Upload Ovis2.6-80B-A3B model files

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

added_tokens.json

@@ -0,0 +1,28 @@
+{
+  "</think>": 151668,
+  "</tool_call>": 151658,
+  "</tool_response>": 151666,
+  "<think>": 151667,
+  "<tool_call>": 151657,
+  "<tool_response>": 151665,
+  "<|box_end|>": 151649,
+  "<|box_start|>": 151648,
+  "<|endoftext|>": 151643,
+  "<|file_sep|>": 151664,
+  "<|fim_middle|>": 151660,
+  "<|fim_pad|>": 151662,
+  "<|fim_prefix|>": 151659,
+  "<|fim_suffix|>": 151661,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|image_pad|>": 151655,
+  "<|object_ref_end|>": 151647,
+  "<|object_ref_start|>": 151646,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|repo_name|>": 151663,
+  "<|video_pad|>": 151656,
+  "<|vision_end|>": 151653,
+  "<|vision_pad|>": 151654,
+  "<|vision_start|>": 151652
+}

chat_template.json

@@ -0,0 +1,3 @@
+{
+  "chat_template": "{%- for message in messages %}\n{{- '<|im_start|>' + message.role + '\\n' }}\n{%- if message.role == 'system' or message.role == 'user' %}\n{%- if message.content is string %}\n{{- message.content }}\n{%- else %}\n{%- for item in message.content %}\n{%- if item.type == 'text' and 'text' in item %}\n{{- item.text }}\n{%- elif item.type == 'image' %}\n{{- '<image>' }}\n{%- elif item.type == 'video' %}\n{{- '<video>' }}\n{%- else %}\n{{- raise_exception('Invalid content type. Supported types for system and user are text, image, video.') }}\n{%- endif %}\n{%- if not loop.last %}{{- '\\n' }}{%- endif %}\n{%- endfor %}\n{%- endif %}\n{%- elif message.role == 'assistant' %}\n{%- set content = '' %}\n{%- if message.content is string %}\n{%- set content = message.content %}\n{%- else %}\n{%- set ns = namespace(content='') -%}\n{%- for item in message.content %}\n{%- if item.type == 'text' and 'text' in item %}\n{%- set ns.content = ns.content ~ item.text %}\n{%- else %}\n{{- raise_exception('Invalid content type. Supported type for assistant is text.') }}\n{%- endif %}\n{%- endfor %}\n{%- set content = ns.content -%}\n{%- endif %}\n{%- set content = content.split('</think>')[-1].lstrip('\\n') %}\n{{- content }}\n{%- else %}\n{{- raise_exception('Invalid role. Supported roles are system, user, assistant.') }}\n{%- endif %}\n{{- '<|im_end|>\\n' }}\n{%- endfor %}\n{%- if add_generation_prompt %}\n{{- '<|im_start|>assistant\\n' }}\n{%- if enable_thinking is defined and enable_thinking is false %}\n{{- '<think>\\n\\n</think>\\n\\n' }}\n{%- endif %}\n{%- endif %}\n"
+}

config.json

@@ -0,0 +1,85 @@
+{
+  "architectures": [
+    "Ovis2_6_NextForCausalLM"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_ovis2_6.Ovis2_6_Next_Config",
+    "AutoModelForCausalLM": "modeling_ovis2_6.Ovis2_6_NextForCausalLM"
+  },
+  "conversation_formatter_class": "Qwen3ConversationFormatter",
+  "hidden_size": 2048,
+  "vocab_size": 151936,
+  "max_position_embeddings": 32768,
+  "llm_config": {
+    "_name_or_path": "Qwen/Qwen3-Next-80B-A3B-Thinking",
+    "architectures": [
+      "Qwen3NextForCausalLM"
+    ],
+    "attention_dropout": 0.0,
+    "bos_token_id": 151643,
+    "decoder_sparse_step": 1,
+    "eos_token_id": 151645,
+    "full_attention_interval": 4,
+    "head_dim": 256,
+    "hidden_act": "silu",
+    "hidden_size": 2048,
+    "initializer_range": 0.02,
+    "intermediate_size": 5120,
+    "linear_conv_kernel_dim": 4,
+    "linear_key_head_dim": 128,
+    "linear_num_key_heads": 16,
+    "linear_num_value_heads": 32,
+    "linear_value_head_dim": 128,
+    "max_position_embeddings": 32768,
+    "mlp_only_layers": [],
+    "model_type": "qwen3_next",
+    "moe_intermediate_size": 512,
+    "norm_topk_prob": true,
+    "num_attention_heads": 16,
+    "num_experts": 512,
+    "num_experts_per_tok": 10,
+    "num_hidden_layers": 48,
+    "num_key_value_heads": 2,
+    "output_router_logits": false,
+    "partial_rotary_factor": 0.25,
+    "rms_norm_eps": 1e-06,
+    "rope_scaling": null,
+    "rope_theta": 10000000,
+    "router_aux_loss_coef": 0.001,
+    "shared_expert_intermediate_size": 512,
+    "tie_word_embeddings": false,
+    "torch_dtype": "bfloat16",
+    "transformers_version": "4.57.0",
+    "use_cache": false,
+    "use_sliding_window": false,
+    "vocab_size": 151936
+  },
+  "model_type": "ovis2_6_next",
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.57.0",
+  "use_cache": false,
+  "visual_vocab_size": 65536,
+  "vit_config": {
+    "_attn_implementation_autoset": true,
+    "_name_or_path": "google/siglip2-so400m-patch16-512",
+    "attention_dropout": 0.0,
+    "fullatt_block_indexes": null,
+    "hidden_act": "gelu_pytorch_tanh",
+    "hidden_size": 1152,
+    "hidden_stride": 2,
+    "image_size": 512,
+    "intermediate_size": 4304,
+    "layer_norm_eps": 1e-06,
+    "model_type": "siglip2_navit",
+    "num_attention_heads": 16,
+    "num_channels": 3,
+    "num_hidden_layers": 27,
+    "num_patches": -1,
+    "patch_size": 16,
+    "preserve_original_pe": true,
+    "temporal_patch_size": 1,
+    "torch_dtype": "bfloat16",
+    "use_rope": true,
+    "window_size": 112
+  }
+}

configuration_ovis2_6.py

@@ -0,0 +1,147 @@
+from typing import Any, Optional, List, Union
+
+from transformers import Qwen3Config, Qwen3MoeConfig, Qwen3NextConfig
+from transformers.configuration_utils import PretrainedConfig
+
+__all__ = ["Siglip2NavitConfig", "Ovis2_6_Config"]
+
+
+class Siglip2NavitConfig(PretrainedConfig):
+    """This is the configuration class to store the configuration of an [`AIMv2Model`].
+
+    Instantiating a configuration with the defaults will yield a similar configuration
+    to that of the [apple/aimv2-large-patch14-224](https://huggingface.co/apple/aimv2-large-patch14-224).
+
+    Args:
+        hidden_size: Dimension of the hidden representations.
+        intermediate_size: Dimension of the SwiGLU representations.
+        num_hidden_layers: Number of hidden layers in the Transformer.
+        num_attention_heads: Number of attention heads for each attention layer
+            in the Transformer.
+        num_channels: Number of input channels.
+        image_size: Image size.
+        patch_size: Patch size.
+        rms_norm_eps: Epsilon value used for the RMS normalization layer.
+        attention_dropout: Dropout ratio for attention probabilities.
+        projection_dropout: Dropout ratio for the projection layer after the attention.
+        qkv_bias: Whether to add a bias to the queries, keys and values.
+        use_bias: Whether to add a bias in the feed-forward and projection layers.
+        kwargs: Keyword arguments for the [`PretrainedConfig`].
+    """
+
+    model_type: str = "siglip2_navit"
+
+    def __init__(
+        self,
+        hidden_size: int = 1024,
+        intermediate_size: int = 4096,
+        num_hidden_layers: int = 24,
+        num_attention_heads: int = 16,
+        num_channels: int = 3,
+        num_patches: int = -1,
+        image_size: int = 512,
+        patch_size: int = 16,
+        hidden_act: str="gelu_pytorch_tanh",
+        layer_norm_eps: float = 1e-6,
+        attention_dropout: float = 0.0,
+        hidden_stride: int = 2,
+        window_size: int = 112,
+        fullatt_block_indexes: Optional[list] = None,
+        temporal_patch_size: int = 1,
+        preserve_original_pe: bool = True,
+        use_rope: bool = True,
+        **kwargs: Any,
+    ):
+        super().__init__(**kwargs)
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_channels = num_channels
+        self.num_patches = num_patches
+        self.patch_size = patch_size
+        self.image_size = image_size
+        self.hidden_act = hidden_act
+        self.attention_dropout = attention_dropout
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_stride = hidden_stride
+        self.window_size = window_size
+        self.fullatt_block_indexes = fullatt_block_indexes
+        self.temporal_patch_size = temporal_patch_size
+        self.preserve_original_pe = preserve_original_pe
+        self.use_rope = use_rope
+
+
+class Ovis2_6_Config(PretrainedConfig):
+    model_type = "ovis2_6"
+    sub_configs = dict(llm_config=Qwen3Config, vit_config=Siglip2NavitConfig)
+
+    def __init__(self,
+        llm_config: Optional[Union[Qwen3Config, dict]] = None,
+        vit_config: Optional[Union[Siglip2NavitConfig, dict]] = None,
+        visual_vocab_size=65536,
+        hidden_size=None,
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+        if isinstance(llm_config, dict):
+            llm_config = Qwen3Config(**llm_config)
+        self.llm_config = llm_config
+        if isinstance(vit_config, dict):
+            vit_config = Siglip2NavitConfig(**vit_config)
+        self.vit_config = vit_config
+        self.visual_vocab_size = visual_vocab_size
+        self.hidden_size = hidden_size
+        if kwargs.get('attn_implementation'):
+            self.llm_config._attn_implementation = kwargs['attn_implementation']
+            self.vit_config._attn_implementation = kwargs['attn_implementation']
+
+
+class Ovis2_6_Moe_Config(PretrainedConfig):
+    model_type = "ovis2_6_moe"
+    sub_configs = dict(llm_config=Qwen3MoeConfig, vit_config=Siglip2NavitConfig)
+
+    def __init__(self,
+        llm_config: Optional[Union[Qwen3MoeConfig, dict]] = None,
+        vit_config: Optional[Union[Siglip2NavitConfig, dict]] = None,
+        visual_vocab_size=65536,
+        hidden_size=None,
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+        if isinstance(llm_config, dict):
+            llm_config = Qwen3MoeConfig(**llm_config)
+        self.llm_config = llm_config
+        if isinstance(vit_config, dict):
+            vit_config = Siglip2NavitConfig(**vit_config)
+        self.vit_config = vit_config
+        self.visual_vocab_size = visual_vocab_size
+        self.hidden_size = hidden_size
+        if kwargs.get('attn_implementation'):
+            self.llm_config._attn_implementation = kwargs['attn_implementation']
+            self.vit_config._attn_implementation = kwargs['attn_implementation']
+
+
+class Ovis2_6_Next_Config(PretrainedConfig):
+    model_type = "ovis2_6_next"
+    sub_configs = dict(llm_config=Qwen3NextConfig, vit_config=Siglip2NavitConfig)
+
+    def __init__(self,
+        llm_config: Optional[Union[Qwen3NextConfig, dict]] = None,
+        vit_config: Optional[Union[Siglip2NavitConfig, dict]] = None,
+        visual_vocab_size=65536,
+        hidden_size=None,
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+        if isinstance(llm_config, dict):
+            llm_config = Qwen3NextConfig(**llm_config)
+        self.llm_config = llm_config
+        if isinstance(vit_config, dict):
+            vit_config = Siglip2NavitConfig(**vit_config)
+        self.vit_config = vit_config
+        self.visual_vocab_size = visual_vocab_size
+        self.hidden_size = hidden_size
+        if kwargs.get('attn_implementation'):
+            self.llm_config._attn_implementation = kwargs['attn_implementation']
+            self.vit_config._attn_implementation = kwargs['attn_implementation']

generation_config.json

@@ -0,0 +1,15 @@
+{
+  "bos_token_id": 151643,
+  "do_sample": true,
+  "eos_token_id": [
+    151645,
+    151643
+  ],
+  "multimodal_max_length": 8192,
+  "pad_token_id": 151643,
+  "repetition_penalty": 1.05,
+  "temperature": 0.6,
+  "top_k": 20,
+  "top_p": 0.95,
+  "transformers_version": "4.57.0"
+}

model-00001-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d9792f5e6046074a725ee88f771a68a7358afe439679a8c32af0e113690b64a8
+size 5000015296

model-00002-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:538ba169d02a4c2bbcbb6e513a773e10a74b67e7fae7c6c94a8ff101aff1df57
+size 5000252384

model-00003-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aab609de4fe9cc063b153735c30af9b8272b27e08cdaf7323be5e1577c2ec4d1
+size 5000263088

model-00004-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8e0ee2f96a1727c5ad5ae8c87868495ee58ceb00b071433d6799a4da6048175c
+size 5000252384

model-00005-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6eb1ecafdfbb8ba943d7f2ca74657216e5a6020a525283d9cda8403a3b997d09
+size 5000263144

model-00006-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48b8c9d9a29cc21aa978df73b253eca23a74a62742316d5125e6561e1e81e671
+size 5000252384

model-00007-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:773bbb0b4173d3a873af3b790c0afe6d74a7b5c79c7573285a2286688e679660
+size 4998492960

model-00008-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4edd37fa232a6a9fea32d840d1428694de8c42f68fff0af052e5405a081144e2
+size 4999928360

model-00009-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e975fe50565c5a334adfc16f44a990e20337b247d7ae3669f542903b54dfbf07
+size 4998494592

model-00010-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b2bc65f0092233077b30a85607824ff1a7f70dcdce83de77bdf4ef0b58e624e8
+size 5000254744

model-00011-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8fd4605f2730180d2f67a9a558c893c57eb8246c57c4e6f5945b186a7ce5d1d9
+size 5000265432

model-00012-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:39d230367bf1f6ec7b17423b3fb111775aa5f8de1c58faadd1b02f8cc7ba7e08
+size 4998494872

model-00013-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6fcc463aad26b77b63efdb4aafca9b616b1ffdcd44b3987bc0dfddb37c0b21d9
+size 4999928032

model-00014-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:db5225c9f89ddba716dd6d431ec8886a4d218d696e1a3d02c6359a72d98c12c7
+size 4998494792

model-00015-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e394e05979d338ae9ad1d7c5d881f0dc7117e520ed7306876484ee6e2eaf9ef8
+size 5000254504

model-00016-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:19eb7d7df3beefb6e3f6f27e7f76e361ff188856682e878d6e1a69eb2e639f97
+size 5000265624

model-00017-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:87b0a5d859bcd8e2070e4d78c39146a804668198f8784d8e179f649c425fc961
+size 5000254568

model-00018-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4dbabdd23b3994b62a13ef130e382158bbe4377e0efc30fcbafa964e35d8a606
+size 5000265568

model-00019-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:56a85fb1f5a77f90f531a707daa75d011d847fa53e0bc01c598bcb51c5ce0816
+size 5000254624

model-00020-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bffeecd813a34e8dff3487a9c4e8c446efeece698f1b5fb2a02184b2260c46f7
+size 4998494616

model-00021-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7360d9f39b23e86a4ad22916ad053a6772a9b0d1ecb132717862288e40e1d245
+size 4999928288

model-00022-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b3b95bb4e89b839e63c4b59dbabebf57f931dde6ec5049811abd13d68574edff
+size 4998494560

model-00023-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:93f6b3e87e0e87335843e8665f8294205003cd3e0085ad92aac9e10468f8ff52
+size 5000254736

model-00024-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7938e466f56bc2b30442ee572b8030ca7e8449b9652ad2fb8983f99f6368e9aa
+size 5000265432

model-00025-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0cf72944222eb6cbcc97503f91537dffe4086b28371bff1287ee942eb04f04ac
+size 5000254744

model-00026-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fafbb99e8a4234bd40f3c89ed17e79104ec672bbabc04b1e4eeb5afd7917d441
+size 5000265488

model-00027-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c3fc67d34a7fb2ef2298e590f00371e1e9348fd2e07b821fc25004c6d76f1a6
+size 5000254744

model-00028-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0299385cf2e2d69816418d7611bc6cfeb38d77000a0bad6a21e98d5c364bd722
+size 4998494592

model-00029-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:51b88ac04a9564cbf65c97f6713dc5631e7818e778f91ee6767f96e4c12b536c
+size 4999928360

model-00030-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5448dfe4410bfd4d9d519ab3c4a8c18cb42771a9e0fd0b6eb7ded16f13fb42bc
+size 4998494592

model-00031-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:706dea8e613b2feb4bf45b290cf5b21c27b83c6d72356108d276d86478e60a30
+size 5000254744

model-00032-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c2c217b90cf4ae3c6be2e1ea5d5d1cd61626c30a741aefa067a5f3c695fb03e6
+size 4990399824

model-00033-of-00033.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:12bf823c96219aa44568257b24fcadf6079a32370267fa849248f695cc5d754e
+size 1075378024

modeling_ovis2_6.py

@@ -0,0 +1,1458 @@
+import math
+from typing import Dict, List, Optional, Tuple, Union
+
+import PIL.Image
+import numpy as np
+import torch
+from torch import Tensor, nn
+from torch.nn import functional as F
+from transformers import (
+    AutoConfig,
+    AutoImageProcessor,
+    AutoModel,
+    AutoModelForCausalLM,
+    AutoTokenizer,
+)
+from transformers.activations import ACT2FN
+from transformers.generation.utils import GenerateOutput
+from transformers.modeling_outputs import BaseModelOutputWithNoAttention
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import is_flash_attn_2_available
+
+from .configuration_ovis2_6 import Siglip2NavitConfig, Ovis2_6_Config, Ovis2_6_Moe_Config, Ovis2_6_Next_Config
+
+
+if is_flash_attn_2_available():
+    from flash_attn import flash_attn_varlen_func
+    from flash_attn.layers.rotary import apply_rotary_emb
+
+
+IMAGE_PLACEHOLDER = "<image>"
+IMAGE_PLACEHOLDER_ID = -200
+VIDEO_PLACEHOLDER = "<video>"
+VIDEO_PLACEHOLDER_ID = -201
+
+VISUAL_ATOM_ID = -300
+INDICATOR_IDS = [-301, -302, -303, -304]
+
+
+# copied from qwen2.5-vl
+class VisionRotaryEmbedding(nn.Module):
+    def __init__(self, dim: int, theta: float = 10000.0) -> None:
+        super().__init__()
+        inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+
+    def forward(self, seqlen: int) -> torch.Tensor:
+        seq = torch.arange(seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
+        freqs = torch.outer(seq, self.inv_freq)
+        return freqs
+
+
+class Siglip2VisionEmbeddings(nn.Module):
+    def __init__(self, config: Siglip2NavitConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.patch_size = config.patch_size
+        self.image_size = config.image_size
+        self.num_patches = config.num_patches
+        self.preserve_original_pe = config.preserve_original_pe
+        self.hidden_stride = config.hidden_stride
+
+
+        # siglip2 naflex
+        if self.num_patches > 0:
+            self.patch_embedding = nn.Linear(
+                    in_features=config.num_channels * self.patch_size * self.patch_size,
+                    out_features=self.embed_dim,
+                )
+            if self.preserve_original_pe:
+                self.position_embedding_size = int(self.num_patches**0.5)
+                self.position_embedding = nn.Embedding(self.num_patches, self.embed_dim)
+
+        else:
+            self.patch_embedding = nn.Conv2d(
+                    in_channels=config.num_channels,
+                    out_channels=self.embed_dim,
+                    kernel_size=self.patch_size,
+                    stride=self.patch_size,
+                    padding="valid",
+                )
+            if self.preserve_original_pe:
+                self.num_patches = (self.image_size // self.patch_size) ** 2
+                self.position_embedding_size = self.image_size // self.patch_size
+                self.position_embedding = nn.Embedding(self.num_patches, self.embed_dim)
+        
+    @staticmethod
+    def resize_positional_embeddings(
+        positional_embeddings: torch.Tensor,
+        spatial_shapes: torch.LongTensor,
+        max_length: int,
+    ) -> torch.Tensor:
+        """
+        Resize positional embeddings to image-specific size and pad to a fixed size.
+        Args:
+            positional_embeddings (`torch.Tensor`):
+                Position embeddings of shape (height, width, embed_dim)
+            spatial_shapes (`torch.LongTensor`):
+                Spatial shapes of shape (batch_size, 2) to resize the positional embeddings to
+            max_length (`int`):
+                Maximum length of the positional embeddings to pad resized positional embeddings to
+        Returns:
+            `torch.Tensor`: Embeddings of shape (batch_size, max_length, embed_dim)
+        """
+        batch_size = spatial_shapes.shape[0]
+        embed_dim = positional_embeddings.shape[-1]
+        source_dtype = positional_embeddings.dtype
+
+        resulted_positional_embeddings = torch.empty(
+            (batch_size, max_length, embed_dim),
+            device=positional_embeddings.device,
+            dtype=source_dtype,
+        )
+
+        # (height, width, embed_dim) -> (1, embed_dim, height, width) for interpolation
+        positional_embeddings = positional_embeddings.permute(2, 0, 1).unsqueeze(0)
+
+        # Upcast to float32 on CPU because antialias is not supported for bfloat16/float16 on CPU
+        if positional_embeddings.device.type == "cpu":
+            positional_embeddings = positional_embeddings.to(torch.float32)
+
+        for i in range(batch_size):
+            # (1, dim, height, width) -> (1, dim, target_height, target_width)
+            height, width = spatial_shapes[i]
+            resized_embeddings = F.interpolate(
+                positional_embeddings,
+                size=(height, width),
+                mode="bilinear",
+                align_corners=False,
+                antialias=True,
+            )
+
+            # (1, dim, target_height, target_width) -> (target_height * target_width, dim)
+            resized_embeddings = resized_embeddings.reshape(embed_dim, height * width).transpose(0, 1)
+
+            # Cast to original dtype
+            resized_embeddings = resized_embeddings.to(source_dtype)
+
+            resulted_positional_embeddings[i, : height * width] = resized_embeddings
+            resulted_positional_embeddings[i, height * width :] = resized_embeddings[0]
+
+        return resulted_positional_embeddings
+
+    def forward(self, pixel_values: torch.FloatTensor, 
+                grid_thws: Optional[torch.LongTensor] = None) -> torch.Tensor:
+        """
+        Args:
+            pixel_values (`torch.FloatTensor`):
+                Pixel values of shape (num_patches, num_channels * temporal_patch_size * patch_size * patch_size)
+            grid_thws: (`torch.LongTensor`):
+                grid shape (num_patches, 3)
+        """
+
+        # Apply patch embeddings to already patchified pixel values
+        target_dtype = self.patch_embedding.weight.dtype
+        if isinstance(self.patch_embedding, nn.Linear):
+            patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))
+        elif isinstance(self.patch_embedding, nn.Conv2d):
+            pixel_values = pixel_values.view(-1, self.config.num_channels * self.config.temporal_patch_size, self.patch_size,
+                   self.patch_size)
+            patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))
+            patch_embeds = patch_embeds.reshape(-1, self.embed_dim)
+
+
+        if self.preserve_original_pe:
+            assert grid_thws is not None
+            pos_embed_new = torch.zeros_like(patch_embeds)
+            ori_h = ori_w = self.position_embedding_size
+            positional_embeddings = self.position_embedding.weight.reshape(
+                                self.position_embedding_size, self.position_embedding_size, -1
+                            ).unsqueeze(0).permute(0,3,1,2)
+            # pos_embed = self.pos_embed.reshape(1, ori_h, ori_w, -1).permute(0, 3, 1, 2)
+            cnt = 0
+            for t, h, w in grid_thws:
+                thw = t * h * w
+                pe = F.interpolate(positional_embeddings, size=(h, w), mode='bicubic', align_corners=False)
+                pe = pe.permute(0, 2, 3, 1).reshape(1, h * w, -1)
+                pe = pe[0].repeat(t, 1)
+                pe = pe.reshape(t, h // self.hidden_stride, self.hidden_stride, w // self.hidden_stride,
+                                self.hidden_stride, -1)
+                pe = pe.permute(0, 1, 3, 2, 4, 5).reshape(thw, -1)
+                pos_embed_new[cnt:cnt + thw] = pe
+                cnt += thw
+            patch_embeds = patch_embeds + pos_embed_new
+
+        return patch_embeds
+
+
+# copied from qwen2.5-vl
+def apply_rotary_pos_emb_flashatt(
+    q: torch.Tensor, k: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    cos = cos.chunk(2, dim=-1)[0].contiguous()
+    sin = sin.chunk(2, dim=-1)[0].contiguous()
+    q_embed = apply_rotary_emb(q.float(), cos.float(), sin.float()).type_as(q)
+    k_embed = apply_rotary_emb(k.float(), cos.float(), sin.float()).type_as(k)
+    return q_embed, k_embed
+
+
+# Copied from transformers.models.llama.modeling_llama.rotate_half
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb_vision(
+    q: torch.Tensor, k: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor
+) -> tuple[torch.Tensor, torch.Tensor]:
+    orig_q_dtype = q.dtype
+    orig_k_dtype = k.dtype
+    q, k = q.float(), k.float()
+    cos, sin = cos.unsqueeze(-2).float(), sin.unsqueeze(-2).float()
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    q_embed = q_embed.to(orig_q_dtype)
+    k_embed = k_embed.to(orig_k_dtype)
+    return q_embed, k_embed
+
+
+class Siglip2Attention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
+                f" {self.num_heads})."
+            )
+        self.scale = self.head_dim**-0.5
+        self.dropout = config.attention_dropout
+        self.is_causal = False
+
+        self.k_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.v_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
+
+        self.use_rope = config.use_rope
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cu_seqlens: torch.Tensor,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+    ) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """Input shape: Batch x Time x Channel"""
+
+        seq_length, embed_dim = hidden_states.shape
+
+        queries = self.q_proj(hidden_states)
+        keys = self.k_proj(hidden_states)
+        values = self.v_proj(hidden_states)
+
+        queries = queries.view(seq_length, self.num_heads, self.head_dim)
+        keys = keys.view(seq_length, self.num_heads, self.head_dim)
+        values = values.view(seq_length, self.num_heads, self.head_dim)
+
+        if self.use_rope:
+            cos, sin = position_embeddings
+            if is_flash_attn_2_available():
+                queries, keys = apply_rotary_pos_emb_flashatt(queries.unsqueeze(0), keys.unsqueeze(0), cos, sin)
+            else:
+                queries, keys = apply_rotary_pos_emb_vision(queries.unsqueeze(0), keys.unsqueeze(0), cos, sin)
+            queries = queries.squeeze(0)
+            keys = keys.squeeze(0)
+
+        max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
+        if is_flash_attn_2_available():
+            attn_output = flash_attn_varlen_func(queries, keys, values, cu_seqlens, cu_seqlens, max_seqlen, max_seqlen).reshape(
+                                                seq_length, -1
+                                            )
+        else:
+            batch_size = cu_seqlens.shape[0] - 1
+            outputs = []
+            cu = cu_seqlens.tolist()
+            for i in range(batch_size):
+                start_idx = cu[i]
+                end_idx = cu[i + 1]
+                # Each sequence is processed independently.
+                q_i = queries[start_idx:end_idx].unsqueeze(0)
+                k_i = keys[start_idx:end_idx].unsqueeze(0)
+                v_i = values[start_idx:end_idx].unsqueeze(0)
+                # (1, seq_len, num_heads, head_dim) ->
+                # (1, num_heads, seq_len, head_dim)
+                q_i, k_i, v_i = [x.transpose(1, 2) for x in (q_i, k_i, v_i)]
+                output_i = F.scaled_dot_product_attention(q_i,
+                                                        k_i,
+                                                        v_i,
+                                                        dropout_p=0.0)
+                # (1, num_heads, seq_len, head_dim) -> (seq_len, embed_dim)
+                output_i = output_i.transpose(1, 2).reshape(-1, self.embed_dim)
+                outputs.append(output_i)
+            attn_output = torch.cat(outputs, dim=0)
+
+        attn_output = self.out_proj(attn_output)
+        return attn_output
+
+class Siglip2MLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.activation_fn = ACT2FN[config.hidden_act]
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+        return hidden_states
+
+
+class Siglip2EncoderLayer(nn.Module):
+    def __init__(self, config: Siglip2NavitConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.self_attn = Siglip2Attention(config)
+        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.mlp = Siglip2MLP(config)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cu_seqlens: torch.Tensor, 
+        position_embeddings: torch.Tensor
+    ) -> tuple[torch.FloatTensor]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`):
+                Input to the layer of shape `(batch, seq_len, embed_dim)`.
+            attention_mask (`torch.FloatTensor`):
+                Attention mask of shape `(batch, 1, q_len, k_v_seq_len)` where padding elements are indicated by very large negative values.
+            output_attentions (`bool`, *optional*, defaults to `False`):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+        """
+        residual = hidden_states
+
+        hidden_states = self.layer_norm1(hidden_states)
+        hidden_states = self.self_attn(
+            hidden_states=hidden_states,
+            cu_seqlens=cu_seqlens, 
+            position_embeddings=position_embeddings
+        )
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.layer_norm2(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        return hidden_states
+
+class Siglip2Encoder(nn.Module):
+    """
+    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
+    [`Siglip2EncoderLayer`].
+    Args:
+        config: Siglip2NavitConfig
+    """
+
+    def __init__(self, config: Siglip2NavitConfig):
+        super().__init__()
+        self.config = config
+        self.layers = nn.ModuleList([Siglip2EncoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+        self.rotary_pos_emb = VisionRotaryEmbedding(config.hidden_size // config.num_attention_heads // 2)
+        self.patch_size = config.patch_size
+        self.hidden_stride = config.hidden_stride
+        self.window_size = config.window_size
+        self.spatial_merge_unit = config.hidden_stride * config.hidden_stride
+        self.fullatt_block_indexes = None if config.fullatt_block_indexes is None else [int(i) for i in config.fullatt_block_indexes.split('|')]
+        
+
+    # copied from qwen2.5_vl
+    def rot_pos_emb(self, grid_thw):
+        pos_ids = []
+        for t, h, w in grid_thw:
+            hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w)
+            hpos_ids = hpos_ids.reshape(
+                h // self.hidden_stride,
+                self.hidden_stride,
+                w // self.hidden_stride,
+                self.hidden_stride,
+            )
+            hpos_ids = hpos_ids.permute(0, 2, 1, 3)
+            hpos_ids = hpos_ids.flatten()
+
+            wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1)
+            wpos_ids = wpos_ids.reshape(
+                h // self.hidden_stride,
+                self.hidden_stride,
+                w // self.hidden_stride,
+                self.hidden_stride,
+            )
+            wpos_ids = wpos_ids.permute(0, 2, 1, 3)
+            wpos_ids = wpos_ids.flatten()
+            pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
+        pos_ids = torch.cat(pos_ids, dim=0)
+        max_grid_size = grid_thw[:, 1:].max()
+        rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
+        rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
+        return rotary_pos_emb
+
+    def get_window_index(self, grid_thw):
+        window_index: list = []
+        cu_window_seqlens: list = [0]
+        window_index_id = 0
+        vit_merger_window_size = self.window_size // self.hidden_stride // self.patch_size  # patch (after merge) number in each window
+
+        for grid_t, grid_h, grid_w in grid_thw:
+            llm_grid_h, llm_grid_w = (
+                grid_h // self.hidden_stride,  # number of patch after merge
+                grid_w // self.hidden_stride,
+            )
+            index = torch.arange(grid_t * llm_grid_h * llm_grid_w).reshape(grid_t, llm_grid_h, llm_grid_w)
+            pad_h = vit_merger_window_size - llm_grid_h % vit_merger_window_size
+            pad_w = vit_merger_window_size - llm_grid_w % vit_merger_window_size
+            num_windows_h = (llm_grid_h + pad_h) // vit_merger_window_size
+            num_windows_w = (llm_grid_w + pad_w) // vit_merger_window_size
+            index_padded = F.pad(index, (0, pad_w, 0, pad_h), "constant", -100)
+            index_padded = index_padded.reshape(
+                grid_t,
+                num_windows_h,
+                vit_merger_window_size,
+                num_windows_w,
+                vit_merger_window_size,
+            )
+            index_padded = index_padded.permute(0, 1, 3, 2, 4).reshape(
+                grid_t,
+                num_windows_h * num_windows_w,
+                vit_merger_window_size,
+                vit_merger_window_size,
+            )
+            seqlens = (index_padded != -100).sum([2, 3]).reshape(-1)
+            index_padded = index_padded.reshape(-1)
+            index_new = index_padded[index_padded != -100]
+            window_index.append(index_new + window_index_id)
+            cu_seqlens_tmp = seqlens.cumsum(0) * self.spatial_merge_unit + cu_window_seqlens[-1]
+            cu_window_seqlens.extend(cu_seqlens_tmp.tolist())
+            window_index_id += (grid_t * llm_grid_h * llm_grid_w).item()
+        window_index = torch.cat(window_index, dim=0)
+
+        return window_index, cu_window_seqlens
+
+    # Ignore copy
+    def forward(
+        self,
+        inputs_embeds,
+        grid_thws: torch.Tensor,
+        output_hidden_states: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor, ...]]]:
+        r"""
+        Args:
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
+            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+                [What are attention masks?](../glossary#attention-mask)
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        """
+
+        rotary_pos_emb = self.rot_pos_emb(grid_thws)
+        window_index, cu_window_seqlens = self.get_window_index(grid_thws)
+        cu_window_seqlens = torch.tensor(
+            cu_window_seqlens,
+            device=inputs_embeds.device,
+            dtype=grid_thws.dtype if torch.jit.is_tracing() else torch.int32,
+        )
+        cu_window_seqlens = torch.unique_consecutive(cu_window_seqlens)
+
+        seq_len, _ = inputs_embeds.size()
+        inputs_embeds = inputs_embeds.reshape(seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
+        inputs_embeds = inputs_embeds[window_index, :, :]
+        inputs_embeds = inputs_embeds.reshape(seq_len, -1)
+        rotary_pos_emb = rotary_pos_emb.reshape(seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
+        rotary_pos_emb = rotary_pos_emb[window_index, :, :]
+        rotary_pos_emb = rotary_pos_emb.reshape(seq_len, -1)
+        emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
+        position_embeddings = (emb.cos(), emb.sin())
+
+        cu_seqlens = torch.repeat_interleave(grid_thws[:, 1] * grid_thws[:, 2], grid_thws[:, 0]).cumsum(
+            dim=0,
+            # Select dtype based on the following factors:
+            #  - FA2 requires that cu_seqlens_q must have dtype int32
+            #  - torch.onnx.export requires that cu_seqlens_q must have same dtype as grid_thw
+            # See https://github.com/huggingface/transformers/pull/34852 for more information
+            dtype=grid_thws.dtype if torch.jit.is_tracing() else torch.int32,
+        )
+        cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
+
+        reverse_indices = torch.argsort(window_index)
+        encoder_states = () if output_hidden_states else None
+
+        hidden_states = inputs_embeds
+        for index, block in enumerate(self.layers):
+            if self.fullatt_block_indexes is None or index in self.fullatt_block_indexes:
+                cu_seqlens_tmp = cu_seqlens
+            else:
+                cu_seqlens_tmp = cu_window_seqlens
+            if self.gradient_checkpointing and self.training:
+                hidden_states = self._gradient_checkpointing_func(block.__call__, hidden_states, cu_seqlens_tmp, position_embeddings)
+            else:
+                hidden_states = block(hidden_states, cu_seqlens_tmp, position_embeddings)
+            if output_hidden_states:
+                hidden_states_ = hidden_states.reshape(seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
+                encoder_states += (hidden_states_[reverse_indices, :].reshape(seq_len, -1),)
+        # tokens = self.post_trunk_norm(tokens)
+        hidden_states = hidden_states.reshape(seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
+        hidden_states = hidden_states[reverse_indices, :].reshape(seq_len, -1)
+
+        return hidden_states, encoder_states
+
+class Siglip2VisionTransformer(nn.Module):
+    def __init__(self, config: Siglip2NavitConfig):
+        super().__init__()
+        self.config = config
+        embed_dim = config.hidden_size
+
+        self.embeddings = Siglip2VisionEmbeddings(config)
+        self.encoder = Siglip2Encoder(config)
+        self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
+
+    def forward(
+        self,
+        pixel_values: torch.FloatTensor,
+        grid_thws: torch.LongTensor,
+        output_hidden_states: Optional[bool] = True,
+        return_dict: Optional[bool] = True,
+    ) -> Union[
+        Tuple[torch.Tensor],
+        Tuple[torch.Tensor, Tuple[torch.Tensor, ...]],
+        BaseModelOutputWithNoAttention,
+    ]:
+        r"""
+        spatial_shapes (`torch.LongTensor` of shape `(batch_size, 2)`):
+            Tensor containing the spatial dimensions (height, width) of the input images.
+        """
+        # 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
+        # )
+
+        hidden_states = self.embeddings(pixel_values, grid_thws)
+
+        last_hidden_state, hidden_states = self.encoder(hidden_states, grid_thws, output_hidden_states)
+        last_hidden_state = self.post_layernorm(last_hidden_state)
+
+        if not return_dict:
+            output = (last_hidden_state,)
+            output += (hidden_states,) if output_hidden_states else ()
+            return output
+        
+        return BaseModelOutputWithNoAttention(
+                last_hidden_state=last_hidden_state, 
+                hidden_states=hidden_states
+            )
+       
+class Siglip2PreTrainedModel(PreTrainedModel):
+    config_class = Siglip2NavitConfig
+    base_model_prefix = "siglip2_navit"
+    supports_gradient_checkpointing = True
+
+    _no_split_modules = [
+        "Siglip2VisionEmbeddings",
+        "Siglip2EncoderLayer",
+    ]
+    _supports_flash_attn_2 = True
+    _supports_sdpa = False
+    _supports_flex_attn = False
+    _supports_attention_backend = True
+
+
+class Siglip2NavitModel(Siglip2PreTrainedModel):
+    config_class = Siglip2NavitConfig
+    main_input_name = "pixel_values"
+
+    def __init__(self, config: Siglip2NavitConfig):
+        super().__init__(config)
+
+        self.vision_model = Siglip2VisionTransformer(config)
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.vision_model.embeddings.patch_embedding
+
+    def forward(
+        self,
+        pixel_values: torch.FloatTensor,
+        grid_thws: torch.LongTensor,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[
+        Tuple[torch.Tensor],
+        Tuple[torch.Tensor, Tuple[torch.Tensor, ...]],
+        BaseModelOutputWithNoAttention,
+    ]:
+        
+        if output_hidden_states is None:
+            output_hidden_states = self.config.output_hidden_states
+        if return_dict is None:
+            return_dict = self.config.use_return_dict
+
+        return self.vision_model(
+            pixel_values=pixel_values,
+            grid_thws=grid_thws,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+class VisualEmbedding(torch.nn.Embedding):
+    """
+    A visual embedding layer that can handle both discrete token IDs (long) and continuous
+    soft-token probabilities (float).
+    """
+
+    def forward(self, visual_tokens: Tensor) -> Tensor:
+        if visual_tokens.dtype in [torch.int8, torch.int16, torch.int32, torch.int64, torch.long]:
+            return super().forward(visual_tokens)
+        # Handle soft tokens (probabilities) by matrix multiplication with the embedding weight
+        return torch.matmul(visual_tokens, self.weight)
+
+
+class VisualTokenizer(torch.nn.Module):
+    """
+    Tokenizes images or videos into a sequence of continuous visual tokens.
+    """
+
+    def __init__(self, vit, visual_vocab_size, image_processor_name_or_path, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.vit = vit
+        self.image_processor = AutoImageProcessor.from_pretrained(image_processor_name_or_path, do_center_crop=False)
+        head_dim = visual_vocab_size - len(INDICATOR_IDS)
+        self.head = torch.nn.Sequential(
+            torch.nn.Linear(self.vit.config.hidden_size * self.vit.config.hidden_stride ** 2, head_dim, bias=False),
+            torch.nn.LayerNorm(head_dim)
+        )
+
+    def _encode(self, pixel_values, grid_thws):
+        output = self.vit(pixel_values, grid_thws, output_hidden_states=True, return_dict=True)
+        features = output.hidden_states[-1]
+        seq_len, _ = features.shape
+        features = features.reshape(seq_len // (self.vit.config.hidden_stride ** 2), -1)
+        return features
+
+    # Adapted from qwen2_vl
+    @staticmethod
+    def smart_resize(
+        height: int, width: int, factor: int = 28, min_pixels: int = 448 * 448, max_pixels: int = 1344 * 1792
+    ):
+        """Rescales the image so that the following conditions are met:
+        1. Both dimensions are divisible by 'factor'.
+        2. The total number of pixels is within ['min_pixels', 'max_pixels'].
+        3. The aspect ratio is maintained as closely as possible.
+        """
+        if height < factor or width < factor:
+            if height < width:
+                width = round(factor / height * width)
+                height = factor
+            else:
+                height = round(factor / width * height)
+                width = factor
+
+        elif max(height, width) / min(height, width) > 200:
+            if height > width:
+                height = 200 * width
+            else:
+                width = 200 * height
+
+        h_bar = round(height / factor) * factor
+        w_bar = round(width / factor) * factor
+        if h_bar * w_bar > max_pixels:
+            beta = math.sqrt((height * width) / max_pixels)
+            h_bar = math.floor(height / beta / factor) * factor
+            w_bar = math.floor(width / beta / factor) * factor
+        elif h_bar * w_bar < min_pixels:
+            beta = math.sqrt(min_pixels / (height * width))
+            h_bar = math.ceil(height * beta / factor) * factor
+            w_bar = math.ceil(width * beta / factor) * factor
+        return h_bar, w_bar
+
+    def preprocess(
+        self,
+        image: Optional[PIL.Image.Image] = None,
+        video: Optional[List[PIL.Image.Image]] = None,
+        min_pixels: Optional[int] = None,
+        max_pixels: Optional[int] = None
+    ):
+        patch_size = self.vit.config.patch_size
+        temporal_patch_size = self.vit.config.temporal_patch_size
+        hidden_stride = self.vit.config.hidden_stride
+        assert (image is None) ^ (video is None), "Invalid input: expect either image or video"
+        if image is not None:
+            images = [image]
+        else:
+            images = video
+        images = [image.convert("RGB") if image.mode != 'RGB' else image for image in images]
+        width, height = images[0].size
+        processed_images = []
+        for image in images:
+            resized_height, resized_width = self.smart_resize(
+                height,
+                width,
+                factor=patch_size * hidden_stride,
+                min_pixels=min_pixels,
+                max_pixels=max_pixels,
+            )
+            new_size = dict(height=resized_height, width=resized_width)
+            new_image = self.image_processor.preprocess(image, size=new_size, return_tensors="np")['pixel_values'][0]
+            processed_images.append(new_image)
+
+        patches = np.array(processed_images)
+        if patches.shape[0] % temporal_patch_size != 0:
+            repeats = np.repeat(patches[-1][np.newaxis], temporal_patch_size - 1, axis=0)
+            patches = np.concatenate([patches, repeats], axis=0)
+        channel = patches.shape[1]
+        grid_t = patches.shape[0] // temporal_patch_size
+        grid_h, grid_w = resized_height // patch_size, resized_width // patch_size
+        grid_thw = torch.tensor([[grid_t, grid_h, grid_w]])
+
+        patches = patches.reshape(
+            grid_t, temporal_patch_size, channel,
+            grid_h // hidden_stride, hidden_stride, patch_size,
+            grid_w // hidden_stride, hidden_stride, patch_size,
+        )
+        patches = patches.transpose(0, 3, 6, 4, 7, 2, 1, 5, 8)
+        flatten_patches = patches.reshape(
+            grid_t * grid_h * grid_w, channel * temporal_patch_size * patch_size * patch_size
+        )
+        flatten_patches = torch.tensor(flatten_patches)
+
+        return flatten_patches, grid_thw
+
+    def forward(
+        self, pixel_values, grid_thws
+    ) -> torch.Tensor:  # [BatchSize, ImageShape] -> [BatchSize, #Token, VocabSize]
+        features = self._encode(pixel_values, grid_thws)
+        logits = self.head(features)
+        tokens = torch.softmax(logits, dim=-1, dtype=torch.float32).to(logits.dtype)
+
+        token_len, _ = tokens.shape
+        padding_tensor = torch.zeros(size=(token_len, len(INDICATOR_IDS)),
+                                     dtype=tokens.dtype,
+                                     device=tokens.device,
+                                     layout=tokens.layout,
+                                     requires_grad=False)
+        tokens = torch.cat((tokens, padding_tensor), dim=1)
+        return tokens
+
+
+class Ovis2_6_PreTrainedModel(PreTrainedModel):
+    config_class = Ovis2_6_Config
+    base_model_prefix = "ovis2_6"
+
+
+class Ovis2_6ForCausalLM(Ovis2_6_PreTrainedModel):
+    _supports_flash_attn_2 = True
+
+    def __init__(self, config: Ovis2_6_Config, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+
+        self.llm = AutoModelForCausalLM.from_config(self.config.llm_config)
+        assert self.config.hidden_size == self.llm.config.hidden_size, "hidden size mismatch"
+        self.text_tokenizer = AutoTokenizer.from_pretrained(self.config.name_or_path)
+        self.visual_tokenizer = VisualTokenizer(vit=AutoModel.from_config(self.config.vit_config),
+                                                visual_vocab_size=self.config.visual_vocab_size,
+                                                image_processor_name_or_path=self.config.name_or_path)
+
+        self.vte = VisualEmbedding(self.config.visual_vocab_size, self.config.hidden_size,
+                                   device=self.visual_tokenizer.vit.device, dtype=self.visual_tokenizer.vit.dtype)
+        indicator_token_indices = torch.arange(
+            self.config.visual_vocab_size - len(INDICATOR_IDS),
+            self.config.visual_vocab_size,
+            dtype=torch.long
+        )
+        self.register_buffer("indicator_token_indices", indicator_token_indices, persistent=False)
+
+        def _merge_modules(modules_list: tuple):
+            merged_modules = []
+            for modules in modules_list:
+                merged_modules.extend(modules if modules else [])
+            return merged_modules
+
+        # Standard model configurations for parallelism and device placement
+        self._no_split_modules = _merge_modules(
+            (self.llm._no_split_modules, self.visual_tokenizer.vit._no_split_modules))
+        self._skip_keys_device_placement = self.llm._skip_keys_device_placement
+        self._keep_in_fp32_modules = _merge_modules(
+            (self.llm._keep_in_fp32_modules, self.visual_tokenizer.vit._keep_in_fp32_modules))
+        self.is_parallelizable = all((self.llm.is_parallelizable, self.visual_tokenizer.vit.is_parallelizable))
+        self.supports_gradient_checkpointing = True
+
+    def tie_weights(self):
+        self.llm.tie_weights()
+
+    def get_wte(self):
+        return self.llm.get_input_embeddings()
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        attention_mask: torch.Tensor,
+        pixel_values: Optional[torch.Tensor],
+        grid_thws: Optional[torch.Tensor],
+        labels: Optional[torch.Tensor] = None,
+        **kwargs
+    ):
+        inputs_embeds = self.merge_multimodal(
+            input_ids=input_ids,
+            pixel_values=pixel_values,
+            grid_thws=grid_thws,
+        )
+        return self.llm(inputs_embeds=inputs_embeds, attention_mask=attention_mask, labels=labels, **kwargs)
+
+    def merge_multimodal(
+        self,
+        input_ids: torch.Tensor,
+        pixel_values: Optional[torch.Tensor],
+        grid_thws: Optional[torch.Tensor],
+    ):
+        placeholder_token_mask = torch.lt(input_ids, 0)
+        multimodal_embeds = self.get_wte()(torch.masked_fill(input_ids, placeholder_token_mask, 0))
+
+        if pixel_values is not None:
+            visual_indicator_embeds = self.vte(self.indicator_token_indices).to(
+                dtype=multimodal_embeds.dtype, device=multimodal_embeds.device
+            )
+            visual_tokens = self.visual_tokenizer(pixel_values, grid_thws)
+            visual_embeds = self.vte(visual_tokens).to(dtype=multimodal_embeds.dtype, device=multimodal_embeds.device)
+
+            for i, indicator_id in enumerate(INDICATOR_IDS):
+                multimodal_embeds[input_ids == indicator_id] = visual_indicator_embeds[i]
+            multimodal_embeds[input_ids == VISUAL_ATOM_ID] = visual_embeds
+
+        return multimodal_embeds
+
+    def _merge_inputs(
+        self, raw_input_ids, placeholder_id, grid_thws, indicator_begin_id, indicator_end_id
+    ):
+        input_ids = []
+        prev_index = 0
+        placeholder_indexes = [i for i, v in enumerate(raw_input_ids) if v == placeholder_id]
+        for placeholder_index, grid_thw in zip(placeholder_indexes, grid_thws):
+            input_ids.extend(raw_input_ids[prev_index:placeholder_index])
+            num_image_atoms = grid_thw.prod().item()
+            num_image_atoms //= self.visual_tokenizer.vit.config.hidden_stride ** 2
+            num_image_atoms //= self.visual_tokenizer.vit.config.temporal_patch_size
+            input_ids.extend([indicator_begin_id] + [VISUAL_ATOM_ID] * num_image_atoms + [indicator_end_id])
+            prev_index = placeholder_index + 1
+        input_ids.extend(raw_input_ids[prev_index:])
+        return input_ids
+
+    def _tokenize_with_visual_placeholder(self, text):
+        placeholder = VIDEO_PLACEHOLDER if VIDEO_PLACEHOLDER in text else IMAGE_PLACEHOLDER
+        placeholder_id = VIDEO_PLACEHOLDER_ID if VIDEO_PLACEHOLDER in text else IMAGE_PLACEHOLDER_ID
+        chunks = [self.text_tokenizer(chunk, add_special_tokens=False).input_ids for chunk in text.split(placeholder)]
+        input_ids = chunks[0]
+        for chunk in chunks[1:]:
+            input_ids.append(placeholder_id)
+            input_ids.extend(chunk)
+        return input_ids
+
+    def preprocess_inputs(
+        self,
+        messages: List[Union[str, Dict]],
+        min_pixels=448 * 448,
+        max_pixels=1792 * 1792,
+        add_generation_prompt=True,
+        enable_thinking=False
+    ):
+        text = self.text_tokenizer.apply_chat_template(
+            messages,
+            tokenize=False,
+            add_generation_prompt=add_generation_prompt,
+            enable_thinking=enable_thinking
+        )
+        input_ids = self._tokenize_with_visual_placeholder(text)
+        images = []
+        videos = []
+        for message in messages:
+            content = message["content"]
+            if isinstance(content, list):
+                images.extend([item["image"] for item in content if item.get("image") is not None])
+                videos.extend([item["video"] for item in content if item.get("video") is not None])
+        if images and videos:
+            raise ValueError(
+                "Multiple visual input data types detected (both image and video provided). "
+                "This model supports only one type of visual input data at a time. "
+                "Please provide either image or video, but not both."
+            )
+
+        pixel_values, grid_thws = None, None
+        if images:
+            pixel_values, grid_thws = zip(
+                *(self.visual_tokenizer.preprocess(image=image, min_pixels=min_pixels, max_pixels=max_pixels)
+                  for image in images)
+            )
+            input_ids = self._merge_inputs(
+                input_ids, IMAGE_PLACEHOLDER_ID, grid_thws, INDICATOR_IDS[0], INDICATOR_IDS[1]
+            )
+            pixel_values = torch.cat(pixel_values, dim=0)
+            grid_thws = torch.cat(grid_thws, dim=0)
+        elif videos:
+            assert len(videos) == 1, "only support single video"
+            pixel_values, grid_thws = self.visual_tokenizer.preprocess(
+                video=videos[0], min_pixels=min_pixels, max_pixels=max_pixels
+            )
+            input_ids = self._merge_inputs(
+                input_ids, VIDEO_PLACEHOLDER_ID, grid_thws, INDICATOR_IDS[2], INDICATOR_IDS[3]
+            )
+
+        input_ids = torch.tensor(input_ids, dtype=torch.long).unsqueeze(0)
+
+        return input_ids, pixel_values, grid_thws
+
+    def generate(
+        self,
+        inputs: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> Union[GenerateOutput, torch.LongTensor]:
+        attention_mask = torch.ne(inputs, self.text_tokenizer.pad_token_id).to(device=inputs.device)
+        inputs_embeds = self.merge_multimodal(
+            input_ids=inputs,
+            pixel_values=kwargs.pop('pixel_values', None),
+            grid_thws=kwargs.pop('grid_thws', None)
+        )
+        enable_thinking = kwargs.pop('enable_thinking', False)
+        enable_thinking_budget = kwargs.pop('enable_thinking_budget', False)
+        thinking_budget = kwargs.pop('thinking_budget', 1024)
+        
+        if enable_thinking and enable_thinking_budget:
+            actual_max_new_tokens = kwargs['max_new_tokens']
+            kwargs['max_new_tokens'] = thinking_budget
+            generated_ids = self.llm.generate(inputs=None, inputs_embeds=inputs_embeds, attention_mask=attention_mask, **kwargs)
+            output_ids = generated_ids
+            output_ids_list = generated_ids[0]
+            
+            # check if the generation has already finished (151645 is <|im_end|>)
+            if 151645 not in output_ids_list:
+                # check if the thinking process has finished (151668 is </think>)
+                # and prepare the second model input
+                if 151668 not in output_ids_list:
+                    early_stopping_text = "\n\nConsidering the limited time by the user, I have to give the solution based on the thinking directly now.\n</think>\n\n"
+                    early_stopping_ids = self.text_tokenizer(early_stopping_text, return_tensors="pt", return_attention_mask=False).input_ids.to(inputs.device)
+                    input_ids_appendent = torch.cat([output_ids, early_stopping_ids], dim=-1)
+                    kwargs['streamer'].put(early_stopping_ids) if 'streamer' in kwargs else None
+                else:
+                    input_ids_appendent = output_ids
+                
+
+                # second generation
+                new_inputs = torch.cat([inputs, input_ids_appendent], dim=-1)
+                attention_mask = torch.ne(new_inputs, self.text_tokenizer.pad_token_id).to(device=inputs.device)
+                inputs_embeds_appendent = self.merge_multimodal(
+                    input_ids=input_ids_appendent,
+                    pixel_values=None,
+                    grid_thws=None
+                )
+                new_inputs_embeds = torch.cat([inputs_embeds, inputs_embeds_appendent], dim=-2)
+                
+                kwargs['max_new_tokens'] = inputs_embeds.size(-2) + actual_max_new_tokens - new_inputs_embeds.size(-2)
+                generated_ids2 = self.llm.generate(inputs=None, inputs_embeds=new_inputs_embeds, attention_mask=attention_mask, **kwargs)
+                kwargs['streamer'].manual_end() if 'streamer' in kwargs else None
+                return torch.cat([input_ids_appendent, generated_ids2], dim=-1)
+            
+            else:
+                kwargs['streamer'].manual_end() if 'streamer' in kwargs else None
+                return generated_ids
+            
+        else:
+            generated_ids = self.llm.generate(inputs=None, inputs_embeds=inputs_embeds, attention_mask=attention_mask, **kwargs)
+            kwargs['streamer'].manual_end() if 'streamer' in kwargs else None
+            return generated_ids
+        
+        
+        
+        
+        
+class Ovis2_6_Moe_PreTrainedModel(PreTrainedModel):
+    config_class = Ovis2_6_Moe_Config
+    base_model_prefix = "ovis2_6_moe"
+
+
+class Ovis2_6_MoeForCausalLM(Ovis2_6_Moe_PreTrainedModel):
+    _supports_flash_attn_2 = True
+
+    def __init__(self, config: Ovis2_6_Moe_Config, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+
+        self.llm = AutoModelForCausalLM.from_config(self.config.llm_config)
+        assert self.config.hidden_size == self.llm.config.hidden_size, "hidden size mismatch"
+        self.text_tokenizer = AutoTokenizer.from_pretrained(self.config.name_or_path)
+        self.visual_tokenizer = VisualTokenizer(vit=AutoModel.from_config(self.config.vit_config),
+                                                visual_vocab_size=self.config.visual_vocab_size,
+                                                image_processor_name_or_path=self.config.name_or_path)
+
+        self.vte = VisualEmbedding(self.config.visual_vocab_size, self.config.hidden_size,
+                                   device=self.visual_tokenizer.vit.device, dtype=self.visual_tokenizer.vit.dtype)
+        indicator_token_indices = torch.arange(
+            self.config.visual_vocab_size - len(INDICATOR_IDS),
+            self.config.visual_vocab_size,
+            dtype=torch.long
+        )
+        self.register_buffer("indicator_token_indices", indicator_token_indices, persistent=False)
+
+        def _merge_modules(modules_list: tuple):
+            merged_modules = []
+            for modules in modules_list:
+                merged_modules.extend(modules if modules else [])
+            return merged_modules
+
+        # Standard model configurations for parallelism and device placement
+        self._no_split_modules = _merge_modules(
+            (self.llm._no_split_modules, self.visual_tokenizer.vit._no_split_modules))
+        self._skip_keys_device_placement = self.llm._skip_keys_device_placement
+        self._keep_in_fp32_modules = _merge_modules(
+            (self.llm._keep_in_fp32_modules, self.visual_tokenizer.vit._keep_in_fp32_modules))
+        self.is_parallelizable = all((self.llm.is_parallelizable, self.visual_tokenizer.vit.is_parallelizable))
+        self.supports_gradient_checkpointing = True
+
+    def tie_weights(self):
+        self.llm.tie_weights()
+
+    def get_wte(self):
+        return self.llm.get_input_embeddings()
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        attention_mask: torch.Tensor,
+        pixel_values: Optional[torch.Tensor],
+        grid_thws: Optional[torch.Tensor],
+        labels: Optional[torch.Tensor] = None,
+        **kwargs
+    ):
+        inputs_embeds = self.merge_multimodal(
+            input_ids=input_ids,
+            pixel_values=pixel_values,
+            grid_thws=grid_thws,
+        )
+        return self.llm(inputs_embeds=inputs_embeds, attention_mask=attention_mask, labels=labels, **kwargs)
+
+    def merge_multimodal(
+        self,
+        input_ids: torch.Tensor,
+        pixel_values: Optional[torch.Tensor],
+        grid_thws: Optional[torch.Tensor],
+    ):
+        placeholder_token_mask = torch.lt(input_ids, 0)
+        multimodal_embeds = self.get_wte()(torch.masked_fill(input_ids, placeholder_token_mask, 0))
+
+        if pixel_values is not None:
+            visual_indicator_embeds = self.vte(self.indicator_token_indices).to(
+                dtype=multimodal_embeds.dtype, device=multimodal_embeds.device
+            )
+            visual_tokens = self.visual_tokenizer(pixel_values, grid_thws)
+            visual_embeds = self.vte(visual_tokens).to(dtype=multimodal_embeds.dtype, device=multimodal_embeds.device)
+
+            for i, indicator_id in enumerate(INDICATOR_IDS):
+                multimodal_embeds[input_ids == indicator_id] = visual_indicator_embeds[i]
+            multimodal_embeds[input_ids == VISUAL_ATOM_ID] = visual_embeds
+
+        return multimodal_embeds
+
+    def _merge_inputs(
+        self, raw_input_ids, placeholder_id, grid_thws, indicator_begin_id, indicator_end_id
+    ):
+        input_ids = []
+        prev_index = 0
+        placeholder_indexes = [i for i, v in enumerate(raw_input_ids) if v == placeholder_id]
+        for placeholder_index, grid_thw in zip(placeholder_indexes, grid_thws):
+            input_ids.extend(raw_input_ids[prev_index:placeholder_index])
+            num_image_atoms = grid_thw.prod().item()
+            num_image_atoms //= self.visual_tokenizer.vit.config.hidden_stride ** 2
+            num_image_atoms //= self.visual_tokenizer.vit.config.temporal_patch_size
+            input_ids.extend([indicator_begin_id] + [VISUAL_ATOM_ID] * num_image_atoms + [indicator_end_id])
+            prev_index = placeholder_index + 1
+        input_ids.extend(raw_input_ids[prev_index:])
+        return input_ids
+
+    def _tokenize_with_visual_placeholder(self, text):
+        placeholder = VIDEO_PLACEHOLDER if VIDEO_PLACEHOLDER in text else IMAGE_PLACEHOLDER
+        placeholder_id = VIDEO_PLACEHOLDER_ID if VIDEO_PLACEHOLDER in text else IMAGE_PLACEHOLDER_ID
+        chunks = [self.text_tokenizer(chunk, add_special_tokens=False).input_ids for chunk in text.split(placeholder)]
+        input_ids = chunks[0]
+        for chunk in chunks[1:]:
+            input_ids.append(placeholder_id)
+            input_ids.extend(chunk)
+        return input_ids
+
+    def preprocess_inputs(
+        self,
+        messages: List[Union[str, Dict]],
+        min_pixels=448 * 448,
+        max_pixels=1792 * 1792,
+        add_generation_prompt=True,
+        enable_thinking=False
+    ):
+        text = self.text_tokenizer.apply_chat_template(
+            messages,
+            tokenize=False,
+            add_generation_prompt=add_generation_prompt,
+            enable_thinking=enable_thinking
+        )
+        input_ids = self._tokenize_with_visual_placeholder(text)
+        images = []
+        videos = []
+        for message in messages:
+            content = message["content"]
+            if isinstance(content, list):
+                images.extend([item["image"] for item in content if item.get("image") is not None])
+                videos.extend([item["video"] for item in content if item.get("video") is not None])
+        if images and videos:
+            raise ValueError(
+                "Multiple visual input data types detected (both image and video provided). "
+                "This model supports only one type of visual input data at a time. "
+                "Please provide either image or video, but not both."
+            )
+
+        pixel_values, grid_thws = None, None
+        if images:
+            pixel_values, grid_thws = zip(
+                *(self.visual_tokenizer.preprocess(image=image, min_pixels=min_pixels, max_pixels=max_pixels)
+                  for image in images)
+            )
+            input_ids = self._merge_inputs(
+                input_ids, IMAGE_PLACEHOLDER_ID, grid_thws, INDICATOR_IDS[0], INDICATOR_IDS[1]
+            )
+            pixel_values = torch.cat(pixel_values, dim=0)
+            grid_thws = torch.cat(grid_thws, dim=0)
+        elif videos:
+            assert len(videos) == 1, "only support single video"
+            pixel_values, grid_thws = self.visual_tokenizer.preprocess(
+                video=videos[0], min_pixels=min_pixels, max_pixels=max_pixels
+            )
+            input_ids = self._merge_inputs(
+                input_ids, VIDEO_PLACEHOLDER_ID, grid_thws, INDICATOR_IDS[2], INDICATOR_IDS[3]
+            )
+
+        input_ids = torch.tensor(input_ids, dtype=torch.long).unsqueeze(0)
+
+        return input_ids, pixel_values, grid_thws
+
+    def generate(
+        self,
+        inputs: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> Union[GenerateOutput, torch.LongTensor]:
+        attention_mask = torch.ne(inputs, self.text_tokenizer.pad_token_id).to(device=inputs.device)
+        inputs_embeds = self.merge_multimodal(
+            input_ids=inputs,
+            pixel_values=kwargs.pop('pixel_values', None),
+            grid_thws=kwargs.pop('grid_thws', None)
+        )
+        enable_thinking = kwargs.pop('enable_thinking', False)
+        enable_thinking_budget = kwargs.pop('enable_thinking_budget', False)
+        thinking_budget = kwargs.pop('thinking_budget', 1024)
+        
+        if enable_thinking and enable_thinking_budget:
+            actual_max_new_tokens = kwargs['max_new_tokens']
+            kwargs['max_new_tokens'] = thinking_budget
+            generated_ids = self.llm.generate(inputs=None, inputs_embeds=inputs_embeds, attention_mask=attention_mask, **kwargs)
+            output_ids = generated_ids
+            output_ids_list = generated_ids[0]
+            
+            # check if the generation has already finished (151645 is <|im_end|>)
+            if 151645 not in output_ids_list:
+                # check if the thinking process has finished (151668 is </think>)
+                # and prepare the second model input
+                if 151668 not in output_ids_list:
+                    early_stopping_text = "\n\nConsidering the limited time by the user, I have to give the solution based on the thinking directly now.\n</think>\n\n"
+                    early_stopping_ids = self.text_tokenizer(early_stopping_text, return_tensors="pt", return_attention_mask=False).input_ids.to(inputs.device)
+                    input_ids_appendent = torch.cat([output_ids, early_stopping_ids], dim=-1)
+                    kwargs['streamer'].put(early_stopping_ids) if 'streamer' in kwargs else None
+                else:
+                    input_ids_appendent = output_ids
+                
+
+                # second generation
+                new_inputs = torch.cat([inputs, input_ids_appendent], dim=-1)
+                attention_mask = torch.ne(new_inputs, self.text_tokenizer.pad_token_id).to(device=inputs.device)
+                inputs_embeds_appendent = self.merge_multimodal(
+                    input_ids=input_ids_appendent,
+                    pixel_values=None,
+                    grid_thws=None
+                )
+                new_inputs_embeds = torch.cat([inputs_embeds, inputs_embeds_appendent], dim=-2)
+                
+                kwargs['max_new_tokens'] = inputs_embeds.size(-2) + actual_max_new_tokens - new_inputs_embeds.size(-2)
+                generated_ids2 = self.llm.generate(inputs=None, inputs_embeds=new_inputs_embeds, attention_mask=attention_mask, **kwargs)
+                kwargs['streamer'].manual_end() if 'streamer' in kwargs else None
+                return torch.cat([input_ids_appendent, generated_ids2], dim=-1)
+            
+            else:
+                kwargs['streamer'].manual_end() if 'streamer' in kwargs else None
+                return generated_ids
+            
+        else:
+            generated_ids = self.llm.generate(inputs=None, inputs_embeds=inputs_embeds, attention_mask=attention_mask, **kwargs)
+            kwargs['streamer'].manual_end() if 'streamer' in kwargs else None
+            return generated_ids
+
+        
+class Ovis2_6_Next_PreTrainedModel(PreTrainedModel):
+    config_class = Ovis2_6_Next_Config
+    base_model_prefix = "ovis2_6_next"
+
+
+class Ovis2_6_NextForCausalLM(Ovis2_6_Next_PreTrainedModel):
+    _supports_flash_attn_2 = True
+
+    def __init__(self, config: Ovis2_6_Next_Config, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+
+        self.llm = AutoModelForCausalLM.from_config(self.config.llm_config)
+        assert self.config.hidden_size == self.llm.config.hidden_size, "hidden size mismatch"
+        self.text_tokenizer = AutoTokenizer.from_pretrained(self.config.name_or_path)
+        self.visual_tokenizer = VisualTokenizer(vit=AutoModel.from_config(self.config.vit_config),
+                                                visual_vocab_size=self.config.visual_vocab_size,
+                                                image_processor_name_or_path=self.config.name_or_path)
+
+        self.vte = VisualEmbedding(self.config.visual_vocab_size, self.config.hidden_size,
+                                   device=self.visual_tokenizer.vit.device, dtype=self.visual_tokenizer.vit.dtype)
+        indicator_token_indices = torch.arange(
+            self.config.visual_vocab_size - len(INDICATOR_IDS),
+            self.config.visual_vocab_size,
+            dtype=torch.long
+        )
+        self.register_buffer("indicator_token_indices", indicator_token_indices, persistent=False)
+
+        def _merge_modules(modules_list: tuple):
+            merged_modules = []
+            for modules in modules_list:
+                merged_modules.extend(modules if modules else [])
+            return merged_modules
+
+        # Standard model configurations for parallelism and device placement
+        self._no_split_modules = _merge_modules(
+            (self.llm._no_split_modules, self.visual_tokenizer.vit._no_split_modules))
+        self._skip_keys_device_placement = self.llm._skip_keys_device_placement
+        self._keep_in_fp32_modules = _merge_modules(
+            (self.llm._keep_in_fp32_modules, self.visual_tokenizer.vit._keep_in_fp32_modules))
+        self.is_parallelizable = all((self.llm.is_parallelizable, self.visual_tokenizer.vit.is_parallelizable))
+        self.supports_gradient_checkpointing = True
+
+    def tie_weights(self):
+        self.llm.tie_weights()
+
+    def get_wte(self):
+        return self.llm.get_input_embeddings()
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        attention_mask: torch.Tensor,
+        pixel_values: Optional[torch.Tensor],
+        grid_thws: Optional[torch.Tensor],
+        labels: Optional[torch.Tensor] = None,
+        **kwargs
+    ):
+        inputs_embeds = self.merge_multimodal(
+            input_ids=input_ids,
+            pixel_values=pixel_values,
+            grid_thws=grid_thws,
+        )
+        return self.llm(inputs_embeds=inputs_embeds, attention_mask=attention_mask, labels=labels, **kwargs)
+
+    def merge_multimodal(
+        self,
+        input_ids: torch.Tensor,
+        pixel_values: Optional[torch.Tensor],
+        grid_thws: Optional[torch.Tensor],
+    ):
+        placeholder_token_mask = torch.lt(input_ids, 0)
+        multimodal_embeds = self.get_wte()(torch.masked_fill(input_ids, placeholder_token_mask, 0))
+
+        if pixel_values is not None:
+            visual_indicator_embeds = self.vte(self.indicator_token_indices).to(
+                dtype=multimodal_embeds.dtype, device=multimodal_embeds.device
+            )
+            visual_tokens = self.visual_tokenizer(pixel_values, grid_thws)
+            visual_embeds = self.vte(visual_tokens).to(dtype=multimodal_embeds.dtype, device=multimodal_embeds.device)
+
+            for i, indicator_id in enumerate(INDICATOR_IDS):
+                multimodal_embeds[input_ids == indicator_id] = visual_indicator_embeds[i]
+            multimodal_embeds[input_ids == VISUAL_ATOM_ID] = visual_embeds
+
+        return multimodal_embeds
+
+    def _merge_inputs(
+        self, raw_input_ids, placeholder_id, grid_thws, indicator_begin_id, indicator_end_id
+    ):
+        input_ids = []
+        prev_index = 0
+        placeholder_indexes = [i for i, v in enumerate(raw_input_ids) if v == placeholder_id]
+        for placeholder_index, grid_thw in zip(placeholder_indexes, grid_thws):
+            input_ids.extend(raw_input_ids[prev_index:placeholder_index])
+            num_image_atoms = grid_thw.prod().item()
+            num_image_atoms //= self.visual_tokenizer.vit.config.hidden_stride ** 2
+            num_image_atoms //= self.visual_tokenizer.vit.config.temporal_patch_size
+            input_ids.extend([indicator_begin_id] + [VISUAL_ATOM_ID] * num_image_atoms + [indicator_end_id])
+            prev_index = placeholder_index + 1
+        input_ids.extend(raw_input_ids[prev_index:])
+        return input_ids
+
+    def _tokenize_with_visual_placeholder(self, text):
+        placeholder = VIDEO_PLACEHOLDER if VIDEO_PLACEHOLDER in text else IMAGE_PLACEHOLDER
+        placeholder_id = VIDEO_PLACEHOLDER_ID if VIDEO_PLACEHOLDER in text else IMAGE_PLACEHOLDER_ID
+        chunks = [self.text_tokenizer(chunk, add_special_tokens=False).input_ids for chunk in text.split(placeholder)]
+        input_ids = chunks[0]
+        for chunk in chunks[1:]:
+            input_ids.append(placeholder_id)
+            input_ids.extend(chunk)
+        return input_ids
+
+    def preprocess_inputs(
+        self,
+        messages: List[Union[str, Dict]],
+        min_pixels=448 * 448,
+        max_pixels=1792 * 1792,
+        add_generation_prompt=True,
+        enable_thinking=False
+    ):
+        text = self.text_tokenizer.apply_chat_template(
+            messages,
+            tokenize=False,
+            add_generation_prompt=add_generation_prompt,
+            enable_thinking=enable_thinking
+        )
+        input_ids = self._tokenize_with_visual_placeholder(text)
+        images = []
+        videos = []
+        for message in messages:
+            content = message["content"]
+            if isinstance(content, list):
+                images.extend([item["image"] for item in content if item.get("image") is not None])
+                videos.extend([item["video"] for item in content if item.get("video") is not None])
+        if images and videos:
+            raise ValueError(
+                "Multiple visual input data types detected (both image and video provided). "
+                "This model supports only one type of visual input data at a time. "
+                "Please provide either image or video, but not both."
+            )
+
+        pixel_values, grid_thws = None, None
+        if images:
+            pixel_values, grid_thws = zip(
+                *(self.visual_tokenizer.preprocess(image=image, min_pixels=min_pixels, max_pixels=max_pixels)
+                  for image in images)
+            )
+            input_ids = self._merge_inputs(
+                input_ids, IMAGE_PLACEHOLDER_ID, grid_thws, INDICATOR_IDS[0], INDICATOR_IDS[1]
+            )
+            pixel_values = torch.cat(pixel_values, dim=0)
+            grid_thws = torch.cat(grid_thws, dim=0)
+        elif videos:
+            assert len(videos) == 1, "only support single video"
+            pixel_values, grid_thws = self.visual_tokenizer.preprocess(
+                video=videos[0], min_pixels=min_pixels, max_pixels=max_pixels
+            )
+            input_ids = self._merge_inputs(
+                input_ids, VIDEO_PLACEHOLDER_ID, grid_thws, INDICATOR_IDS[2], INDICATOR_IDS[3]
+            )
+
+        input_ids = torch.tensor(input_ids, dtype=torch.long).unsqueeze(0)
+
+        return input_ids, pixel_values, grid_thws
+
+    def generate(
+        self,
+        inputs: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> Union[GenerateOutput, torch.LongTensor]:
+        attention_mask = torch.ne(inputs, self.text_tokenizer.pad_token_id).to(device=inputs.device)
+        inputs_embeds = self.merge_multimodal(
+            input_ids=inputs,
+            pixel_values=kwargs.pop('pixel_values', None),
+            grid_thws=kwargs.pop('grid_thws', None)
+        )
+        enable_thinking = kwargs.pop('enable_thinking', False)
+        enable_thinking_budget = kwargs.pop('enable_thinking_budget', False)
+        thinking_budget = kwargs.pop('thinking_budget', 1024)
+        
+        if enable_thinking and enable_thinking_budget:
+            actual_max_new_tokens = kwargs['max_new_tokens']
+            kwargs['max_new_tokens'] = thinking_budget
+            generated_ids = self.llm.generate(inputs=None, inputs_embeds=inputs_embeds, attention_mask=attention_mask, **kwargs)
+            output_ids = generated_ids
+            output_ids_list = generated_ids[0]
+            
+            # check if the generation has already finished (151645 is <|im_end|>)
+            if 151645 not in output_ids_list:
+                # check if the thinking process has finished (151668 is </think>)
+                # and prepare the second model input
+                if 151668 not in output_ids_list:
+                    early_stopping_text = "\n\nConsidering the limited time by the user, I have to give the solution based on the thinking directly now.\n</think>\n\n"
+                    early_stopping_ids = self.text_tokenizer(early_stopping_text, return_tensors="pt", return_attention_mask=False).input_ids.to(inputs.device)
+                    input_ids_appendent = torch.cat([output_ids, early_stopping_ids], dim=-1)
+                    kwargs['streamer'].put(early_stopping_ids) if 'streamer' in kwargs else None
+                else:
+                    input_ids_appendent = output_ids
+                
+
+                # second generation
+                new_inputs = torch.cat([inputs, input_ids_appendent], dim=-1)
+                attention_mask = torch.ne(new_inputs, self.text_tokenizer.pad_token_id).to(device=inputs.device)
+                inputs_embeds_appendent = self.merge_multimodal(
+                    input_ids=input_ids_appendent,
+                    pixel_values=None,
+                    grid_thws=None
+                )
+                new_inputs_embeds = torch.cat([inputs_embeds, inputs_embeds_appendent], dim=-2)
+                
+                kwargs['max_new_tokens'] = inputs_embeds.size(-2) + actual_max_new_tokens - new_inputs_embeds.size(-2)
+                generated_ids2 = self.llm.generate(inputs=None, inputs_embeds=new_inputs_embeds, attention_mask=attention_mask, **kwargs)
+                kwargs['streamer'].manual_end() if 'streamer' in kwargs else None
+                return torch.cat([input_ids_appendent, generated_ids2], dim=-1)
+            
+            else:
+                kwargs['streamer'].manual_end() if 'streamer' in kwargs else None
+                return generated_ids
+            
+        else:
+            generated_ids = self.llm.generate(inputs=None, inputs_embeds=inputs_embeds, attention_mask=attention_mask, **kwargs)
+            kwargs['streamer'].manual_end() if 'streamer' in kwargs else None
+            return generated_ids
+        
+AutoConfig.register('siglip2_navit', Siglip2NavitConfig)
+AutoModel.register(Siglip2NavitConfig, Siglip2NavitModel)
+
+AutoConfig.register("ovis2_6", Ovis2_6_Config)
+AutoModelForCausalLM.register(Ovis2_6_Config, Ovis2_6ForCausalLM)
+
+AutoConfig.register("ovis2_6_moe", Ovis2_6_Moe_Config)
+AutoModelForCausalLM.register(Ovis2_6_Moe_Config, Ovis2_6_MoeForCausalLM)
+
+AutoConfig.register("ovis2_6_next", Ovis2_6_Next_Config)
+AutoModelForCausalLM.register(Ovis2_6_Next_Config, Ovis2_6_NextForCausalLM)

preprocessor_config.json

@@ -0,0 +1,24 @@
+{
+  "do_convert_rgb": null,
+  "do_normalize": true,
+  "do_rescale": true,
+  "do_resize": true,
+  "image_mean": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "image_processor_type": "SiglipImageProcessor",
+  "image_std": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "processor_class": "SiglipProcessor",
+  "resample": 2,
+  "rescale_factor": 0.00392156862745098,
+  "size": {
+    "height": 512,
+    "width": 512
+  }
+}

special_tokens_map.json

@@ -0,0 +1,31 @@
+{
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4
+size 11422654

tokenizer_config.json

@@ -0,0 +1,247 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151665": {
+      "content": "<tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151666": {
+      "content": "</tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151667": {
+      "content": "<think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151668": {
+      "content": "</think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>",
+    "<image>",
+    "<video>",
+    "<ovis_visual_atom>",
+    "<ovis_image_start>",
+    "<ovis_image_end>",
+    "<ovis_video_start>",
+    "<ovis_video_end>"
+  ],
+  "bos_token": null,
+  "chat_template": "{%- for message in messages %}\n{{- '<|im_start|>' + message.role + '\\n' }}\n{%- if message.role == 'system' or message.role == 'user' %}\n{%- if message.content is string %}\n{{- message.content }}\n{%- else %}\n{%- for item in message.content %}\n{%- if item.type == 'text' and 'text' in item %}\n{{- item.text }}\n{%- elif item.type == 'image' %}\n{{- '<image>' }}\n{%- elif item.type == 'video' %}\n{{- '<video>' }}\n{%- else %}\n{{- raise_exception('Invalid content type. Supported types for system and user are text, image, video.') }}\n{%- endif %}\n{%- if not loop.last %}{{- '\\n' }}{%- endif %}\n{%- endfor %}\n{%- endif %}\n{%- elif message.role == 'assistant' %}\n{%- set content = '' %}\n{%- if message.content is string %}\n{%- set content = message.content %}\n{%- else %}\n{%- set ns = namespace(content='') -%}\n{%- for item in message.content %}\n{%- if item.type == 'text' and 'text' in item %}\n{%- set ns.content = ns.content ~ item.text %}\n{%- else %}\n{{- raise_exception('Invalid content type. Supported type for assistant is text.') }}\n{%- endif %}\n{%- endfor %}\n{%- set content = ns.content -%}\n{%- endif %}\n{%- set content = content.split('</think>')[-1].lstrip('\\n') %}\n{{- content }}\n{%- else %}\n{{- raise_exception('Invalid role. Supported roles are system, user, assistant.') }}\n{%- endif %}\n{{- '<|im_end|>\\n' }}\n{%- endfor %}\n{%- if add_generation_prompt %}\n{{- '<|im_start|>assistant\\n' }}\n{%- if enable_thinking is defined and enable_thinking is false %}\n{{- '<think>\\n\\n</think>\\n\\n' }}\n{%- endif %}\n{%- endif %}\n",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}