Upload folder using huggingface_hub

chat_template.jinja

@@ -0,0 +1,48 @@
+{% macro render_content(msg) -%}
+    {%- set c = msg.get('content') -%}
+    {%- if c is string -%}
+      {{ c }}
+    {%- elif c is not none -%}
+      {% for content in c -%}
+        {% if content['type'] == 'image' or 'image' in content or 'image_url' in content -%}
+          <|media_start|>image<|media_content|><|media_pad|><|media_end|>
+        {% else -%}
+          {{ content['text'] }}
+        {%- endif -%}
+      {%- endfor -%}
+    {%- endif -%}
+{%- endmacro %}
+
+
+{%- if tools -%}
+  <|im_system|>tool_declare<|im_middle|>{{ tools | tojson(separators=(',', ':')) }}<|im_end|>
+{%- endif -%}
+{% for message in messages %}  
+  {%- set role_name =  message.get('name') or  message['role'] -%}
+  {%- if message['role'] == 'user' -%}
+    <|im_user|>{{role_name}}<|im_middle|>
+  {%- elif message['role'] == 'assistant' -%}
+    <|im_assistant|>{{role_name}}<|im_middle|>
+  {%- else -%}
+    <|im_system|>{{role_name}}<|im_middle|>
+  {%- endif -%}
+
+  {%- if message['role'] == 'assistant' and message.get('tool_calls') -%}
+    {{render_content(message)}}<|tool_calls_section_begin|>
+    {%- for tool_call in message['tool_calls'] -%}
+        {%- set formatted_id = tool_call['id'] -%}
+      <|tool_call_begin|>{{ formatted_id }}<|tool_call_argument_begin|>{% if tool_call['function']['arguments'] is string %}{{ tool_call['function']['arguments'] }}{% else %}{{ tool_call['function']['arguments'] | tojson }}{% endif %}<|tool_call_end|>
+    {%- endfor -%}
+    <|tool_calls_section_end|>
+  {%- elif message['role'] == 'tool' -%}
+    {%- set tool_call_id = message.tool_call_id -%}
+    ## Return of {{ tool_call_id }}
+{{render_content(message)}}
+  {%- elif message['content'] is not none -%}
+    {{render_content(message)}}
+  {%- endif -%}
+  <|im_end|>
+{%- endfor -%}
+{%- if add_generation_prompt -%}
+  <|im_assistant|>assistant<|im_middle|>
+{%- endif -%}

config.json

@@ -0,0 +1,75 @@
+{
+  "architectures": [
+    "KimiLinearForCausalLM"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_kimi.KimiLinearConfig",
+    "AutoModel": "modeling_kimi.KimiLinearModel",
+    "AutoModelForCausalLM": "modeling_kimi.KimiLinearForCausalLM"
+  },
+  "bos_token_id": 163584,
+  "dtype": "bfloat16",
+  "eos_token_id": 200002,
+  "first_k_dense_replace": 1,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 1024,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "kv_lora_rank": 512,
+  "linear_attn_config": {
+    "full_attn_layers": [
+      4,
+      8,
+      12,
+      16
+    ],
+    "head_dim": 128,
+    "kda_layers": [
+      1,
+      2,
+      3,
+      5,
+      6,
+      7,
+      9,
+      10,
+      11,
+      13,
+      14,
+      15
+    ],
+    "num_heads": 32,
+    "short_conv_kernel_size": 4
+  },
+  "mla_use_nope": true,
+  "model_max_length": 1048576,
+  "model_type": "kimi_linear",
+  "moe_intermediate_size": 2048,
+  "moe_layer_freq": 1,
+  "moe_renormalize": true,
+  "moe_router_activation_func": "sigmoid",
+  "num_attention_heads": 16,
+  "num_expert_group": 1,
+  "num_experts": 16,
+  "num_experts_per_token": 2,
+  "num_hidden_layers": 18,
+  "num_key_value_heads": 2,
+  "num_nextn_predict_layers": 0,
+  "num_shared_experts": 1,
+  "pad_token_id": 199999,
+  "q_lora_rank": null,
+  "qk_nope_head_dim": 128,
+  "qk_rope_head_dim": 64,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "rope_theta": 10000.0,
+  "routed_scaling_factor": 2.446,
+  "tie_word_embeddings": false,
+  "topk_group": 1,
+  "transformers_version": "4.57.1",
+  "use_cache": true,
+  "use_grouped_topk": true,
+  "v_head_dim": 128,
+  "vocab_size": 201088
+}

configuration_kimi.py

@@ -0,0 +1,140 @@
+# coding=utf-8
+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,
+        num_nextn_predict_layers: int = 0,
+        linear_attn_config: Optional[dict] = None,
+        **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
+        # 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
+
+        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"]
+        )

generation_config.json

@@ -0,0 +1,7 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 163584,
+  "eos_token_id": 200002,
+  "pad_token_id": 199999,
+  "transformers_version": "4.57.1"
+}

model-00001-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2b46dce368ffa84238ec6704b4aa4af05d919f4f999b7df347d9e1fedf4ddfc0
+size 996133416

model-00002-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bb8d14d02a6a03b684d1b21faa157a2c381a76ea4dc378999863cf840a53c93d
+size 997527352

model-00003-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c088e0d0ff87e21d36a78b5e4f2c263cd69a3142cf9c978607f0d9cbe48afe31
+size 997527440

model-00004-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d97415e8d3a757bb58f92e3a8de8b1088c5b9a429373e176741b22f9f20b01f5
+size 997527624

model-00005-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2c505e76bf66a850d78c70cbe007fbf2def8ff8a8dc57b5afe3f9624e1fd2ffc
+size 610668816

model-00006-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f2c821308069fd4f4b7b9a1df4604fcd22cc52f0ce3460e828d2127f3da34c7
+size 411828352

modeling_kimi.py

@@ -0,0 +1,1099 @@
+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, repeat
+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, auto_docstring, can_return_tuple, logging
+from transformers.utils.generic import OutputRecorder, check_model_inputs
+
+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
+    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`")
+
+from .configuration_kimi import KimiLinearConfig
+
+assert version.parse(transformers.__version__) >= version.parse("4.56.0"), \
+    "Please upgrade transformers to >= 4.56.0"
+
+logger = logging.get_logger(__name__)
+
+
+def index_first_axis(x, indices):
+    other_shape = x.shape[1:]
+    second_dim = other_shape.numel()
+    return torch.gather(
+        rearrange(x, "b ... -> b (...)"), 0, repeat(indices, "z -> z d", d=second_dim),
+    ).reshape(-1, *other_shape)
+
+
+def index_put_first_axis(x, indices, first_axis_dim):
+    y = torch.zeros(first_axis_dim, *x.shape[1:], device=x.device, dtype=x.dtype)
+    # TODO [2022-03-04] For some reason torch.scatter is a bit faster than indexing.
+    y[indices] = x
+    # y.scatter_(0, repeat(indices, 'z -> z d', d=x.shape[1]), x)
+    return y
+
+
+@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 unpad_input(
+    q: torch.Tensor,
+    states: tuple[torch.Tensor],
+    attention_mask: torch.Tensor,
+    q_len: int,
+    keepdim: bool = False,
+):
+    indices_k, cu_seqlens_k, max_seqlen_in_batch_k = get_unpad_data(attention_mask)
+    batch_size, seq_len, *_ = states[0].shape
+
+    state = tuple(
+        index_first_axis(rearrange(s, "b s ... -> (b s) ..."), indices_k)
+        for s in states
+    )
+
+    if q_len == seq_len:
+        q = index_first_axis(rearrange(q, "b s ... -> (b s) ..."), indices_k)
+        cu_seqlens_q = cu_seqlens_k
+        max_seqlen_in_batch_q = max_seqlen_in_batch_k
+        indices_q = indices_k
+    elif q_len == 1:
+        max_seqlen_in_batch_q = 1
+        cu_seqlens_q = torch.arange(batch_size + 1, dtype=torch.int32, device=q.device)
+        indices_q = cu_seqlens_q[:-1]
+        q = q.squeeze(1)
+    else:
+        raise NotImplementedError("We only support either q_len == k_len (prefilling) or q_len == 1 (decoding)")
+
+    if keepdim:
+        q = q.unsqueeze(0)
+        state = tuple(s.unsqueeze(0) for s in state)
+
+    return (
+        q,
+        state,
+        indices_q,
+        (cu_seqlens_q, cu_seqlens_k),
+        (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
+    )
+
+
+def pad_input(
+    hidden_states: torch.Tensor,
+    indices: torch.LongTensor,
+    batch_size: int,
+    seq_len: int,
+) -> torch.Tensor:
+    output = index_put_first_axis(hidden_states, indices, batch_size * seq_len)
+    return rearrange(output, "(b s) ... -> b s ...", b=batch_size)
+
+
+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_mask_sizes(self, cache_position: torch.Tensor, 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
+        query_length = 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):
+        """
+        KimiRMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * \
+            torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_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
+
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, hidden_states):
+        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)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        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:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(
+        batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+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_states = repeat_kv(key, module.num_key_value_groups)
+    value_states = repeat_kv(value, module.num_key_value_groups)
+
+    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
+    if attention_mask is not None:
+        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        attn_weights = attn_weights + causal_mask
+
+    attn_weights = nn.functional.softmax(
+        attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
+    attn_weights = nn.functional.dropout(
+        attn_weights, p=dropout, training=module.training)
+    attn_output = torch.matmul(attn_weights, value_states)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+
+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.rope_theta = config.rope_theta
+        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}")
+
+        assert self.q_lora_rank is None
+        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
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor | 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)
+
+        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()
+        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 suppoerted 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)).view(1, 1, -1, 1))
+
+        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.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 q_len <= 64 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, conv_state_q = self.q_conv1d(
+            x=self.q_proj(hidden_states),
+            cache=conv_state_q,
+            output_final_state=use_cache,
+            cu_seqlens=cu_seqlens,
+        )
+        k, conv_state_k = self.k_conv1d(
+            x=self.k_proj(hidden_states),
+            cache=conv_state_k,
+            output_final_state=use_cache,
+            cu_seqlens=cu_seqlens,
+        )
+        v, conv_state_v = self.v_conv1d(
+            x=self.v_proj(hidden_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 = fused_kda_gate(g, self.A_log, self.head_dim, g_bias=self.dt_bias)
+        beta = self.b_proj(hidden_states).float().sigmoid()
+
+        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,
+                initial_state=recurrent_state,
+                output_final_state=True,
+                use_qk_l2norm_in_kernel=True,
+                cu_seqlens=cu_seqlens,
+            )
+        else:
+            o, recurrent_state = fused_recurrent_kda(
+                q=q,
+                k=k,
+                v=v,
+                g=g,
+                beta=beta,
+                initial_state=recurrent_state,
+                output_final_state=True,
+                use_qk_l2norm_in_kernel=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)
+
+        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
+        assert not self.training
+        scores_for_choice = scores.view(bsz * seq_len, -1)
+        scores_for_choice += self.e_score_correction_bias.unsqueeze(0)
+        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(), 0.0)  # [n, e]
+        _, 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.ep_size = 1
+        self.experts_per_rank = config.num_experts
+        self.ep_rank = 0
+        self.experts = nn.ModuleList(
+            [
+                KimiBlockSparseMLP(
+                    config, 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,
+            )
+
+    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 not self.training:
+            y = self.moe_infer(hidden_states, topk_idx, topk_weight).view(*orig_shape)
+        else:
+            raise NotImplementedError("Training mode is not supported in KimiSparseMoeBlock")
+        if self.config.num_shared_experts is not None:
+            y = y + self.shared_experts(identity)
+        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
+        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)
+
+    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,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> tuple[torch.FloatTensor, tuple[torch.FloatTensor, torch.FloatTensor] | None]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+        """
+
+        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
+
+
+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):
+        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_()
+
+
+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)
+
+        if getattr(config, "_attn_implementation", None) is not None:
+            if config._attn_implementation != "flash_attention_2":
+                logger.warning_once(
+                    f"Ignoring the provided attention implementation {config._attn_implementation}")
+                logger.warning_once("Using flash_attention_2 backend instead.")
+                config._attn_implementation = "flash_attention_2"
+        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
+    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,
+            input_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            cache_position=cache_position,
+            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)
+
+        for decoder_layer in self.layers:
+            layer_mask = linear_attn_mask if decoder_layer.is_linear_attn else causal_mask
+
+            hidden_states = decoder_layer(
+                hidden_states,
+                attention_mask=layer_mask,
+                past_key_values=past_key_values,
+                cache_position=cache_position,
+                **kwargs,
+            )
+
+        hidden_states = self.norm(hidden_states)
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+        )
+
+
+class KimiLinearForCausalLM(KimiPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.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
+    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]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, KimiLinearForCausalLM
+
+        >>> model = KimiLinearForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
+        >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```"""
+
+        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,
+        )

tokenization_kimi_py

@@ -0,0 +1,347 @@
+import os
+import tiktoken
+
+from logging import getLogger
+from pathlib import Path
+from typing import (
+    cast,
+    Tuple,
+    Dict,
+    Iterator,
+    List,
+    Union,
+    Optional,
+)
+from shutil import copyfile
+from tiktoken.load import load_tiktoken_bpe
+from tokenizers import AddedToken, pre_tokenizers, Regex
+from transformers.tokenization_utils import PreTrainedTokenizer
+from transformers.models.gpt2.tokenization_gpt2 import bytes_to_unicode
+from typing import Any
+
+
+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|>",
+            ] 
+        
+        special_tokens_mapping = {
+        i: added_tokens_decoder[i].content for i in added_tokens_decoder
+    }
+
+        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 + 2
+            )
+        }
+
+
+
+        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,
+            **kwargs,
+        )
+        self.all_special_ids_set = set(self.all_special_ids)
+
+    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
+
+        # 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
+
+        texts = self.pre_tokenizer_process(text)
+
+        all_substrs = []
+        for text in texts:
+            substrs = (
+                substr
+                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
+                )
+            )
+            all_substrs.extend(substrs)
+
+        t: List[int] = []
+        for substr in all_substrs:
+            if allow_special_tokens:
+                t.extend(
+                    # we should consider special token as a common token
+                    self.model.encode(
+                        substr,
+                        allowed_special="all",
+                    )
+                )
+            else:
+                t.extend(
+                    # we should consider special token as a common token
+                    self.model.encode(
+                        substr,
+                        disallowed_special=(),
+                    )
+                )
+
+        return t
+
+    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 pre_tokenizer_process(self, text: str) -> List[str]:
+        """
+        pre-tokenizes the input text into a list of tokens.
+        This method is used to split the input text into smaller chunks for internal processing.
+        """
+        return [text]
+
+
+    """ ----- 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, 
+        **kwargs
+    ):
+        tools = deep_sort_dict(tools)
+        return super().apply_chat_template(conversation, 
+                                           tools=tools,
+                                           tokenize=tokenize,
+                                           add_generation_prompt=add_generation_prompt, 
+                                           **kwargs)
+    
+
+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

tokenizer_config.json

@@ -0,0 +1,164 @@
+{
+  "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": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163587": {
+      "content": "<|im_user|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163588": {
+      "content": "<|im_assistant|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "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
+    },
+    "163594": {
+      "content": "<|im_system|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163595": {
+      "content": "<|tool_calls_section_begin|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "163596": {
+      "content": "<|tool_calls_section_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "163597": {
+      "content": "<|tool_call_begin|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "163598": {
+      "content": "<|tool_call_argument_begin|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "163599": {
+      "content": "<|tool_call_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "163601": {
+      "content": "<|im_middle|>",
+      "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
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_end|>",
+    "<|im_user|>",
+    "<|im_assistant|>",
+    "<|start_header_id|>",
+    "<|end_header_id|>",
+    "[EOT]",
+    "<|im_system|>",
+    "<|im_middle|>"
+  ],
+  "bos_token": "[BOS]",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "[EOS]",
+  "extra_special_tokens": {},
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "[PAD]",
+  "tokenizer_class": "TikTokenTokenizer",
+  "unk_token": "[UNK]",
+  "auto_map": {
+    "AutoTokenizer": [
+      "tokenization_kimi.TikTokenTokenizer",
+      null
+    ]
+  }
+}