k,
+        causal=False,
+        deterministic=deterministic,
+    )
+    if isinstance(attn_out, tuple):
+        attn_out = attn_out[0]
+
+    attn_out = attn_out.flatten(start_dim=-2)
+
+    return attn_out
+
+
+def eager_attention(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    q_cu_seqlens: Optional[torch.Tensor] = None,
+    k_cu_seqlens: Optional[torch.Tensor] = None,
+    **kwargs,
+) -> torch.Tensor:
+    seq_length = q.shape[0]
+    attention_mask = torch.zeros(
+        [1, seq_length, seq_length], device=q.device, dtype=torch.bool
+    )
+    for i in range(1, len(q_cu_seqlens)):
+        attention_mask[
+            ...,
+            q_cu_seqlens[i - 1] : q_cu_seqlens[i],
+            q_cu_seqlens[i - 1] : q_cu_seqlens[i],
+        ] = True
+    q = q.transpose(0, 1)
+    k = k.transpose(0, 1)
+    v = v.transpose(0, 1)
+
+    attn_weight = q @ k.transpose(-2, -1) / math.sqrt(q.shape[-1])
+    attn_weight = attn_weight.masked_fill(
+        ~attention_mask, torch.finfo(attn_weight.dtype).min
+    )
+    attn_weight = torch.softmax(attn_weight, dim=-1, dtype=torch.float32).to(q.dtype)
+
+    attn_output = attn_weight @ v
+    attn_output = attn_output.transpose(0, 1)
+    attn_output = attn_output.reshape(seq_length, -1)
+    return attn_output
+
+
+VL_VISION_ATTENTION_FUNCTIONS = {
+    "flash_attention_2": multihead_attention,
+    "eager": eager_attention,
+}
+
+
+def _apply_rope_input_validation(x, freqs_cis):
+    assert x.ndim == freqs_cis.ndim + 1, (x.shape, freqs_cis.shape)
+    assert x.shape[:-2] == freqs_cis.shape[:-1], (x.shape, freqs_cis.shape)
+    assert x.shape[-1] == 2 * freqs_cis.shape[-1], (x.shape, freqs_cis.shape)
+    assert freqs_cis.dtype == torch.complex64, freqs_cis.dtype
+
+
+def get_rope_shape_decorate(func):
+    _get_rope_shape_first_call_flag = set()
+
+    def wrapper(org, interpolation_mode, shape):
+        key = (org.requires_grad, torch.is_grad_enabled(), interpolation_mode)
+        if key not in _get_rope_shape_first_call_flag:
+            _get_rope_shape_first_call_flag.add(key)
+            _ = func(org, interpolation_mode, shape=(64, 64))
+        return func(org, interpolation_mode, shape)
+
+    return wrapper
+
+
+@get_rope_shape_decorate
+@torch.compile(dynamic=True)
+def get_rope_shape(org, interpolation_mode, shape):
+    return (
+        F.interpolate(
+            org.permute((2, 0, 1)).unsqueeze(0),
+            size=shape,
+            mode=interpolation_mode,
+        )
+        .squeeze(0)
+        .permute((1, 2, 0))
+        .flatten(end_dim=1)
+    )
+
+
+def apply_rope(
+    xq: torch.Tensor, xk: torch.Tensor, freqs_cis: torch.Tensor
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """
+    Args: (The leading dimensions of all inputs should be the same)
+        xq: query, tensor of shape (..., num_heads, head_dim)
+        xk: key, tensor of shape (..., num_heads, head_dim)
+        freqs_cis: tensor of shape (..., head_dim/2), dtype=torch.complex64. It contains the precomputed cis(freqs) for each position in the 2D grid.
+    Returns:
+        xq_out, xk_out: tensors of shape (..., num_heads, head_dim)
+    """
+    _apply_rope_input_validation(xq, freqs_cis)
+    _apply_rope_input_validation(xk, freqs_cis)
+
+    freqs_cis = freqs_cis.unsqueeze(-2)  # ..., 1, head_dim/2
+    # ..., num_heads, head_dim/2
+    xq_ = torch.view_as_complex(xq.float().view(*xq.shape[:-1], -1, 2))
+    xk_ = torch.view_as_complex(xk.float().view(*xq.shape[:-1], -1, 2))
+    xq_out = torch.view_as_real(xq_ * freqs_cis).flatten(-2)  # ..., num_heads, head_dim
+    xk_out = torch.view_as_real(xk_ * freqs_cis).flatten(-2)  # ..., num_heads, head_dim
+    return xq_out.type_as(xq), xk_out.type_as(xk)
+
+
+def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
+    """
+    From:
+    https://github.com/OpenGVLab/InternVideo/blob/421f6d2361fc8f61a3394244571f2601a4e99e29/InternVideo2/multi_modality/models/backbones/internvideo2/pos_embed.py#L86
+    embed_dim: output dimension for each position
+    pos: a list of positions to be encoded: size (M,)
+    out: (M, D)
+    """
+    assert embed_dim % 2 == 0
+    omega = np.arange(embed_dim // 2, dtype=np.float32)
+    omega /= embed_dim / 2.0
+    omega = 1.0 / 10000**omega  # (D/2,)
+
+    pos = pos.reshape(-1)  # (M,)
+    out = np.einsum("m,d->md", pos, omega)  # (M, D/2), outer product
+
+    emb_sin = np.sin(out)  # (M, D/2)
+    emb_cos = np.cos(out)  # (M, D/2)
+
+    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
+    return emb
+
+
+def get_1d_sincos_pos_embed(embed_dim, t_size, cls_token=False):
+    """
+    t_size: int of the temporal size
+    return:
+    pos_embed: [t_size, embed_dim] or [1+t_size, embed_dim] (w/ or w/o cls_token)
+    """
+    grid_t = np.arange(t_size, dtype=np.float32)
+    pos_embed = get_1d_sincos_pos_embed_from_grid(embed_dim, grid_t)
+    if cls_token:
+        pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed], axis=0)
+    return pos_embed
+
+
+class Learnable2DInterpPosEmbDivided_fixed(nn.Module):
+    def __init__(
+        self,
+        height: int,
+        width: int,
+        num_frames: int,
+        dim: int,
+        interpolation_mode: str = "bicubic",
+    ) -> None:
+        super().__init__()
+        self.height = height
+        self.width = width
+        self.num_frames = num_frames
+        self.dim = dim
+        self.interpolation_mode = interpolation_mode
+        self.weight = nn.Parameter(torch.empty(height, width, dim))
+        self.register_buffer(
+            "time_weight",
+            torch.from_numpy(get_1d_sincos_pos_embed(self.dim, self.num_frames))
+            .float()
+            .unsqueeze(1),
+            persistent=False,
+        )
+
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        nn.init.normal_(self.weight)
+
+    def forward(self, x: torch.Tensor, grid_thws: torch.Tensor) -> torch.Tensor:
+        pos_embs = []
+        for t, h, w in grid_thws.tolist():
+            assert t <= self.num_frames, f"t:{t} > self.num_frames:{self.num_frames}"
+            if (h, w) == self.weight.shape[:-1]:
+                pos_emb_2d = self.weight.flatten(end_dim=1)
+            else:
+                pos_emb_2d = get_rope_shape(
+                    self.weight,
+                    interpolation_mode=self.interpolation_mode,
+                    shape=(h, w),
+                )
+
+            if t == 1:
+                pos_emb_3d = pos_emb_2d
+            else:
+                pos_emb_3d = (
+                    pos_emb_2d.unsqueeze(0).repeat(t, 1, 1) + self.time_weight[0:t]
+                )
+
+            pos_embs.append(pos_emb_3d.reshape(-1, pos_emb_3d.shape[-1]))
+
+        out = x + torch.cat(pos_embs)
+        return out
+
+
+class MoonVision3dPatchEmbed(nn.Module):
+    def __init__(
+        self,
+        out_dim: int,
+        in_dim: int = 3,
+        patch_size: int | tuple[int, int] = (14, 14),
+        pos_emb_height: int = 14,
+        pos_emb_width: int = 14,
+        pos_emb_time: int = 4,
+        pos_emb_type: str = "divided_fixed",
+        patch_embed_proj_bias: bool = True,
+        pos_emb_interpolation_mode: str = "bicubic",
+    ):
+        super().__init__()
+        assert isinstance(
+            patch_size, int | Sequence
+        ), f"Invalid patch_size type: {type(patch_size)}"
+        if isinstance(patch_size, int):
+            patch_size = (patch_size, patch_size)
+        assert (
+            len(patch_size) == 2
+        ), f"Expected patch_size to be a tuple of 2, got {patch_size}"
+        self.patch_size = patch_size
+
+        self.proj = nn.Conv2d(
+            in_dim,
+            out_dim,
+            kernel_size=patch_size,
+            stride=patch_size,
+            bias=patch_embed_proj_bias,
+        )
+
+        if pos_emb_type == "divided_fixed":
+            self.pos_emb = Learnable2DInterpPosEmbDivided_fixed(
+                height=pos_emb_height,
+                width=pos_emb_width,
+                num_frames=pos_emb_time,
+                dim=out_dim,
+                interpolation_mode=pos_emb_interpolation_mode,
+            )
+        else:
+            raise NotImplementedError(f"Not support pos_emb_type: {pos_emb_type}")
+
+    def forward(self, x: torch.Tensor, grid_thws: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+            x (L, Channels): input tensor
+            grid_hws (N, 3): temporal, height and width
+
+        Returns:
+            (L, Cout) tensor
+        """
+        x = self.proj(x).view(x.size(0), -1)
+        # apply positional embedding
+        x = self.pos_emb(x, grid_thws)
+        return x
+
+
+class Rope2DPosEmbRepeated(nn.Module):
+    """2D rotary position embedding with multi-resolution support.
+
+    This class is intended to be used in the following way:
+    1. Before training, create an instance of Rope2DPosEmb. This instance will hold the precomputed cis.
+    2. Before each forward pass, call `get_freqs_cis_by_*` to get the `freqs_cis` tensor for this iteration.
+    3. During the forward pass, pass the `freqs_cis` tensor to each attention layer, and call `apply` just before each attention operation.
+        The rope is shared across all attention layers and all heads.
+
+    Refs:
+    - RoFormer: https://arxiv.org/abs/2104.09864
+    - VisionLLaMA: https://arxiv.org/abs/2403.00522
+    - https://github.com/Meituan-AutoML/VisionLLaMA/blob/main/dit/models.py
+
+    Args:
+        dim (int): usually the multi-head attention dimension, should be divisible by 4 (TODO: relax this constraint if needed)
+        max_height (int): the maximum height of the 2D grid
+        max_width (int): the maximum width of the 2D grid
+        theta_base (float): the base of the theta
+        device (str): the device to store the precomputed cis
+    """
+
+    def __init__(self, dim: int, max_height: int, max_width: int, theta_base=10000):
+        super().__init__()
+        self.dim = dim
+        assert self.dim % 4 == 0, "dim must be divisible by 4"
+        self.max_height = max_height
+        self.max_width = max_width
+        self.theta_base = theta_base
+
+    def extra_repr(self):
+        return f"dim={self.dim}, max_height={self.max_height}, max_width={self.max_width}, theta_base={self.theta_base}"
+
+    def _precompute_freqs_cis(self, device: torch.device) -> torch.Tensor:
+        """Calculate the cis(freqs) for each position in the 2D grid.
+
+        Return: complex tensor of shape (max_height, max_width, dim//2) and value:
+            height axis: ret[h, w, 2*i] = cis(h * theta_base**(-4*i/dim))
+            weight axis: ret[h, w, 2*i+1] = cis(w * theta_base**(-4*i/dim))   with (i in [0, dim//4))
+            note: `cis` is a mathematical notation defined by cis x = cos x + i sin x,
+        """
+        N = self.max_height * self.max_width
+        flat_pos = torch.arange(0, N).float().to(device)
+        x_pos = flat_pos % self.max_width
+        y_pos = flat_pos // self.max_width
+        dim_range = (
+            torch.arange(0, self.dim, 4)[: (self.dim // 4)].float().to(device)
+        )  # C/4
+        freqs = 1.0 / (self.theta_base ** (dim_range / self.dim))
+        x_freqs = torch.outer(x_pos, freqs).float()  # N, C/4
+        y_freqs = torch.outer(y_pos, freqs).float()  # N, C/4
+        x_cis = torch.polar(torch.ones_like(x_freqs), x_freqs)  # N, C/4
+        y_cis = torch.polar(torch.ones_like(y_freqs), y_freqs)  # N, C/4
+        # N, C/4, 2
+        freqs_cis = torch.cat(
+            [x_cis.unsqueeze(dim=-1), y_cis.unsqueeze(dim=-1)], dim=-1
+        )
+        # max_height, max_width, C/2
+        freqs_cis = freqs_cis.reshape(self.max_height, self.max_width, -1)
+        return freqs_cis
+
+    def get_freqs_cis(
+        self, grid_thws: torch.Tensor, device: torch.device
+    ) -> torch.Tensor:
+        """
+        Args:
+            grid_thws (torch.Tensor): grid time, height and width
+
+        Returns:
+            freqs_cis: tensor of shape (sum(t * height * width), dim//2)
+        """
+        if not hasattr(self, "freqs_cis"):
+            self.register_buffer(
+                "freqs_cis", self._precompute_freqs_cis(device), persistent=False
+            )
+
+        shapes = grid_thws.tolist()
+        assert all(
+            1 <= h <= self.max_height and 1 <= w <= self.max_width for t, h, w in shapes
+        ), (
+            shapes,
+            self.max_height,
+            self.max_width,
+        )
+        freqs_cis = torch.cat(
+            [
+                self.freqs_cis[:h, :w].reshape(-1, self.dim // 2).repeat(t, 1)
+                for t, h, w in shapes
+            ],
+            dim=0,
+        )
+        return freqs_cis
+
+
+class MLP2(nn.Module):
+    """
+    Args:
+        dims: [in_dim, hidden_dim, out_dim]
+        bias: whether to use bias in linear layer.
+    """
+
+    def __init__(self, dims: list[int], activation, bias=True):
+        super().__init__()
+        assert len(dims) == 3
+        self.fc0 = nn.Linear(dims[0], dims[1], bias=bias)
+        self.fc1 = nn.Linear(dims[1], dims[2], bias=bias)
+        self.activation = activation
+        for m in [self.fc0, self.fc1]:
+            nn.init.trunc_normal_(m.weight, std=math.sqrt(2 / m.in_features))
+            if m.bias is not None:
+                nn.init.zeros_(m.bias)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.fc0(x)
+        x = self.activation(x)
+        return self.fc1(x)
+
+
+class MoonViTEncoderLayer(nn.Module):
+    def __init__(
+        self,
+        num_heads: int,
+        hidden_dim: int,
+        mlp_dim: int,
+        qkv_hidden_size: int | None = None,
+        norm_type: str = "layernorm",
+        mlp_type: str = "mlp2",
+        *,
+        attn_implementation: str = "flash_attention_2",
+        activation=F.gelu,
+        attn_bias: bool = False,
+        linear_bias: bool = True,
+        use_deterministic_attn: bool = False,
+    ):
+        super().__init__()
+        self.num_heads = num_heads
+        self.hidden_dim = hidden_dim
+        self.qkv_hidden_size = (
+            hidden_dim if qkv_hidden_size is None else qkv_hidden_size
+        )
+        self.hidden_size_per_attention_head = self.qkv_hidden_size // self.num_heads
+        self.attn_implementation = attn_implementation
+        self.use_deterministic_attn = use_deterministic_attn
+
+        if norm_type == "layernorm":
+            self.norm0 = nn.LayerNorm(hidden_dim)
+            self.norm1 = nn.LayerNorm(hidden_dim)
+        elif norm_type == "rmsnorm":
+            self.norm0 = nn.RMSNorm(hidden_dim)
+            self.norm1 = nn.RMSNorm(hidden_dim)
+        else:
+            raise NotImplementedError(f"Not support norm_type: {norm_type}")
+
+        if mlp_type == "mlp2":
+            self.mlp = MLP2(
+                [hidden_dim, mlp_dim, hidden_dim], activation, bias=linear_bias
+            )
+        else:
+            raise NotImplementedError(f"Not support mlp_type: {mlp_type}")
+
+        self.wqkv = nn.Linear(hidden_dim, self.qkv_hidden_size * 3, bias=attn_bias)
+        self.wo = nn.Linear(self.qkv_hidden_size, hidden_dim, bias=attn_bias)
+
+    def attention_qkvpacked(
+        self,
+        x: torch.Tensor,
+        cu_seqlens: torch.Tensor,
+        max_seqlen: torch.Tensor,
+        rope_freqs_cis: torch.Tensor | None = None,
+    ):
+        """
+        Args:
+            x (torch.Tensor): (batch_size, seqlen, hidden_dim)
+            cu_seqlens (torch.Tensor):
+        """
+        xqkv = self.wqkv(x)
+
+        qkv_shape = xqkv.size()[:-1] + (
+            3,
+            self.num_heads,
+            self.hidden_size_per_attention_head,
+        )
+        # xqkv: (batch_size, seqlen, 3, nheads, headdim)
+        xqkv = xqkv.view(*qkv_shape)
+        xq, xk, xv = torch.unbind(xqkv, dim=-3)
+
+        xq, xk = apply_rope(xq, xk, rope_freqs_cis)
+
+        attn_func = VL_VISION_ATTENTION_FUNCTIONS[self.attn_implementation]
+        attn_out = attn_func(
+            xq,
+            xk,
+            xv,
+            q_cu_seqlens=cu_seqlens,
+            k_cu_seqlens=cu_seqlens,
+            max_seqlen_k=max_seqlen,
+            max_seqlen_q=max_seqlen,
+            deterministic=self.use_deterministic_attn,
+        )
+
+        attn_out = self.wo(attn_out)
+        return attn_out
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cu_seqlens: torch.Tensor,
+        max_seqlen: int,
+        rope_freqs_cis: torch.Tensor | None = None,
+    ):
+        residual = hidden_states
+        hidden_states = self.norm0(hidden_states)
+
+        hidden_states = self.attention_qkvpacked(
+            hidden_states, cu_seqlens, max_seqlen, rope_freqs_cis
+        )
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.norm1(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        return hidden_states
+
+
+class MoonViT3dEncoder(nn.Module):
+    def __init__(
+        self,
+        hidden_dim: int,
+        num_layers: int,
+        block_cfg: dict,
+        use_deterministic_attn: bool = False,
+    ) -> None:
+        super().__init__()
+        self.use_deterministic_attn = use_deterministic_attn
+
+        qkv_hidden_size = (
+            block_cfg["hidden_dim"]
+            if block_cfg.get("qkv_hidden_size") is None
+            else block_cfg["qkv_hidden_size"]
+        )
+        self.rope_2d = Rope2DPosEmbRepeated(
+            qkv_hidden_size // block_cfg["num_heads"], 512, 512
+        )
+        self.blocks = nn.ModuleList(
+            [
+                MoonViTEncoderLayer(
+                    **block_cfg, use_deterministic_attn=self.use_deterministic_attn
+                )
+                for _ in range(num_layers)
+            ]
+        )
+        norm_type = block_cfg.get("norm_type", "layernorm")
+        if norm_type == "layernorm":
+            self.final_layernorm = nn.LayerNorm(hidden_dim)
+        elif norm_type == "rmsnorm":
+            self.final_layernorm = nn.RMSNorm(hidden_dim)
+        else:
+            raise NotImplementedError(f"Not support norm_type: {norm_type}")
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        grid_thws: torch.Tensor,
+    ) -> torch.Tensor:
+        rope_freqs_cis = self.rope_2d.get_freqs_cis(
+            grid_thws=grid_thws, device=hidden_states.device
+        )
+
+        lengths = torch.cat(
+            (
+                torch.zeros(1, dtype=grid_thws.dtype, device=grid_thws.device),
+                grid_thws[:, 0] * grid_thws[:, 1] * grid_thws[:, 2],
+            )
+        )
+
+        max_seqlen = lengths.max()
+        cu_seqlens = lengths.to(hidden_states.device).cumsum(dim=0, dtype=torch.int32)
+        for block in self.blocks:
+            hidden_states = block(
+                hidden_states, cu_seqlens, max_seqlen, rope_freqs_cis=rope_freqs_cis
+            )
+
+        hidden_states = self.final_layernorm(hidden_states)
+        return hidden_states
+
+
+def tpool_patch_merger(
+    x: torch.Tensor,
+    grid_thws: torch.Tensor,
+    merge_kernel_size: tuple[int, int] = (2, 2),
+) -> list[torch.Tensor]:
+    d_model = x.size(-1)
+
+    outputs = []
+    pre_sum = 0
+    for t, h, w in grid_thws.tolist():
+        # Get the current sequence
+        seq = x[pre_sum : pre_sum + t * h * w]
+        # Reshape along self.merge_kernel_size and concat to the last dimension
+        kernel_height, kernel_width = merge_kernel_size
+        new_height, new_width = h // kernel_height, w // kernel_width
+        reshaped_seq = seq.view(
+            t, new_height, kernel_height, new_width, kernel_width, d_model
+        )
+        reshaped_seq = (
+            reshaped_seq.permute(0, 1, 3, 2, 4, 5).contiguous().mean(dim=0)
+        )  # temporal pooling
+        padded_seq = reshaped_seq.view(
+            new_height * new_width, kernel_height * kernel_width, -1
+        )
+        outputs.append(padded_seq)
+        pre_sum += t * h * w
+
+    return outputs
+
+
+class MoonViT3dPretrainedModel(PreTrainedModel):
+    config_class = None
+    model_type = "moonvit3d"
+    _no_split_modules = ["MoonViTEncoderLayer"]
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+
+    def __init__(self, config, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+        config = deepcopy(config)
+        self.merge_kernel_size = config.merge_kernel_size
+        self.patch_size = config.patch_size
+        self.merge_type = config.merge_type
+
+        self.patch_embed = MoonVision3dPatchEmbed(
+            out_dim=config.hidden_size,
+            patch_size=config.patch_size,
+            pos_emb_height=config.init_pos_emb_height,
+            pos_emb_width=config.init_pos_emb_width,
+            pos_emb_time=config.init_pos_emb_time,
+            pos_emb_type=config.pos_emb_type,
+            patch_embed_proj_bias=getattr(config, "patch_embed_proj_bias", True),
+            pos_emb_interpolation_mode=getattr(
+                config, "pos_emb_interpolation_mode", "bicubic"
+            ),
+        )
+
+        self.encoder = MoonViT3dEncoder(
+            hidden_dim=config.hidden_size,
+            num_layers=config.num_hidden_layers,
+            block_cfg={
+                "num_heads": config.num_attention_heads,
+                "hidden_dim": config.hidden_size,
+                "qkv_hidden_size": getattr(config, "qkv_hidden_size", None),
+                "mlp_dim": config.intermediate_size,
+                "norm_type": getattr(config, "norm_type", "layernorm"),
+                "mlp_type": getattr(config, "mlp_type", "mlp2"),
+                "activation": PytorchGELUTanh(),
+                "attn_bias": getattr(config, "attn_bias", True),
+                "linear_bias": getattr(config, "linear_bias", True),
+                "attn_implementation": config._attn_implementation,
+            },
+            use_deterministic_attn=getattr(self, "use_deterministic_attn", False),
+        )
+
+    def forward(
+        self, pixel_values: torch.Tensor, grid_thws: torch.Tensor
+    ) -> torch.Tensor:
+        """
+        Args:
+            pixel_values (torch.Tensor): The input pixel values.
+            grid_thws (torch.Tensor): Temporal, height and width.
+
+        Returns:
+            torch.Tensor: The output tokens.
+        """
+        # grid_thws = grid_thws.to('cpu')
+        assert grid_thws.ndim == 2, f"grid_thws should be 2D, got {grid_thws.ndim}"
+        assert grid_thws.size(1) == 3, f"No support for thw: {grid_thws}"
+        hidden_states = self.patch_embed(pixel_values, grid_thws)
+        hidden_states = self.encoder(hidden_states, grid_thws)
+        if (
+            self.merge_type == "sd2_tpool"
+        ):  # spatial downsampling 2x with temporal pooling all
+            hidden_states = tpool_patch_merger(
+                hidden_states, grid_thws, merge_kernel_size=self.merge_kernel_size
+            )
+        else:
+            raise NotImplementedError(f"Not support {self.merge_type}")
+
+        return hidden_states
+
+
+# ============================================================================
+# MM Projector Helper Classes (from mm_projector/modeling_mm_projectors.py)
+# ============================================================================
+
+
+class IdentityMap(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x, *args, **kwargs):
+        return x
+
+
+class MLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        # TODO, use faster LayerNorm
+        self.pre_norm = nn.LayerNorm(config.mm_hidden_size)
+        self.proj = nn.Sequential(
+            nn.Linear(config.mm_hidden_size, config.hidden_size),
+            nn.GELU(),
+            nn.Linear(config.hidden_size, config.hidden_size),
+        )
+
+    def forward(self, x, *args, **kwargs):
+        assert isinstance(x, list | tuple), f"x is not a list or tuple: {type(x)}"
+        lengths = [item.shape[0] for item in x]
+        x = torch.cat(x, dim=0)
+        x = self.pre_norm(x)
+        x = self.proj(x)
+        x = torch.split(x, lengths, dim=0)
+
+        return x
+
+
+class PatchMergerMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        eps = config.projector_ln_eps
+        self.hidden_size = config.mm_hidden_size * (
+            config.merge_kernel_size[0] * config.merge_kernel_size[1]
+        )
+        self.pre_norm = nn.LayerNorm(config.mm_hidden_size, eps=eps)
+        self.proj = nn.Sequential(
+            nn.Linear(self.hidden_size, self.hidden_size),
+            nn.GELU(),
+            nn.Linear(self.hidden_size, config.hidden_size),
+        )
+
+    def forward(self, x, *args, **kwargs):
+        if isinstance(x, list) or isinstance(x, tuple):
+            x = [self.proj(self.pre_norm(item).view(item.shape[0], -1)) for item in x]
+        else:
+            # B, N, N_k, C = x.shape
+            B = x.shape[0]
+            x = self.proj(self.pre_norm(x).view(B, -1, self.hidden_size))
+        return x
+
+
+class PatchMergerMLPV2(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        eps = config.projector_ln_eps
+        self.hidden_size = config.mm_hidden_size * (
+            config.merge_kernel_size[0] * config.merge_kernel_size[1]
+        )
+        self.proj = nn.Sequential(
+            nn.Linear(self.hidden_size, self.hidden_size, bias=False),
+            nn.GELU(),
+            nn.Linear(self.hidden_size, config.hidden_size, bias=False),
+        )
+        self.post_norm = nn.RMSNorm(config.hidden_size, eps=eps)
+        for m in self.proj.modules():
+            if isinstance(m, nn.Linear):
+                nn.init.trunc_normal_(m.weight, std=math.sqrt(2 / m.in_features))
+                if m.bias is not None:
+                    nn.init.zeros_(m.bias)
+
+    def forward(self, x, *args, **kwargs):
+        if isinstance(x, list) or isinstance(x, tuple):
+            lengths = [item.shape[0] for item in x]
+            x = torch.concat([item.view(item.shape[0], -1) for item in x], dim=0)
+            x = self.post_norm(self.proj(x))
+            x = torch.split(x, lengths, dim=0)
+        else:
+            # B, N, N_k, C = x.shape
+            B = x.shape[0]
+            x = self.proj(x.view(B, -1, self.hidden_size))
+            x = self.post_norm(x)
+        return x
+
+
+class KimiK3PreTrainedModel(PreTrainedModel):
+    config_class = KimiK3Config
+    base_model_prefix = "model"
+    _no_split_modules = [
+        "MoonViT3dPretrainedModel",
+        "MoonViTEncoderLayer",
+        "KimiDecoderLayer",
+        "PatchMergerMLP",
+        "PatchMergerMLPV2",
+    ]
+    _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
+    _supports_sdpa = False
+
+    def _init_weights(self, module):
+        # HOTFIX: disk offloading attempts to initialize the meta tensors
+        # but this is bad programming: we shouldn't be initializing these
+        # params in the first place
+        # the init attempt attempts to get `module.weight`, which DNE for qmodels
+
+        return
+
+        # important: this ported version of Llava isn't meant for training from scratch - only
+        # inference and fine-tuning - so the proper init weights code has been removed - the original codebase
+        # https://github.com/haotian-liu/LLaVA/tree/main/llava should serve for that purpose
+        std = (
+            self.config.initializer_range
+            if hasattr(self.config, "initializer_range")
+            else self.config.text_config.initializer_range
+        )
+
+        if hasattr(module, "class_embedding"):
+            module.class_embedding.data.normal_(mean=0.0, std=std)
+
+        if isinstance(module, (nn.Linear, nn.Conv2d)):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+
+class VisionTowerConfig(PretrainedConfig):
+    model_type = "moonvit3d"
+
+    def __init__(self, config: KimiK3Config, **kwargs):
+        super().__init__(**kwargs)
+        self.patch_size = config.patch_size
+        self.init_pos_emb_height = config.init_pos_emb_height
+        self.init_pos_emb_width = config.init_pos_emb_width
+        self.init_pos_emb_time = config.init_pos_emb_time
+        self.pos_emb_type = config.pos_emb_type
+        self.num_attention_heads = config.vt_num_attention_heads
+        self.num_hidden_layers = config.vt_num_hidden_layers
+        self.hidden_size = config.vt_hidden_size
+        self.intermediate_size = config.vt_intermediate_size
+        self.merge_kernel_size = config.merge_kernel_size
+        self.merge_type = config.merge_type
+        self._attn_implementation = config._attn_implementation
+        self.qkv_hidden_size = getattr(config, "qkv_hidden_size", None)
+        self.norm_type = getattr(config, "norm_type", "layernorm")
+        self.attn_bias = getattr(config, "attn_bias", True)
+        self.patch_embed_proj_bias = getattr(config, "patch_embed_proj_bias", True)
+        self.mlp_type = getattr(config, "mlp_type", "mlp2")
+        self.linear_bias = getattr(config, "linear_bias", True)
+        self.pos_emb_interpolation_mode = getattr(
+            config, "pos_emb_interpolation_mode", "bilinear"
+        )
+
+
+class ProjectorConfig:
+    def __init__(self, config: KimiK3Config):
+        self.mm_projector_type = config.mm_projector_type
+        self.mm_hidden_size = config.mm_hidden_size
+        self.hidden_size = config.text_hidden_size
+        self.merge_kernel_size = config.merge_kernel_size
+        self.projector_hidden_act = config.projector_hidden_act
+        self.projector_ln_eps = config.projector_ln_eps
+
+
+# ref https://github.com/huggingface/transformers/blob/78b2929c0554b79e0489b451ce4ece14d265ead2/src/transformers/models/llava/modeling_llava.py#L240
+class KimiK3ForConditionalGeneration(KimiK3PreTrainedModel):
+    @classmethod
+    def _supports_default_dynamic_cache(cls) -> bool:
+        return False
+
+    def __init__(self, config: KimiK3Config):
+        super().__init__(config)
+
+        vt_config = VisionTowerConfig(config.vision_config)
+        self.vision_tower = MoonViT3dPretrainedModel(vt_config)
+
+        proj_config = ProjectorConfig(config.vision_config)
+        if proj_config.mm_projector_type == "identity":
+            self.mm_projector = IdentityMap()
+        elif proj_config.mm_projector_type == "mlp":
+            self.mm_projector = MLP(proj_config)
+        elif proj_config.mm_projector_type == "patchmerger":
+            self.mm_projector = PatchMergerMLP(proj_config)
+        elif proj_config.mm_projector_type == "patchmergerv2":
+            self.mm_projector = PatchMergerMLPV2(proj_config)
+        else:
+            raise ValueError(
+                f"Unsupported mm_projector_type: {proj_config.mm_projector_type}"
+            )
+
+        self.language_model = KimiLinearForCausalLM(config.text_config)
+        self.post_init()
+
+        if hasattr(self.language_model, "dtype"):
+            target_dtype = self.language_model.dtype
+            self.vision_tower = self.vision_tower.to(dtype=target_dtype)
+            self.mm_projector = self.mm_projector.to(dtype=target_dtype)
+
+    def get_input_embeddings(self):
+        return self.language_model.get_input_embeddings()
+
+    def set_input_embeddings(self, value):
+        self.language_model.set_input_embeddings(value)
+
+    def get_output_embeddings(self):
+        return self.language_model.get_output_embeddings()
+
+    def set_output_embeddings(self, new_embeddings):
+        self.language_model.set_output_embeddings(new_embeddings)
+
+    def set_decoder(self, decoder):
+        self.language_model.set_decoder(decoder)
+
+    def get_decoder(self):
+        return self.language_model.get_decoder()
+
+    def tie_weights(self, **kwargs):
+        return self.language_model.tie_weights(**kwargs)
+
+    def resize_token_embeddings(
+        self, new_num_tokens: int | None = None, pad_to_multiple_of=None
+    ) -> nn.Embedding:
+        model_embeds = self.language_model.resize_token_embeddings(
+            new_num_tokens, pad_to_multiple_of
+        )
+        # update vocab size
+        self.config.text_config.vocab_size = model_embeds.num_embeddings
+        self.vocab_size = model_embeds.num_embeddings
+        return model_embeds
+
+    def _merge_input_ids_with_image_features(
+        self,
+        image_features: list[torch.Tensor],
+        inputs_embeds: torch.Tensor,
+        input_ids: torch.Tensor,
+        attention_mask: torch.Tensor,
+        labels: torch.Tensor | None = None,
+    ):
+        """
+        Args:
+            image_features (:obj:`torch.Tensor` of shape :obj:`(num_image_tokens, embed_dim)`):
+                The image features to merge with the input embeddings.
+            inputs_embeds (:obj:`torch.Tensor` of shape :obj:`(batch_size, sequence_length, embed_dim)`):
+                The input embeddings.
+            input_ids (:obj:`torch.Tensor` of shape :obj:`(batch_size, sequence_length)`):
+                The input ids.
+            attention_mask (:obj:`torch.Tensor` of shape :obj:`(batch_size, sequence_length)`):
+                The attention mask.
+            labels (:obj:`torch.Tensor` of shape :obj:`(batch_size, sequence_length)`, *optional*):
+                The labels.
+        """
+        _, embed_dim = image_features[0].shape
+        feature_lengths = [x.shape[0] for x in image_features]
+        image_features = torch.cat(image_features, dim=0)
+
+        image_token_index: int = self.config.media_placeholder_token_id
+        pad_token_id: int = self.config.pad_token_id
+        ignore_index: int = self.config.ignore_index
+
+        batch_size, sequence_length = input_ids.shape
+        left_padding = not torch.sum(input_ids[:, -1] == torch.tensor(pad_token_id))
+
+        # 1. Create a mask to know where special image tokens are
+        _token_occupation_table = torch.ones_like(input_ids.flatten())
+        _token_occupation_table[input_ids.flatten() == image_token_index] = (
+            torch.tensor(feature_lengths, dtype=torch.long, device=input_ids.device)
+        )
+        _token_occupation_table = _token_occupation_table.reshape(input_ids.shape)
+
+        max_embed_dim = _token_occupation_table.sum(-1).max().item()
+        assert (
+            max_embed_dim >= sequence_length
+        ), f"The maximum embedding dimension ({max_embed_dim}) is less than the sequence length ({sequence_length})"
+        batch_indices, non_image_indices = torch.where(input_ids != image_token_index)
+
+        # 2. Compute the positions where text should be written
+        # Calculate new positions for text tokens in merged image-text sequence.
+        new_token_positions = torch.cumsum(_token_occupation_table, -1) - 1
+        nb_image_pad = max_embed_dim - 1 - new_token_positions[:, -1]
+        if left_padding:
+            new_token_positions += nb_image_pad[:, None]  # offset for left padding
+        text_to_overwrite = new_token_positions[batch_indices, non_image_indices]
+
+        # 3. Create the full embedding, already padded to the maximum position
+        final_embedding = torch.zeros(
+            batch_size,
+            max_embed_dim,
+            embed_dim,
+            dtype=inputs_embeds.dtype,
+            device=inputs_embeds.device,
+        )
+        final_attention_mask = torch.zeros(
+            batch_size,
+            max_embed_dim,
+            dtype=attention_mask.dtype,
+            device=inputs_embeds.device,
+        )
+        if labels is not None:
+            final_labels = torch.full(
+                (batch_size, max_embed_dim),
+                ignore_index,
+                dtype=input_ids.dtype,
+                device=input_ids.device,
+            )
+        # In case the Vision model or the Language model has been offloaded to CPU, we need to manually
+        # set the corresponding tensors into their correct target device.
+        target_device = inputs_embeds.device
+        batch_indices, non_image_indices, text_to_overwrite = (
+            batch_indices.to(target_device),
+            non_image_indices.to(target_device),
+            text_to_overwrite.to(target_device),
+        )
+        attention_mask = attention_mask.to(target_device)
+
+        # 4. Fill the embeddings based on the mask.
+        final_embedding[batch_indices, text_to_overwrite] = inputs_embeds[
+            batch_indices, non_image_indices
+        ]
+        final_attention_mask[batch_indices, text_to_overwrite] = attention_mask[
+            batch_indices, non_image_indices
+        ]
+        if labels is not None:
+            final_labels[batch_indices, text_to_overwrite] = labels[
+                batch_indices, non_image_indices
+            ]
+
+        # 5. Fill the embeddings corresponding to the images. Anything that is not `text_positions` needs filling (#29835)
+        image_to_overwrite = torch.full(
+            (batch_size, max_embed_dim),
+            True,
+            dtype=torch.bool,
+            device=inputs_embeds.device,
+        )
+        image_to_overwrite[batch_indices, text_to_overwrite] = False
+        image_to_overwrite &= image_to_overwrite.cumsum(-1) - 1 >= nb_image_pad[
+            :, None
+        ].to(target_device)
+
+        if image_to_overwrite.sum() != image_features.shape[:-1].numel():
+            raise ValueError(
+                f"The input provided to the model are wrong. The number of image tokens is {image_to_overwrite.sum()} while"
+                f" the number of image features given to the model is {image_features.shape[:-1].numel()}. "
+                "This prevents correct indexing and breaks batch generation."
+            )
+
+        final_embedding[image_to_overwrite] = (
+            image_features.contiguous().reshape(-1, embed_dim).to(target_device)
+        )
+        final_attention_mask |= image_to_overwrite
+        position_ids = (final_attention_mask.cumsum(-1) - 1).masked_fill_(
+            (final_attention_mask == 0), 1
+        )
+
+        # 6. Mask out the embedding at padding positions, as we later use the past_key_value value to determine the non-attended tokens.
+        batch_indices, pad_indices = torch.where(input_ids == pad_token_id)
+        indices_to_mask = new_token_positions[batch_indices, pad_indices]
+
+        final_embedding[batch_indices, indices_to_mask] = 0
+
+        if labels is None:
+            final_labels = None
+
+        return final_embedding, final_attention_mask, final_labels, position_ids
+
+    def _extract_image_features(
+        self, pixel_values: torch.Tensor, grid_thws: torch.Tensor
+    ) -> list[torch.Tensor]:
+        """
+        Args:
+            pixel_values (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, num_channels, height, width)`):
+                The pixel values of the images processed by image processor.
+            grid_thws (:obj:`torch.Tensor` of shape :obj:`(batch_size, 3)`):
+                The grid, height, width of the images.
+
+        Returns:
+            selected_image_feature (:obj:`torch.FloatTensor` of shape :obj:`(num_image_tokens, embed_dim)`):
+                The selected image features to use as input to the projector head.
+
+        """
+
+        target_dtype = self.vision_tower.patch_embed.proj.weight.dtype
+        pixel_values = pixel_values.to(target_dtype)
+
+        image_features = self.vision_tower(pixel_values, grid_thws)
+        return image_features
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor | None = None,
+        pixel_values: torch.FloatTensor | list[torch.FloatTensor] | None = None,
+        grid_thws: torch.Tensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: list[torch.FloatTensor] | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        labels: torch.LongTensor | None = None,
+        use_cache: bool | None = None,
+        output_attentions: bool | None = None,
+        output_hidden_states: bool | None = None,
+        return_dict: bool | None = None,
+    ) -> tuple | LlavaCausalLMOutputWithPast:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        ```"""
+        assert self.vision_tower is not None, "vision_tower is not loaded"
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        if inputs_embeds is None:
+            # 1. Extra the input embeddings
+            inputs_embeds = self.get_input_embeddings()(input_ids)
+
+            # 2. Merge text and images
+            if (
+                pixel_values is not None
+                and len(pixel_values) > 0
+                and input_ids.shape[1] != 1
+            ):
+                image_features = self._extract_image_features(pixel_values, grid_thws)
+                if self.mm_projector:
+                    image_features = self.mm_projector(image_features)
+
+                inputs_embeds = inputs_embeds.to(
+                    image_features[0].dtype
+                )  # num_tokens, embed_dim
+                inputs_embeds, attention_mask, labels, position_ids = (
+                    self._merge_input_ids_with_image_features(
+                        image_features,
+                        inputs_embeds,
+                        input_ids,
+                        attention_mask,
+                        labels,
+                    )
+                )
+
+            # In case input_ids.shape[1] == 1 & pixel_values==None & past_key_values != None, we are in the case of
+            # generation with cache
+            elif (
+                past_key_values is not None
+                and pixel_values is not None
+                and input_ids.shape[1] == 1
+            ):
+                # Retrieve the first layer to inspect the logits and mask out the hidden states
+                # that are set to 0
+                first_layer_past_key_value = past_key_values[0][0][:, :, :, 0]
+
+                # Sum all dimensions of head_dim (-2) to avoid random errors such as: https://github.com/huggingface/transformers/pull/28032#issuecomment-1863691941
+                batch_index, non_attended_tokens = torch.where(
+                    first_layer_past_key_value.float().sum(-2) == 0
+                )
+
+                # Get the target length
+                target_length = input_ids.shape[1]
+                past_length = first_layer_past_key_value.shape[-1]
+
+                extended_attention_mask = torch.ones(
+                    (attention_mask.shape[0], past_length),
+                    dtype=attention_mask.dtype,
+                    device=attention_mask.device,
+                )
+
+                # Filter out only the tokens that can be un-attended, this can happen
+                # if one uses Llava + Fused modules where the cache on the
+                # first iteration is already big enough, or if one passes custom cache
+                valid_indices = non_attended_tokens < extended_attention_mask.size(-1)
+                new_batch_index = batch_index[valid_indices]
+                new_non_attended_tokens = non_attended_tokens[valid_indices]
+
+                # Zero-out the places where we don't need to attend
+                extended_attention_mask[new_batch_index, new_non_attended_tokens] = 0
+
+                attention_mask = torch.cat(
+                    (extended_attention_mask, attention_mask[:, -target_length:]), dim=1
+                )
+                position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1
+
+        outputs = self.language_model(
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        logits = outputs[0]
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            if attention_mask is not None:
+                shift_attention_mask = attention_mask[..., 1:]
+                shift_logits = logits[..., :-1, :][
+                    shift_attention_mask.to(logits.device) != 0
+                ].contiguous()
+                shift_labels = labels[..., 1:][
+                    shift_attention_mask.to(labels.device) != 0
+                ].contiguous()
+            else:
+                shift_logits = logits[..., :-1, :].contiguous()
+                shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = nn.CrossEntropyLoss()
+            loss = loss_fct(
+                shift_logits.view(-1, shift_logits.size(-1)),
+                shift_labels.view(-1).to(shift_logits.device),
+            )
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return LlavaCausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        inputs_embeds=None,
+        pixel_values=None,
+        grid_thws=None,
+        attention_mask=None,
+        **kwargs,
+    ):
+        if past_key_values is not None:
+            if hasattr(past_key_values, "get_seq_length"):
+                cache_length = past_key_values.get_seq_length()
+                past_length = getattr(past_key_values, "seen_tokens", cache_length)
+            else:
+                cache_length = past_length = past_key_values[0][0].shape[2]
+
+            # Keep only the unprocessed tokens:
+            # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
+            # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
+            # input)
+            if (
+                attention_mask is not None
+                and attention_mask.shape[1] > input_ids.shape[1]
+            ):
+                input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
+            # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
+            # input_ids based on the past_length.
+            elif past_length < input_ids.shape[1]:
+                input_ids = input_ids[:, past_length:]
+            # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
+            elif self.config.media_placeholder_token_id in input_ids:
+                input_ids = input_ids[:, input_ids.shape[1] - 1 :]
+            # If the cache has seen more tokens than it can hold, then the cache has a size limit. Let's discard the
+            # older attention values, as their corresponding values are not part of the input.
+            if cache_length < past_length and attention_mask is not None:
+                attention_mask = attention_mask[
+                    :, -(cache_length + input_ids.shape[1]) :
+                ]
+
+        position_ids = kwargs.get("position_ids", None)
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -input_ids.shape[1] :]
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+                "pixel_values": pixel_values,
+                "grid_thws": grid_thws,
+            }
+        )
+        return model_inputs
+
+    def _reorder_cache(self, *args, **kwargs):
+        return self.language_model._reorder_cache(*args, **kwargs)

modeling_kimi_k3_linear.py

@@ -0,0 +1,1459 @@
+# coding=utf-8
+# Copyright 2025-2026 The Moonshot AI Team, DeepSeek-AI, and HuggingFace Inc. team. All rights reserved.
+#
+# The multi-head latent attention, MoE gating and sparse MoE block in this file are
+# adapted from DeepSeek-V3 (DeepSeek-V3/modeling_deepseek.py). They have been
+# extensively modified and extended for the Kimi-Linear architecture.
+#
+# Licensing Information:
+# - Code adapted from DeepSeek-V3 (DeepSeek-V3/modeling_deepseek.py) is licensed under the Apache License, Version 2.0.
+# - Other parts of the code are licensed under the Kimi K3 License (see the LICENSE file in this repository).
+#
+# Apache License, Version 2.0:
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import math
+from collections.abc import Callable
+from typing import Any
+
+import torch
+import torch.nn.functional as F
+import transformers
+from einops import rearrange
+from packaging import version
+from torch import nn
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache
+from transformers.generation import GenerationMixin
+from transformers.masking_utils import create_causal_mask
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+)
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.processing_utils import Unpack
+from transformers.pytorch_utils import ALL_LAYERNORM_LAYERS
+from transformers.utils import (
+    TransformersKwargs,
+    can_return_tuple,
+    logging,
+)
+from transformers.utils.generic import check_model_inputs
+from transformers.utils.output_capturing import OutputRecorder
+
+try:
+    from fla.modules import FusedRMSNormGated, ShortConvolution
+    from fla.ops.kda import chunk_kda, fused_recurrent_kda
+
+    # from fla.ops.kda.gate import fused_kda_gate  # deprecated, gate is now computed inside chunk_kda/fused_recurrent_kda
+    from fla.ops.utils.index import prepare_cu_seqlens_from_mask, prepare_lens_from_mask
+    from fla.utils import tensor_cache
+except ImportError:
+    raise ImportError("Plese run `pip install -U fla-core`")
+
+def get_calibrate_all_experts_flag() -> bool:
+    return False
+
+from .configuration_kimi_k3 import KimiLinearConfig
+
+assert version.parse(transformers.__version__) >= version.parse(
+    "4.56.0"
+), "Please upgrade transformers to >= 4.56.0"
+
+logger = logging.get_logger(__name__)
+
+
+# Register Moonshot-specific activation functions
+class SituAndMul(nn.Module):
+    """
+    SituAndMul activation: beta * tanh(gate / beta) * sigmoid(gate) * up
+    When linear_beta is set, up is also transformed by linear_beta * tanh(up / linear_beta).
+    """
+
+    def __init__(self, beta: float = 1.0, linear_beta: float | None = None):
+        super().__init__()
+        self.beta = beta
+        self.linear_beta = linear_beta
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        d = x.shape[-1] // 2
+        gate = x[..., :d].to(torch.float32)
+        up = x[..., d:].to(torch.float32)
+        situ_a = self.beta * torch.tanh(gate / self.beta) * torch.sigmoid(gate)
+        if self.linear_beta is not None:
+            up = self.linear_beta * torch.tanh(up / self.linear_beta)
+        return (situ_a * up).to(x.dtype)
+
+
+ACT2FN["situ"] = SituAndMul
+
+
+def _get_situ_activation_params(config: KimiLinearConfig):
+    beta = getattr(config, "activation_situ_beta", None)
+    linear_beta = getattr(config, "activation_situ_linear_beta", None)
+    return beta or 1.0, linear_beta
+
+
+def index_first_axis(x, indices):
+    return x[indices]
+
+
+@tensor_cache
+def get_unpad_data(
+    attention_mask: torch.Tensor,
+) -> tuple[torch.Tensor, torch.Tensor, int]:
+    lens = prepare_lens_from_mask(attention_mask)
+    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = lens.max().item()
+    cu_seqlens = prepare_cu_seqlens_from_mask(attention_mask)
+    return indices, cu_seqlens, max_seqlen_in_batch
+
+
+def pad_input(
+    hidden_states: torch.Tensor,
+    indices: torch.LongTensor,
+    batch_size: int,
+    seq_len: int,
+) -> torch.Tensor:
+    out = hidden_states.new_zeros((batch_size * seq_len, *hidden_states.shape[1:]))
+    out[indices] = hidden_states
+    return out.view(batch_size, seq_len, *hidden_states.shape[1:])
+
+
+class KimiDynamicCache:
+    """
+    Dynamic cache for Kimi model.
+    Inspired by Qwen3-Next
+    """
+
+    is_compileable = False
+
+    def __init__(self, config: KimiLinearConfig):
+        super().__init__()
+        self.config = config
+
+        if config.linear_attn_config is not None:
+            self.layer_types = []
+            for i in range(config.num_hidden_layers):
+                if config.is_kda_layer(i):
+                    self.layer_types.append("linear_attention")
+                else:
+                    self.layer_types.append("full_attention")
+        else:
+            self.layer_types = ["full_attention"] * config.num_hidden_layers
+
+        self.transformer_layers = [
+            i
+            for i in range(config.num_hidden_layers)
+            if self.layer_types[i] == "full_attention"
+        ]
+
+        linear_layers = [
+            i
+            for i in range(config.num_hidden_layers)
+            if self.layer_types[i] == "linear_attention"
+        ]
+        self.last_linear_layer = linear_layers[-1] if linear_layers else -1
+
+        self.conv_states = [None for _ in range(config.num_hidden_layers)]
+        self.recurrent_states = [None for _ in range(config.num_hidden_layers)]
+        self.key_cache = [None for _ in range(config.num_hidden_layers)]
+        self.value_cache = [None for _ in range(config.num_hidden_layers)]
+
+    def __len__(self):
+        return len(self.layer_types)
+
+    def update(
+        self,
+        key_states: torch.Tensor,
+        value_states: torch.Tensor,
+        layer_idx: int,
+        cache_kwargs: dict[str, Any] | None = None,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        if self.key_cache[layer_idx] is None:
+            self.key_cache[layer_idx] = key_states
+            self.value_cache[layer_idx] = value_states
+        else:
+            self.key_cache[layer_idx] = torch.cat(
+                [self.key_cache[layer_idx], key_states], dim=2
+            )
+            self.value_cache[layer_idx] = torch.cat(
+                [self.value_cache[layer_idx], value_states], dim=2
+            )
+
+        return self.key_cache[layer_idx], self.value_cache[layer_idx]
+
+    def reorder_cache(self, beam_idx: torch.LongTensor):
+        """Reorders the cache for beam search, given the selected beam indices."""
+        for layer_idx in range(len(self.key_cache)):
+            if self.key_cache[layer_idx] is not None:
+                device = self.key_cache[layer_idx].device
+                beam_idx = beam_idx.to(device)
+                self.key_cache[layer_idx] = self.key_cache[layer_idx].index_select(
+                    0, beam_idx
+                )
+                self.value_cache[layer_idx] = self.value_cache[layer_idx].index_select(
+                    0, beam_idx
+                )
+
+            if self.conv_states[layer_idx] is not None:
+                device = self.conv_states[layer_idx][0].device
+                beam_idx = beam_idx.to(device)
+                q_conv, k_conv, v_conv = self.conv_states[layer_idx]
+                self.conv_states[layer_idx] = (
+                    q_conv.index_select(0, beam_idx),
+                    k_conv.index_select(0, beam_idx),
+                    v_conv.index_select(0, beam_idx),
+                )
+                self.recurrent_states[layer_idx] = self.recurrent_states[
+                    layer_idx
+                ].index_select(0, beam_idx)
+
+    def get_seq_length(self, layer_idx: int | None = 0) -> int:
+        """Returns the sequence length of the cached states. A layer index can be optionally passed."""
+        # take any layer that contains cache and not empty tensor
+        layer_idx = (
+            self.transformer_layers[0]
+            if layer_idx not in self.transformer_layers
+            else layer_idx
+        )
+        if len(self.key_cache) <= layer_idx or self.key_cache[layer_idx] is None:
+            return 0
+        return self.key_cache[layer_idx].shape[-2]
+
+    def get_query_offset(self, layer_idx: int = 0) -> int:
+        return self.get_seq_length(layer_idx=layer_idx)
+
+    def get_mask_sizes(self, cache_position, layer_idx: int) -> tuple[int, int]:
+        """
+        Return a tuple (kv_length, kv_offset) corresponding to the length and offset that will be returned for
+        the given layer at `layer_idx`.
+        The masks are then prepared according to the given lengths (kv_length, kv_offset) and patterns for each layer.
+        """
+        kv_offset = 0
+        # cache_position may be an int (new API) or a 1-D tensor (old API)
+        query_length = (
+            cache_position
+            if isinstance(cache_position, int)
+            else cache_position.shape[0]
+        )
+        past_seen_tokens = self.get_seq_length(layer_idx)
+        kv_length = query_length + past_seen_tokens
+        return kv_length, kv_offset
+
+    @property
+    def has_previous_state(self):
+        """We have a previous state if the last linear (conv) layer was already updated."""
+        if self.last_linear_layer == -1:
+            return False
+        return self.conv_states[self.last_linear_layer] is not None
+
+
+class KimiRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        dtype = hidden_states.dtype
+        x = hidden_states.float()
+        x = x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.variance_epsilon)
+        return self.weight * x.to(dtype)
+
+
+ALL_LAYERNORM_LAYERS.append(KimiRMSNorm)
+
+
+class KimiBlockSparseMLP(nn.Module):
+    def __init__(
+        self, config: KimiLinearConfig, hidden_size=None, intermediate_size=None
+    ):
+        super().__init__()
+        self.config = config
+        self.ffn_dim = (
+            config.intermediate_size if intermediate_size is None else intermediate_size
+        )
+        self.hidden_dim = config.hidden_size if hidden_size is None else hidden_size
+
+        self.w1 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)  # gate
+        self.w2 = nn.Linear(self.ffn_dim, self.hidden_dim, bias=False)  # down
+        self.w3 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)  # up
+
+        if config.hidden_act == "situ":
+            beta, linear_beta = _get_situ_activation_params(config)
+            self.act_fn = SituAndMul(
+                beta=beta,
+                linear_beta=linear_beta,
+            )
+        else:
+            self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, hidden_states):
+        if self.config.hidden_act == "situ":
+            gate_up = torch.cat(
+                [self.w1(hidden_states), self.w3(hidden_states)], dim=-1
+            )
+            current_hidden_states = self.act_fn(gate_up)
+        else:
+            current_hidden_states = self.act_fn(self.w1(hidden_states)) * self.w3(
+                hidden_states
+            )
+        current_hidden_states = self.w2(current_hidden_states)
+        return current_hidden_states
+
+
+class KimiMLP(nn.Module):
+    def __init__(
+        self, config: KimiLinearConfig, hidden_size=None, intermediate_size=None
+    ):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size if hidden_size is None else hidden_size
+        self.intermediate_size = (
+            config.intermediate_size if intermediate_size is None else intermediate_size
+        )
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        if config.hidden_act == "situ":
+            beta, linear_beta = _get_situ_activation_params(config)
+            self.act_fn = SituAndMul(
+                beta=beta,
+                linear_beta=linear_beta,
+            )
+        else:
+            self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        if self.config.hidden_act == "situ":
+            gate_up = torch.cat([self.gate_proj(x), self.up_proj(x)], dim=-1)
+            down_proj = self.down_proj(self.act_fn(gate_up))
+        else:
+            down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """Expand the key/value heads from `num_key_value_heads` to `num_attention_heads`."""
+    if n_rep == 1:
+        return hidden_states
+    return torch.repeat_interleave(hidden_states, dim=1, repeats=n_rep)
+
+
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: torch.Tensor | None,
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs: Unpack[TransformersKwargs],
+):
+    key = repeat_kv(key, module.num_key_value_groups)
+    value = repeat_kv(value, module.num_key_value_groups)
+
+    scores = torch.einsum("bhqd,bhkd->bhqk", query, key) * scaling
+    if attention_mask is not None:
+        scores = scores + attention_mask[:, :, :, : key.shape[-2]]
+
+    probs = F.softmax(scores, dim=-1, dtype=torch.float32).to(query.dtype)
+    probs = F.dropout(probs, p=dropout, training=module.training)
+    out = torch.einsum("bhqk,bhkd->bhqd", probs, value).transpose(1, 2).contiguous()
+
+    return out, probs
+
+
+class KimiMLAAttention(nn.Module):
+    """
+    Multi-Latent Attention adapted from deepseek-v3
+    """
+
+    def __init__(self, config: KimiLinearConfig, layer_idx: int):
+        nn.Module.__init__(self)
+        self.config = config
+        self.layer_idx = layer_idx
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+
+        self.attention_dropout = getattr(config, "attention_dropout", 0.0)
+
+        try:
+            self.q_lora_rank = config.q_lora_rank
+            self.qk_rope_head_dim = config.qk_rope_head_dim
+            self.kv_lora_rank = config.kv_lora_rank
+            self.v_head_dim = config.v_head_dim
+            self.qk_nope_head_dim = config.qk_nope_head_dim
+            self.q_head_dim = self.qk_nope_head_dim + self.qk_rope_head_dim
+            self.use_nope = config.mla_use_nope
+            self.scaling = self.q_head_dim ** (-0.5)
+        except Exception as e:
+            raise ValueError(
+                f"Kimi MLA config is not found or not properly formatted: {e}"
+            )
+
+        if self.q_lora_rank is not None:
+            self.q_a_proj = nn.Linear(
+                self.hidden_size,
+                self.q_lora_rank,
+                bias=False,
+            )
+            self.q_a_layernorm = KimiRMSNorm(self.q_lora_rank)
+            self.q_b_proj = nn.Linear(
+                self.q_lora_rank,
+                self.num_heads * self.q_head_dim,
+                bias=False,
+            )
+        else:
+            self.q_proj = nn.Linear(
+                self.hidden_size,
+                self.num_heads * self.q_head_dim,
+                bias=False,
+            )
+        self.kv_a_proj_with_mqa = nn.Linear(
+            self.hidden_size,
+            self.kv_lora_rank + self.qk_rope_head_dim,
+            bias=False,
+        )
+        self.kv_a_layernorm = KimiRMSNorm(self.kv_lora_rank)
+        self.kv_b_proj = nn.Linear(
+            self.kv_lora_rank,
+            self.num_heads
+            * (self.q_head_dim - self.qk_rope_head_dim + self.v_head_dim),
+            bias=False,
+        )
+        self.o_proj = nn.Linear(
+            self.num_heads * self.v_head_dim,
+            self.hidden_size,
+            bias=False,
+        )
+        self.is_causal = True
+        assert self.use_nope
+
+        self.use_output_gate = getattr(config, "mla_use_output_gate", False)
+        if self.use_output_gate:
+            projection_size = self.num_heads * self.v_head_dim
+            self.g_proj = nn.Linear(self.hidden_size, projection_size, bias=False)
+
+        self.rotary_emb = None
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        **kwargs,
+    ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[torch.Tensor] | None]:
+        batch_size, seq_length = hidden_states.shape[:-1]
+        query_shape = (batch_size, seq_length, -1, self.q_head_dim)
+        key_shape = (
+            batch_size,
+            seq_length,
+            -1,
+            self.qk_nope_head_dim + self.v_head_dim,
+        )
+
+        if self.q_lora_rank is not None:
+            q_states = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states)))
+        else:
+            q_states = self.q_proj(hidden_states)
+        q_states = q_states.view(query_shape).transpose(1, 2)
+        q_pass, q_rot = torch.split(
+            q_states, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1
+        )
+
+        compressed_kv = self.kv_a_proj_with_mqa(hidden_states)
+        k_pass, k_rot = torch.split(
+            compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1
+        )
+
+        k_pass = (
+            self.kv_b_proj(self.kv_a_layernorm(k_pass)).view(key_shape).transpose(1, 2)
+        )
+        k_pass, value_states = torch.split(
+            k_pass, [self.qk_nope_head_dim, self.v_head_dim], dim=-1
+        )
+
+        k_rot = k_rot.view(batch_size, 1, seq_length, self.qk_rope_head_dim)
+
+        k_rot = k_rot.expand(*k_pass.shape[:-1], -1)
+
+        query_states = torch.cat((q_pass, q_rot), dim=-1)
+        key_states = torch.cat((k_pass, k_rot), dim=-1)
+
+        if past_key_values is not None:
+            key_states, value_states = past_key_values.update(
+                key_states, value_states, self.layer_idx
+            )
+
+        if (
+            self.config._attn_implementation == "flash_attention_2"
+            and self.q_head_dim != self.v_head_dim
+        ):
+            value_states = F.pad(value_states, [0, self.q_head_dim - self.v_head_dim])
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            attention_interface = ALL_ATTENTION_FUNCTIONS[
+                self.config._attn_implementation
+            ]
+
+        attn_output, _ = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            **kwargs,
+        )
+
+        if (
+            self.config._attn_implementation == "flash_attention_2"
+            and self.q_head_dim != self.v_head_dim
+        ):
+            attn_output = attn_output[:, :, :, : self.v_head_dim]
+
+        attn_output = attn_output.reshape(batch_size, seq_length, -1).contiguous()
+        if self.use_output_gate:
+            g = self.g_proj(hidden_states).sigmoid()
+            attn_output = attn_output * g
+        attn_output = self.o_proj(attn_output)
+        return attn_output
+
+
+class KimiDeltaAttention(nn.Module):
+    def __init__(self, config: KimiLinearConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.mode = "chunk"
+
+        self.hidden_size = config.hidden_size
+        self.conv_size = config.linear_attn_config["short_conv_kernel_size"]
+        self.head_dim = config.linear_attn_config["head_dim"]
+        self.num_heads = config.linear_attn_config["num_heads"]
+        self.head_k_dim = self.head_dim
+        self.num_k_heads = self.num_heads
+
+        self.layer_idx = layer_idx
+
+        assert self.mode in [
+            "chunk",
+            "fused_recurrent",
+        ], f"Not supported mode `{self.mode}`."
+
+        projection_k_size = self.head_k_dim * self.num_k_heads
+        projection_size = self.head_dim * self.num_heads
+
+        self.q_proj = nn.Linear(self.hidden_size, projection_k_size, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, projection_k_size, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, projection_size, bias=False)
+
+        self.q_conv1d = ShortConvolution(
+            hidden_size=projection_k_size,
+            kernel_size=self.conv_size,
+            activation="silu",
+        )
+        self.k_conv1d = ShortConvolution(
+            hidden_size=projection_k_size,
+            kernel_size=self.conv_size,
+            activation="silu",
+        )
+        self.v_conv1d = ShortConvolution(
+            hidden_size=projection_size,
+            kernel_size=self.conv_size,
+            activation="silu",
+        )
+
+        self.A_log = torch.nn.Parameter(
+            torch.log(torch.empty(self.num_heads, dtype=torch.float32).uniform_(1, 16))
+        )
+
+        self.f_a_proj = nn.Linear(self.hidden_size, self.head_dim, bias=False)
+        self.f_b_proj = nn.Linear(self.head_dim, projection_size, bias=False)
+
+        self.dt_bias = nn.Parameter(torch.empty(projection_size, dtype=torch.float32))
+
+        self.b_proj = nn.Linear(self.hidden_size, self.num_heads, bias=False)
+
+        self.use_full_rank_gate = config.linear_attn_config.get(
+            "use_full_rank_gate", False
+        )
+        self.gate_lower_bound = config.linear_attn_config.get("gate_lower_bound", None)
+        if self.use_full_rank_gate:
+            self.g_proj = nn.Linear(self.hidden_size, projection_size, bias=False)
+        else:
+            self.g_a_proj = nn.Linear(self.hidden_size, self.head_dim, bias=False)
+            self.g_b_proj = nn.Linear(self.head_dim, projection_size, bias=False)
+
+        self.o_norm = FusedRMSNormGated(
+            self.head_dim, eps=config.rms_norm_eps, activation="sigmoid"
+        )
+        self.o_proj = nn.Linear(projection_size, self.hidden_size, bias=False)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
+        cache_params: KimiDynamicCache | None = None,
+        **kwargs: Unpack[dict],
+    ) -> tuple[torch.Tensor, torch.Tensor | None, Cache | None]:
+        if attention_mask is not None:
+            if attention_mask.dim() != 2:
+                attention_mask = kwargs.get("padding_mask")
+
+            if attention_mask is not None and attention_mask.dim() != 2:
+                raise ValueError(
+                    "attention_mask must be a 0-1 matrix of shape [batch_size, seq_len] "
+                    "(0 = padding). 3D masks are not supported here.",
+                )
+        use_cache = cache_params is not None
+        batch_size, q_len, _ = hidden_states.shape
+        mode = "fused_recurrent" if use_cache and q_len == 1 else self.mode
+        if self.training:
+            assert mode == "chunk", "Only chunk mode is supported in training."
+
+        cu_seqlens = kwargs.get("cu_seqlens")
+        indices = None
+        if attention_mask is not None:
+            indices, cu_seqlens, _ = get_unpad_data(attention_mask[:, -q_len:])
+            hidden_states = index_first_axis(
+                rearrange(hidden_states, "b s ... -> (b s) ..."), indices
+            ).unsqueeze(0)
+
+        conv_state_q, conv_state_k, conv_state_v = None, None, None
+        recurrent_state = None
+        if cache_params is not None:
+            if cache_params.conv_states[self.layer_idx] is not None:
+                conv_state_q, conv_state_k, conv_state_v = cache_params.conv_states[
+                    self.layer_idx
+                ]
+            recurrent_state = cache_params.recurrent_states[self.layer_idx]
+
+        q_proj_states = self.q_proj(hidden_states)
+        k_proj_states = self.k_proj(hidden_states)
+        v_proj_states = self.v_proj(hidden_states)
+        q, conv_state_q = self.q_conv1d(
+            x=q_proj_states,
+            cache=conv_state_q,
+            output_final_state=use_cache,
+            cu_seqlens=cu_seqlens,
+        )
+        k, conv_state_k = self.k_conv1d(
+            x=k_proj_states,
+            cache=conv_state_k,
+            output_final_state=use_cache,
+            cu_seqlens=cu_seqlens,
+        )
+        v, conv_state_v = self.v_conv1d(
+            x=v_proj_states,
+            cache=conv_state_v,
+            output_final_state=use_cache,
+            cu_seqlens=cu_seqlens,
+        )
+        g = self.f_b_proj(self.f_a_proj(hidden_states))
+        g = rearrange(g, "... (h d) -> ... h d", d=self.head_dim)
+        beta = self.b_proj(hidden_states).float()
+
+        q, k = map(
+            lambda x: rearrange(x, "... (h d) -> ... h d", d=self.head_k_dim), (q, k)
+        )
+        v = rearrange(v, "... (h d) -> ... h d", d=self.head_dim)
+
+        if mode == "chunk":
+            o, recurrent_state = chunk_kda(
+                q=q,
+                k=k,
+                v=v,
+                g=g,
+                beta=beta,
+                A_log=self.A_log,
+                dt_bias=self.dt_bias,
+                initial_state=recurrent_state,
+                output_final_state=True,
+                use_qk_l2norm_in_kernel=True,
+                use_gate_in_kernel=True,
+                use_beta_sigmoid_in_kernel=True,
+                safe_gate=self.gate_lower_bound is not None,
+                lower_bound=self.gate_lower_bound,
+                transpose_state_layout=True,
+                cu_seqlens=cu_seqlens,
+            )
+        else:
+            o, recurrent_state = fused_recurrent_kda(
+                q=q,
+                k=k,
+                v=v,
+                g=g,
+                beta=beta,
+                A_log=self.A_log,
+                dt_bias=self.dt_bias,
+                initial_state=recurrent_state,
+                output_final_state=True,
+                use_qk_l2norm_in_kernel=True,
+                use_gate_in_kernel=True,
+                use_beta_sigmoid_in_kernel=True,
+                lower_bound=self.gate_lower_bound,
+                transpose_state_layout=True,
+                cu_seqlens=cu_seqlens,
+            )
+        if cache_params is not None:
+            cache_params.recurrent_states[self.layer_idx] = recurrent_state
+            cache_params.conv_states[self.layer_idx] = (
+                conv_state_q,
+                conv_state_k,
+                conv_state_v,
+            )
+
+        if self.use_full_rank_gate:
+            g = self.g_proj(hidden_states)
+        else:
+            g = self.g_b_proj(self.g_a_proj(hidden_states))
+        g = rearrange(g, "... (h d) -> ... h d", d=self.head_dim)
+        o = self.o_norm(o, g)
+
+        o = rearrange(o, "b t h d -> b t (h d)")
+        o = self.o_proj(o)
+        if attention_mask is not None:
+            o = pad_input(o.squeeze(0), indices, batch_size, q_len)
+
+        return o
+
+
+class KimiMoEGate(nn.Module):
+    """
+    MoEGate adapted from Deepseek-V3.
+    Parameter correspondences:
+        num_experts -> n_routed_experts
+        num_experts_per_token -> num_experts_per_tok
+        num_expert_group -> n_group
+        moe_router_activation_func -> scoring_func
+    """
+
+    def __init__(self, config: KimiLinearConfig):
+        super().__init__()
+        self.config = config
+        self.top_k = config.num_experts_per_token
+        self.num_experts = config.num_experts
+        self.routed_scaling_factor = config.routed_scaling_factor
+        self.moe_router_activation_func = config.moe_router_activation_func
+        self.num_expert_group = getattr(config, "num_expert_group", 1)
+        self.topk_group = getattr(config, "topk_group", 1)
+
+        # topk selection algorithm
+        self.moe_renormalize = config.moe_renormalize
+        self.gating_dim = config.hidden_size
+        self.weight = nn.Parameter(
+            torch.empty((self.num_experts, self.gating_dim)),
+        )
+
+        self.e_score_correction_bias = nn.Parameter(
+            torch.empty(self.num_experts),
+        )
+        self.reset_parameters()
+
+    def reset_parameters(self) -> None:
+        import torch.nn.init as init
+
+        init.kaiming_uniform_(self.weight, a=math.sqrt(5))
+
+    def forward(self, hidden_states):
+        bsz, seq_len, h = hidden_states.shape
+        # compute gating score
+        hidden_states = hidden_states.view(-1, h)
+        logits = F.linear(
+            hidden_states.type(torch.float32),
+            self.weight.type(torch.float32),
+            None,
+        )
+        if self.moe_router_activation_func == "sigmoid":
+            scores = logits.sigmoid()
+        elif self.moe_router_activation_func == "softmax":
+            scores = logits.softmax(dim=1)
+        else:
+            raise NotImplementedError(
+                f"insupportable scoring function for MoE gating: {self.moe_router_activation_func}",
+            )
+
+        # select top-k experts
+        scores = scores.view(bsz * seq_len, -1)
+        scores_for_choice = scores + self.e_score_correction_bias.unsqueeze(0)
+        if self.num_expert_group > 1 and self.num_expert_group > self.topk_group:
+            group_scores = (
+                scores_for_choice.view(bsz * seq_len, self.num_expert_group, -1)
+                .topk(2, dim=-1)[0]
+                .sum(dim=-1)
+            )  # [n, num_expert_group]
+            group_idx = torch.topk(
+                group_scores,
+                k=self.topk_group,
+                dim=-1,
+                sorted=False,
+            )[1]  # [n, top_k_group]
+            group_mask = torch.zeros_like(group_scores)  # [n, num_expert_group]
+            group_mask.scatter_(1, group_idx, 1)  # [n, num_expert_group]
+            score_mask = (
+                group_mask.unsqueeze(-1)
+                .expand(
+                    bsz * seq_len,
+                    self.num_expert_group,
+                    self.num_experts // self.num_expert_group,
+                )
+                .reshape(bsz * seq_len, -1)
+            )  # [n, e]
+            tmp_scores = scores_for_choice.masked_fill(
+                ~score_mask.bool(), float("-inf")
+            )  # [n, e]
+        else:
+            tmp_scores = scores_for_choice
+        _, topk_idx = torch.topk(
+            tmp_scores,
+            k=self.top_k,
+            dim=-1,
+            sorted=False,
+        )
+        topk_weight = scores.gather(1, topk_idx)
+
+        # norm gate to sum 1
+        if self.top_k > 1 and self.moe_renormalize:
+            denominator = topk_weight.sum(dim=-1, keepdim=True) + 1e-20
+            topk_weight = topk_weight / denominator
+        # must multiply the scaling factor
+        topk_weight = topk_weight * self.routed_scaling_factor
+
+        return topk_idx, topk_weight
+
+
+class KimiSparseMoeBlock(nn.Module):
+    """
+    Adapted from Deepseek-V3's MOE implementation
+    The namings are consistent with Kimi's version.
+    """
+
+    def __init__(self, config: KimiLinearConfig):
+        super().__init__()
+        self.config = config
+        self.hidden_dim = config.hidden_size
+        self.num_experts = config.num_experts
+        self.top_k = config.num_experts_per_token
+        self.moe_renormalize = config.moe_renormalize
+
+        self.use_latent_moe = (
+            getattr(config, "routed_expert_hidden_size", None) is not None
+        )
+        self.moe_hidden_size = (
+            config.routed_expert_hidden_size
+            if self.use_latent_moe
+            else config.hidden_size
+        )
+        self.latent_moe_use_norm = getattr(config, "latent_moe_use_norm", False)
+
+        self.ep_size = 1
+        self.experts_per_rank = config.num_experts
+        self.ep_rank = 0
+        self.experts = nn.ModuleList(
+            [
+                KimiBlockSparseMLP(
+                    config,
+                    hidden_size=self.moe_hidden_size,
+                    intermediate_size=config.moe_intermediate_size,
+                )
+                for _ in range(config.num_experts)
+            ],
+        )
+        self.gate = KimiMoEGate(config)
+        if config.num_shared_experts is not None:
+            intermediate_size = config.moe_intermediate_size * config.num_shared_experts
+            self.shared_experts = KimiMLP(
+                config=config,
+                intermediate_size=intermediate_size,
+            )
+
+        if self.use_latent_moe:
+            self.routed_expert_down_proj = nn.Linear(
+                config.hidden_size,
+                self.moe_hidden_size,
+                bias=False,
+            )
+            self.routed_expert_up_proj = nn.Linear(
+                self.moe_hidden_size,
+                config.hidden_size,
+                bias=False,
+            )
+            if self.latent_moe_use_norm:
+                self.routed_expert_norm = KimiRMSNorm(
+                    self.moe_hidden_size,
+                    eps=config.rms_norm_eps,
+                )
+
+    def forward(self, hidden_states):
+        identity = hidden_states
+        orig_shape = hidden_states.shape
+        topk_idx, topk_weight = self.gate(hidden_states)
+        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+
+        if self.use_latent_moe:
+            hidden_states = self.routed_expert_down_proj(hidden_states)
+
+        if not self.training:
+            y = self.moe_infer(hidden_states, topk_idx, topk_weight)
+        else:
+            y = self.moe_train(hidden_states, topk_idx, topk_weight)
+
+        if self.use_latent_moe:
+            if self.latent_moe_use_norm:
+                y = self.routed_expert_norm(y)
+            y = self.routed_expert_up_proj(y)
+
+        y = y.view(*orig_shape)
+
+        if self.config.num_shared_experts is not None:
+            y = y + self.shared_experts(identity)
+        return y
+
+    def moe_train(self, x, topk_ids, topk_weight):
+        """Training-compatible MoE dispatch with gradient flow."""
+        y = torch.zeros_like(x)
+
+        with torch.no_grad():
+            expert_mask = F.one_hot(topk_ids, self.num_experts).permute(2, 1, 0)
+
+        for expert_idx, expert in enumerate(self.experts):
+            top_k_pos, token_indices = torch.where(expert_mask[expert_idx])
+
+            if get_calibrate_all_experts_flag():
+                expert_out = expert(x)[token_indices]
+            else:
+                expert_out = expert(x[token_indices])
+
+            expert_weights = topk_weight[token_indices, top_k_pos, None]
+            y.index_add_(0, token_indices, (expert_out * expert_weights).to(y.dtype))
+
+        return y
+
+    @torch.no_grad()
+    def moe_infer(self, x, topk_ids, topk_weight):
+        cnts = topk_ids.new_zeros((topk_ids.shape[0], len(self.experts)))
+        cnts.scatter_(1, topk_ids, 1)
+        tokens_per_expert = cnts.sum(dim=0)
+        idxs = topk_ids.view(-1).argsort()
+        sorted_tokens = x[idxs // topk_ids.shape[1]]
+
+        tokens_per_expert = tokens_per_expert.cpu().numpy()
+
+        outputs = []
+        start_idx = 0
+        for i, num_tokens in enumerate(tokens_per_expert):
+            end_idx = start_idx + num_tokens
+            if num_tokens == 0:
+                continue
+            expert = self.experts[i + self.ep_rank * self.experts_per_rank]
+            tokens_for_this_expert = sorted_tokens[start_idx:end_idx]
+            expert_out = expert(tokens_for_this_expert)
+            outputs.append(expert_out)
+            start_idx = end_idx
+
+        outs = torch.cat(outputs, dim=0) if len(outputs) else sorted_tokens.new_empty(0)
+
+        new_x = torch.empty_like(outs)
+        new_x[idxs] = outs
+        final_out = (
+            new_x.view(*topk_ids.shape, -1)
+            .type(topk_weight.dtype)
+            .mul_(topk_weight.unsqueeze(dim=-1))
+            .sum(dim=1)
+            .type(new_x.dtype)
+        )
+        return final_out
+
+
+class KimiDecoderLayer(nn.Module):
+    def __init__(self, config: KimiLinearConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.config = config
+        self.layer_idx = layer_idx
+        if config.is_kda_layer(layer_idx):
+            self.is_linear_attn = True
+            self.self_attn = KimiDeltaAttention(config=config, layer_idx=layer_idx)
+        elif config.is_mla:
+            self.is_linear_attn = False
+            self.self_attn = KimiMLAAttention(config=config, layer_idx=layer_idx)
+        else:
+            raise NotImplementedError
+        if (
+            config.num_experts is not None
+            and layer_idx >= config.first_k_dense_replace
+            and layer_idx % getattr(config, "moe_layer_freq", 1) == 0
+        ):
+            self.block_sparse_moe = KimiSparseMoeBlock(config)
+        else:
+            self.mlp = KimiMLP(config)
+        self.input_layernorm = KimiRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = KimiRMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+
+        # Attention residual
+        self.use_attn_residuals = (
+            getattr(config, "attn_res_block_size", None) is not None
+        )
+        if self.use_attn_residuals:
+            self.attn_res_block_size = config.attn_res_block_size
+            self.self_attention_res_norm = KimiRMSNorm(
+                config.hidden_size, eps=config.rms_norm_eps
+            )
+            self.mlp_res_norm = KimiRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+            self.self_attention_res_proj = nn.Linear(config.hidden_size, 1, bias=False)
+            self.mlp_res_proj = nn.Linear(config.hidden_size, 1, bias=False)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: tuple[torch.Tensor] | None = None,
+        output_attentions: bool | None = False,
+        use_cache: bool | None = False,
+        block_residual: torch.Tensor | None = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ):
+        if self.use_attn_residuals:
+            return self._forward_attn_residual(
+                hidden_states,
+                attention_mask,
+                position_ids,
+                past_key_values,
+                output_attentions,
+                use_cache,
+                block_residual,
+                **kwargs,
+            )
+
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        if self.is_linear_attn is False:
+            hidden_states = self.self_attn(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_values=past_key_values,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+                **kwargs,
+            )
+        else:
+            hidden_states = self.self_attn(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                cache_params=past_key_values,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+                **kwargs,
+            )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        if hasattr(self, "block_sparse_moe"):
+            hidden_states = self.block_sparse_moe(hidden_states)
+        else:
+            hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        return hidden_states
+
+    def _forward_attn_residual(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: tuple[torch.Tensor] | None = None,
+        output_attentions: bool | None = False,
+        use_cache: bool | None = False,
+        block_residual: torch.Tensor | None = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ):
+        batch_size, seq_len, hidden_size = hidden_states.shape
+        prefix_sum = hidden_states
+
+        if block_residual is not None and block_residual.shape[1] > 0:
+            hidden_states = _apply_attn_res(
+                prefix_sum.view(-1, hidden_size),
+                block_residual,
+                self.self_attention_res_proj,
+                self.self_attention_res_norm,
+            ).view(batch_size, seq_len, hidden_size)
+
+        if self.layer_idx % self.attn_res_block_size == 0:
+            block_residual = torch.cat(
+                [block_residual, prefix_sum.view(-1, hidden_size).unsqueeze(1)], dim=1
+            )
+            prefix_sum = None
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        if self.is_linear_attn is False:
+            hidden_states = self.self_attn(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_values=past_key_values,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+                **kwargs,
+            )
+        else:
+            hidden_states = self.self_attn(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                cache_params=past_key_values,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+                **kwargs,
+            )
+
+        if prefix_sum is not None:
+            prefix_sum = prefix_sum + hidden_states
+        else:
+            prefix_sum = hidden_states
+
+        hidden_states = _apply_attn_res(
+            prefix_sum.view(-1, hidden_size),
+            block_residual,
+            self.mlp_res_proj,
+            self.mlp_res_norm,
+        ).view(batch_size, seq_len, hidden_size)
+
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        if hasattr(self, "block_sparse_moe"):
+            hidden_states = self.block_sparse_moe(hidden_states)
+        else:
+            hidden_states = self.mlp(hidden_states)
+
+        if prefix_sum is None:
+            prefix_sum = hidden_states
+        else:
+            prefix_sum = prefix_sum + hidden_states
+
+        return prefix_sum, block_residual
+
+
+class KimiPreTrainedModel(PreTrainedModel):
+    config_class = KimiLinearConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["KimiDecoderLayer"]
+    _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
+    _can_record_outputs = {
+        "router_logits": OutputRecorder(KimiBlockSparseMLP, index=1),
+        "hidden_states": KimiDecoderLayer,
+        "attentions": KimiMLAAttention,
+    }
+    _is_stateful = True
+
+    def _init_weights(self, module):
+        # HOTFIX: disk offloading attempts to initialize the meta tensors
+        # but this is bad programming: we shouldn't be initializing these
+        # params in the first place
+        # the init attempt attempts to get `module.weight`, which DNE for qmodels
+
+        return
+
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+
+def _apply_attn_res(prefix_sum, block_residual, proj, norm):
+    """
+    prefix_sum:     (num_tokens, hidden_size)
+    block_residual: (num_tokens, num_blocks, hidden_size)
+    """
+    v = torch.cat((block_residual, prefix_sum.unsqueeze(1)), dim=1)
+    v_float = v.float()
+    variance = v_float.pow(2).mean(-1, keepdim=True)
+    k = v_float * torch.rsqrt(variance + norm.variance_epsilon)
+    score_weight = norm.weight.float() * proj.weight.squeeze(0).float()
+    scores = (k * score_weight).sum(-1)
+    probs = scores.softmax(-1).unsqueeze(1)
+    hidden_states = torch.matmul(probs, v_float).squeeze(1)
+    return hidden_states.to(v.dtype)
+
+
+class KimiLinearModel(KimiPreTrainedModel):
+    def __init__(self, config: KimiLinearConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(
+            config.vocab_size, config.hidden_size, self.padding_idx
+        )
+        self.layers = nn.ModuleList(
+            [
+                KimiDecoderLayer(config, layer_idx)
+                for layer_idx in range(config.num_hidden_layers)
+            ]
+        )
+        self.norm = KimiRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        self.use_attn_residuals = (
+            getattr(config, "attn_res_block_size", None) is not None
+        )
+        if self.use_attn_residuals:
+            self.output_attn_res_norm = KimiRMSNorm(
+                config.hidden_size, eps=config.rms_norm_eps
+            )
+            self.output_attn_res_proj = nn.Linear(config.hidden_size, 1, bias=False)
+
+        from transformers.utils import is_flash_attn_2_available as _fa2_avail
+
+        _requested = getattr(config, "_attn_implementation", None)
+        if _requested not in (None, "flash_attention_2") or not _fa2_avail():
+            # Fall back gracefully when flash-attn2 is unavailable or a different impl is requested
+            if _requested == "flash_attention_2" and not _fa2_avail():
+                logger.warning_once(
+                    "flash_attention_2 requested but not available; falling back to sdpa."
+                )
+            config._attn_implementation = (
+                _requested if _requested not in (None, "flash_attention_2") else "eager"
+            )
+        else:
+            config._attn_implementation = "flash_attention_2"
+
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def _update_linear_attn_mask(self, attention_mask, cache_position):
+        """
+        NOTE: Left-padding is used for linear attention mask.
+        No need for zeroing states when
+            1. Cached forward
+            2. Attending to all inputs
+        """
+        linear_attn_mask = attention_mask
+        if cache_position[0] > 0 or (
+            attention_mask is not None and torch.all(attention_mask == 1)
+        ):
+            linear_attn_mask = None
+        return linear_attn_mask
+
+    @check_model_inputs
+    # @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        cache_position: torch.LongTensor | None = None,
+        use_cache: bool | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple | BaseModelOutputWithPast:
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        if (input_ids is None) and (inputs_embeds is None):
+            raise ValueError(
+                "You must specify exactly one of input_ids or inputs_embeds"
+            )
+
+        # Get inputs_embeds
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if use_cache and past_key_values is None:
+            past_key_values = KimiDynamicCache(config=self.config)
+
+        if cache_position is None:
+            past_seen_tokens = (
+                past_key_values.get_seq_length() if past_key_values is not None else 0
+            )
+            cache_position: torch.Tensor = torch.arange(
+                past_seen_tokens,
+                past_seen_tokens + inputs_embeds.shape[1],
+                device=inputs_embeds.device,
+            )
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        causal_mask = create_causal_mask(
+            config=self.config,
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            position_ids=position_ids,
+        )
+        linear_attn_mask = self._update_linear_attn_mask(attention_mask, cache_position)
+
+        hidden_states = inputs_embeds
+        if past_key_values is not None:
+            assert isinstance(past_key_values, KimiDynamicCache)
+
+        block_residual = None
+        if self.use_attn_residuals:
+            block_residual = hidden_states.new_zeros(
+                hidden_states.shape[0] * hidden_states.shape[1],
+                0,
+                hidden_states.shape[2],
+            )
+
+        for decoder_layer in self.layers:
+            layer_mask = (
+                linear_attn_mask if decoder_layer.is_linear_attn else causal_mask
+            )
+
+            if self.use_attn_residuals:
+                hidden_states, block_residual = decoder_layer(
+                    hidden_states,
+                    attention_mask=layer_mask,
+                    past_key_values=past_key_values,
+                    cache_position=cache_position,
+                    block_residual=block_residual,
+                    **kwargs,
+                )
+            else:
+                hidden_states = decoder_layer(
+                    hidden_states,
+                    attention_mask=layer_mask,
+                    past_key_values=past_key_values,
+                    cache_position=cache_position,
+                    **kwargs,
+                )
+
+        if self.use_attn_residuals:
+            hidden_states = self._apply_output_attn_res(hidden_states, block_residual)
+
+        hidden_states = self.norm(hidden_states)
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+        )
+
+    def _apply_output_attn_res(self, hidden_states, block_residual):
+        batch_size, seq_len, hidden_size = hidden_states.shape
+        return _apply_attn_res(
+            hidden_states.view(-1, hidden_size),
+            block_residual,
+            self.output_attn_res_proj,
+            self.output_attn_res_norm,
+        ).view(batch_size, seq_len, hidden_size)
+
+
+class KimiLinearForCausalLM(KimiPreTrainedModel, GenerationMixin):
+    @classmethod
+    def _supports_default_dynamic_cache(cls) -> bool:
+        return False
+
+    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = KimiLinearModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @can_return_tuple
+    # @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: list[torch.FloatTensor] | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        labels: torch.LongTensor | None = None,
+        use_cache: bool | None = None,
+        output_attentions: bool | None = None,
+        output_hidden_states: bool | None = None,
+        generation_mode: bool | None = None,
+        return_dict: bool | None = None,
+        cache_position: torch.LongTensor | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple | CausalLMOutputWithPast:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+        """
+
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            cache_position=cache_position,
+        )
+
+        logits = outputs[0]
+        if generation_mode:
+            logits = logits[:, -1:]
+        logits = self.lm_head(logits)
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits, labels, self.vocab_size, **kwargs)
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

modeling_kimi_linear.py

@@ -0,0 +1 @@
+from .modeling_kimi_k3_linear import *

recipe.yaml

@@ -0,0 +1,7 @@
+default_stage:
+  default_modifiers:
+    QuantizationModifier:
+      targets: [Linear]
+      ignore: ['re:.*self_attn.*', 're:.*shared_experts.*', 're:.*lm_head.*', 're:.*vision_tower.*']
+      scheme: MXFP4A16
+      bypass_divisibility_checks: false

tiktoken.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b6c497a7469b33ced9c38afb1ad6e47f03f5e5dc05f15930799210ec050c5103
+size 2795286

tokenization_kimi.py

@@ -0,0 +1,430 @@
+import os
+from logging import getLogger
+from pathlib import Path
+from shutil import copyfile
+from typing import Dict, Iterator, List, Optional, Tuple, Union, cast
+
+import tiktoken
+from tiktoken.load import load_tiktoken_bpe
+from tokenizers import AddedToken
+from transformers.convert_slow_tokenizer import bytes_to_unicode
+from transformers.tokenization_utils import PreTrainedTokenizer
+
+try:
+    from .encoding_k3 import build_chat_segments, is_batched_conversation
+except ImportError:  # pragma: no cover - supports direct file execution/import.
+    from encoding_k3 import build_chat_segments, is_batched_conversation
+
+logger = getLogger(__name__)
+VOCAB_FILES_NAMES = {"vocab_file": "tiktoken.model"}
+
+
+class TikTokenTokenizer(PreTrainedTokenizer):
+    """
+    Tokenizing and encoding/decoding text using the Tiktoken tokenizer. See megatron/tokenizer/tiktoken_tokenizer.py.
+
+    This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer to
+    this superclass for more information regarding those methods.
+
+    Args:
+        vocab_file (`str`):
+            The path to the Tiktoken model file.
+        bos_token (`str` or `tokenizers.AddedToken`, *optional*, defaults to `"<|begin_of_text|>",`):
+            The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.
+        eos_token (`str` or `tokenizers.AddedToken`, *optional*, defaults to `"<|end_of_text|>"`):
+            The end of sequence token.
+        unk_token (`str` or `tokenizers.AddedToken`, *optional*, defaults to `"<|reserved_special_token_249|>"`):
+            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+            token instead. The second to last item in special_tokens.
+        pad_token (`str` or `tokenizers.AddedToken`, *optional*, defaults to `"<|reserved_special_token_250|>"`):
+            The token used for padding, for example when batching sequences of different lengths.
+        additional_special_tokens (list of `str`, *optional*):
+            A tuple or a list of additional tokens, which will be marked as `special`, meaning that they will be
+            skipped when decoding if `skip_special_tokens` is set to `True`.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+
+    model_input_names = ["input_ids", "attention_mask"]
+
+    special_tokens: Dict[str, int]
+
+    num_reserved_special_tokens = 256
+
+    pat_str = "|".join(
+        [
+            r"""[\p{Han}]+""",
+            r"""[^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]*[\p{Ll}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]+(?i:'s|'t|'re|'ve|'m|'ll|'d)?""",
+            r"""[^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]+[\p{Ll}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]*(?i:'s|'t|'re|'ve|'m|'ll|'d)?""",
+            r"""\p{N}{1,3}""",
+            r""" ?[^\s\p{L}\p{N}]+[\r\n]*""",
+            r"""\s*[\r\n]+""",
+            r"""\s+(?!\S)""",
+            r"""\s+""",
+        ]
+    )
+
+    def __init__(
+        self,
+        vocab_file,
+        bos_token: Union[str, AddedToken] = "[BOS]",
+        eos_token: Union[str, AddedToken] = "[EOS]",
+        unk_token: Union[str, AddedToken, None] = None,
+        pad_token: Union[str, AddedToken, None] = None,
+        additional_special_tokens: List[str] = None,
+        added_tokens_decoder: Optional[dict] = None,
+        **kwargs,
+    ):
+        assert os.path.isfile(vocab_file), vocab_file
+
+        if additional_special_tokens is None:
+            additional_special_tokens = [
+                "<|im_end|>",
+                "<|im_user|>",
+                "<|im_assistant|>",
+                "<|start_header_id|>",
+                "<|end_header_id|>",
+                "[EOT]",
+                "<|im_system|>",
+                "<|im_middle|>",
+            ]
+
+        if added_tokens_decoder:
+            special_tokens_mapping = {
+                i: added_tokens_decoder[i].content for i in added_tokens_decoder
+            }
+        else:
+            special_tokens_mapping = {}
+
+        self.vocab_file = vocab_file
+        mergeable_ranks = load_tiktoken_bpe(vocab_file)
+        num_base_tokens = len(mergeable_ranks)
+        self.special_tokens = {
+            special_tokens_mapping.get(i, f"<|reserved_token_{i}|>"): i
+            for i in range(
+                num_base_tokens, num_base_tokens + self.num_reserved_special_tokens
+            )
+        }
+
+        self.model = tiktoken.Encoding(
+            name=Path(vocab_file).name,
+            pat_str=self.pat_str,
+            mergeable_ranks=mergeable_ranks,
+            special_tokens=self.special_tokens,
+        )
+        logger.info(f"Reloaded tiktoken model from {vocab_file}")
+
+        self.n_words: int = self.model.n_vocab
+        # BOS / EOS token IDs
+        self.bos_id: int = self.special_tokens[str(bos_token)]
+        self.eos_id: int = self.special_tokens[str(eos_token)]
+        logger.info(
+            f"#words: {self.n_words} - BOS ID: {self.bos_id} - EOS ID: {self.eos_id}"
+        )
+
+        self.pad_id: int = self.special_tokens[str(pad_token)]
+        self.unk_id: int = self.special_tokens[str(unk_token)]
+
+        self.byte_encoder = bytes_to_unicode()
+        self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
+
+        self.decoder = {}
+        for i in range(self.n_words):
+            # Taken from https://gist.github.com/xenova/a452a6474428de0182b17605a98631ee
+            decoding = "".join(
+                [
+                    self.byte_encoder[ord(char)]
+                    for char in self.model.decode_single_token_bytes(i).decode(
+                        "latin-1"
+                    )
+                ]
+            )
+            self.decoder[i] = decoding
+
+        self.encoder = {}
+        for i in range(self.n_words):
+            if i in self.decoder:
+                self.encoder[self.decoder[i]] = i
+
+        super().__init__(
+            bos_token=bos_token,
+            eos_token=eos_token,
+            unk_token=unk_token,
+            pad_token=pad_token,
+            additional_special_tokens=additional_special_tokens,
+            added_tokens_decoder=added_tokens_decoder,
+            **kwargs,
+        )
+        self.all_special_ids_set = set(self.all_special_ids)
+
+    def _encode_text_piece(
+        self, text: str, allow_special_tokens: bool = True
+    ) -> List[int]:
+        # The tiktoken tokenizer can handle <=400k chars without
+        # pyo3_runtime.PanicException.
+        TIKTOKEN_MAX_ENCODE_CHARS = 400_000
+
+        # https://github.com/openai/tiktoken/issues/195
+        # Here we iterate over subsequences and split if we exceed the limit
+        # of max consecutive non-whitespace or whitespace characters.
+        MAX_NO_WHITESPACES_CHARS = 25_000
+
+        t: List[int] = []
+        for i in range(0, len(text), TIKTOKEN_MAX_ENCODE_CHARS):
+            for substr in self._split_whitespaces_or_nonwhitespaces(
+                text[i : i + TIKTOKEN_MAX_ENCODE_CHARS],
+                MAX_NO_WHITESPACES_CHARS,
+            ):
+                if allow_special_tokens:
+                    t.extend(
+                        # structural markers: encode <|...|> as their special token IDs
+                        self.model.encode(
+                            substr,
+                            allowed_special="all",
+                        )
+                    )
+                else:
+                    t.extend(
+                        # user/tool text: encode any <|...|> as ordinary BPE tokens (never as control tokens)
+                        self.model.encode(
+                            substr,
+                            disallowed_special=(),
+                        )
+                    )
+
+        return t
+
+    def encode(
+        self, text: str, allow_special_tokens: bool = True, **kwargs
+    ) -> List[int]:
+        """
+        Encodes a string into a list of token IDs.
+
+        Args:
+            text (str): The input string to be encoded.
+
+        Returns:
+            list[int]: A list of token IDs.
+        """
+        # If there are other args, we should call super().encode because there are a lot of code
+        # to handle those args. supper().encode finally will call _tokenize and _convert_token_to_id.
+        # NOTE: our encode method is not compatible with the super().encode method,
+        #   e.g. split_special_tokens' default is True in our encode method.
+        if len(kwargs) > 0:
+            logger.warning(f"Calling super().encode with {kwargs}")
+            return super().encode(text, **kwargs)
+
+        assert type(text) is str
+        return self._encode_text_piece(text, allow_special_tokens=allow_special_tokens)
+
+    def decode(self, token_ids: Union[int, List[int]], **kwargs) -> str:
+        """
+        Decodes a list of token IDs into a string.
+
+        Args:
+            token_ids (List[int]): The list of token IDs to be decoded.
+
+        Returns:
+            str: The decoded string.
+        """
+        # If there are other args, we should call super().decode because there are a lot of code
+        # to handle those args. supper().encode finally will call convert_tokens_to_string and _convert_id_to_token.
+        if len(kwargs) > 0:
+            return super().decode(token_ids, **kwargs)
+
+        if type(token_ids) is int:
+            token_ids = [token_ids]
+
+        return self.model.decode(cast(List[int], token_ids))
+
+    @staticmethod
+    def _split_whitespaces_or_nonwhitespaces(
+        s: str, max_consecutive_slice_len: int
+    ) -> Iterator[str]:
+        """
+        Splits the string `s` so that each substring contains no more than `max_consecutive_slice_len`
+        consecutive whitespaces or consecutive non-whitespaces.
+        """
+        current_slice_len = 0
+        current_slice_is_space = s[0].isspace() if len(s) > 0 else False
+        slice_start = 0
+
+        for i in range(len(s)):
+            is_now_space = s[i].isspace()
+
+            if current_slice_is_space ^ is_now_space:
+                current_slice_len = 1
+                current_slice_is_space = is_now_space
+            else:
+                current_slice_len += 1
+                if current_slice_len > max_consecutive_slice_len:
+                    yield s[slice_start:i]
+                    slice_start = i
+                    current_slice_len = 1
+        yield s[slice_start:]
+
+    def _encode_chat_segments(self, segments) -> List[int]:
+        token_ids: List[int] = []
+        for segment in segments:
+            token_ids.extend(
+                self._encode_text_piece(
+                    segment.text,
+                    allow_special_tokens=segment.allow_special,
+                )
+            )
+        return token_ids
+
+    @staticmethod
+    def _truncate(
+        ids: List[int], truncation: bool = False, max_length: Optional[int] = None
+    ) -> List[int]:
+        if truncation and max_length is not None:
+            return ids[:max_length]
+        return ids
+
+    def _format_chat_token_output(
+        self,
+        encoded_inputs: List[List[int]],
+        *,
+        is_batched: bool,
+        padding=False,
+        truncation: bool = False,
+        max_length: Optional[int] = None,
+        return_tensors=None,
+        return_dict: bool = False,
+    ):
+        encoded_inputs = [
+            self._truncate(ids, truncation=truncation, max_length=max_length)
+            for ids in encoded_inputs
+        ]
+
+        needs_batch_encoding = (
+            is_batched or padding or return_tensors is not None or return_dict
+        )
+        if not needs_batch_encoding:
+            return encoded_inputs[0]
+
+        features = [
+            {"input_ids": ids, "attention_mask": [1] * len(ids)}
+            for ids in encoded_inputs
+        ]
+        batch = self.pad(
+            features,
+            padding=padding,
+            max_length=max_length if padding else None,
+            return_attention_mask=True,
+            return_tensors=return_tensors,
+        )
+
+        if return_dict:
+            return batch
+        if is_batched:
+            return batch["input_ids"]
+        return batch["input_ids"][0] if return_tensors is None else batch["input_ids"]
+
+    """ ----- Below are the abstract methods required by PreTrainedTokenizer ----- """
+
+    @property
+    def vocab_size(self) -> int:
+        return self.n_words
+
+    def get_vocab(self) -> Dict[str, int]:
+        return self.encoder
+
+    def _tokenize(self, text: str, **kwargs) -> List[str]:
+        return [self.decoder[t] for t in self.encode(text)]
+
+    def _convert_token_to_id(self, token: str) -> int:
+        return self.encoder.get(token, self.unk_id)
+
+    def _convert_id_to_token(self, index: int) -> str:
+        return self.decoder.get(index)
+
+    @staticmethod
+    def clean_up_tokenization(out_string: str) -> str:
+        return out_string
+
+    def convert_tokens_to_string(self, tokens: List[str]) -> str:
+        text = "".join(tokens)
+        text = bytearray([self.byte_decoder[c] for c in text]).decode(
+            "utf-8", "replace"
+        )
+        return text
+
+    def save_vocabulary(
+        self, save_directory: str, filename_prefix: Optional[str] = None
+    ) -> Tuple[str]:
+        if not os.path.isdir(save_directory):
+            raise ValueError(
+                f"vocabulary path ({save_directory}) should be a directory"
+            )
+        out_vocab_file = os.path.join(
+            save_directory,
+            (filename_prefix + "-" if filename_prefix else "")
+            + VOCAB_FILES_NAMES["vocab_file"],
+        )
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(
+            out_vocab_file
+        ) and os.path.isfile(self.vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+
+        return (out_vocab_file,)
+
+    def apply_chat_template(
+        self,
+        conversation,
+        tools: Optional[list[dict]] = None,
+        tokenize: bool = False,
+        add_generation_prompt: bool = True,
+        thinking: bool = True,
+        padding=False,
+        truncation: bool = False,
+        max_length: Optional[int] = None,
+        return_tensors=None,
+        return_dict: bool = False,
+        **kwargs,
+    ):
+        # Tokenizer-level rendering reorders tool result messages to match
+        # assistant tool_calls, normalizes per-call arguments and response
+        # schema, then encodes the resulting XTML structure segment-by-segment.
+        is_batched = is_batched_conversation(conversation)
+        conversations = conversation if is_batched else [conversation]
+        image_prompts = kwargs.pop("image_prompts", None)
+        if is_batched and image_prompts is not None:
+            raise ValueError("image_prompts is only supported for one chat.")
+
+        # by default set thinking effort to max
+        kwargs.setdefault("thinking_effort", "max")
+
+        segment_batches = [
+            build_chat_segments(
+                messages,
+                tools=tools,
+                add_generation_prompt=add_generation_prompt,
+                thinking=thinking,
+                image_prompts=image_prompts,
+                **kwargs,
+            )
+            for messages in conversations
+        ]
+
+        if not tokenize:
+            rendered = [
+                "".join(segment.text for segment in segments)
+                for segments in segment_batches
+            ]
+            return rendered if is_batched else rendered[0]
+
+        encoded_inputs = [
+            self._encode_chat_segments(segments) for segments in segment_batches
+        ]
+        return self._format_chat_token_output(
+            encoded_inputs,
+            is_batched=is_batched,
+            padding=padding,
+            truncation=truncation,
+            max_length=max_length,
+            return_tensors=return_tensors,
+            return_dict=return_dict,
+        )

tokenizer_config.json

@@ -0,0 +1,224 @@
+{
+  "added_tokens_decoder": {
+    "163584": {
+      "content": "[BOS]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163585": {
+      "content": "[EOS]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163586": {
+      "content": "<|end_of_msg|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163587": {
+      "content": "<|open|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "163588": {
+      "content": "<|close|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "163589": {
+      "content": "<|sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "163590": {
+      "content": "[start_header_id]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163591": {
+      "content": "[end_header_id]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163593": {
+      "content": "[EOT]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163602": {
+      "content": "<|media_begin|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163603": {
+      "content": "<|media_content|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163604": {
+      "content": "<|media_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163605": {
+      "content": "<|media_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163649": {
+      "content": "<osagent_mode>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163838": {
+      "content": "[UNK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163839": {
+      "content": "[PAD]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163840": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163841": {
+      "content": "<|im_user|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163842": {
+      "content": "<|im_assistant|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163843": {
+      "content": "<|start_header_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163844": {
+      "content": "<|end_header_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163845": {
+      "content": "<|im_system|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163846": {
+      "content": "<|im_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_end|>",
+    "<|im_user|>",
+    "<|im_assistant|>",
+    "<|start_header_id|>",
+    "<|end_header_id|>",
+    "[EOT]",
+    "<|im_system|>",
+    "<|im_middle|>"
+  ],
+  "auto_map": {
+    "AutoTokenizer": [
+      "tokenization_kimi.TikTokenTokenizer",
+      null
+    ]
+  },
+  "backend": "custom",
+  "bos_token": "[BOS]",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "[EOS]",
+  "extra_special_tokens": [
+    "<|im_end|>",
+    "<|im_user|>",
+    "<|im_assistant|>",
+    "<|start_header_id|>",
+    "<|end_header_id|>",
+    "[EOT]",
+    "<|im_system|>",
+    "<|im_middle|>"
+  ],
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "[PAD]",
+  "tokenizer_class": "TikTokenTokenizer",
+  "unk_token": "[UNK]"
+}