Upload folder using huggingface_hub

.mv

@@ -0,0 +1 @@
+Revision:master,CreatedAt:1719926951

LICENSE

@@ -0,0 +1,75 @@
+The CogVLM License
+
+1. Definitions
+
+“Licensor” means the CogVLM Model Team that distributes its Software.
+
+“Software” means the CogVLM model parameters made available under this license.
+
+2. License Grant
+
+Under the terms and conditions of this license, the Licensor hereby grants you a non-exclusive, worldwide, non-transferable, non-sublicensable, revocable, royalty-free copyright license.
+This license permits you to use all open-source models in this repository for academic research free. Users who wish to use the models for commercial purposes must register [here](https://open.bigmodel.cn/mla/form).
+Registered users may use the models for commercial activities free of charge, but must comply with all terms and conditions of this license.
+The license notice shall be included in all copies or substantial portions of the Software.
+
+3. Restriction
+
+You will not use, copy, modify, merge, publish, distribute, reproduce, or create derivative works of the Software, in whole or in part, for any military, or illegal purposes.
+
+You will not use the Software for any act that may undermine China's national security and national unity, harm the public interest of society, or infringe upon the rights and interests of human beings.
+
+4. Disclaimer
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+5. Limitation of Liability
+
+EXCEPT TO THE EXTENT PROHIBITED BY APPLICABLE LAW, IN NO EVENT AND UNDER NO LEGAL THEORY, WHETHER BASED IN TORT, NEGLIGENCE, CONTRACT, LIABILITY, OR OTHERWISE WILL ANY LICENSOR BE LIABLE TO YOU FOR ANY DIRECT, INDIRECT, SPECIAL, INCIDENTAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES, OR ANY OTHER COMMERCIAL LOSSES, EVEN IF THE LICENSOR HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGES.
+
+6. Dispute Resolution
+
+This license shall be governed and construed in accordance with the laws of People’s Republic of China. Any dispute arising from or in connection with this License shall be submitted to Haidian District People's Court in Beijing.
+
+Note that the license is subject to update to a more comprehensive version.  For any questions related to the license and copyright, please contact us at [email protected].
+
+7. Llama3 and EVA-CLIP2 License
+
+For the CogVLM2 open source model based on the LLama3 series model as the base model, the Llama3 license conditions (https://llama.meta.com/llama3/license/, a copy of this repository license conditions) and the EVA-CLIP2 license conditions (MIT , https://github.com/baaivision/EVA/blob/master/LICENSE) for model weights.
+
+1. 定义
+
+“许可方”是指分发其软件的 CogVLM 模型团队。
+
+“软件”是指根据本许可提供的 CogVLM 模型参数。
+
+2. 许可授予
+
+根据本许可的条款和条件，许可方特此授予您非排他性、全球性、不可转让、不可再许可、可撤销、免版税的版权许可。
+本许可允许您免费使用本仓库中的所有开源模型进行学术研究，对于希望将模型用于商业目的的用户，需在[这里](https://open.bigmodel.cn/mla/form)完成登记。
+经过登记的用户可以免费使用本模型进行商业活动，但必须遵守本许可的所有条款和条件。
+上述版权声明和本许可声明应包含在本软件的所有副本或重要部分中。
+
+3.限制
+
+您不得出于任何军事或非法目的使用、复制、修改、合并、发布、分发、复制或创建本软件的全部或部分衍生作品。
+
+您不得利用本软件从事任何危害国家安全和国家统一、危害社会公共利益、侵犯人身权益的行为。
+
+4.免责声明
+
+本软件“按原样”提供，不提供任何明示或暗示的保证，包括但不限于对适销性、特定用途的适用性和非侵权性的保证。 在任何情况下，作者或版权持有人均不对任何索赔、损害或其他责任负责，无论是在合同诉讼、侵权行为还是其他方面，由软件或软件的使用或其他交易引起、由软件引起或与之相关 软件。
+
+5. 责任限制
+
+除适用法律禁止的范围外，在任何情况下且根据任何法律理论，无论是基于侵权行为、疏忽、合同、责任或其他原因，任何许可方均不对您承担任何直接、间接、特殊、偶然、示范性、 或间接损害，或任何其他商业损失，即使许可人已被告知此类损害的可能性。
+
+6.争议解决
+
+本许可受中华人民共和国法律管辖并按其解释。 因本许可引起的或与本许可有关的任何争议应提交北京市海淀区人民法院。
+
+请注意，许可证可能会更新到更全面的版本。 有关许可和版权的任何问题，请���过 [email protected] 与我们联系。
+
+7. Llama3 和 EVA-CLIP2 许可
+
+针对基于以 LLama3 系列模型作为基座模型的 CogVLM2 开源模型， Llama3 许可条件 (https://llama.meta.com/llama3/license/ ，本仓库副本一份许可条件) 和 EVA-CLIP2 许可条件 (MIT, https://github.com/baaivision/EVA/blob/master/LICENSE) 适用于模型权重。

LLAMA3_LICENSE

@@ -0,0 +1,117 @@
+META LLAMA 3 COMMUNITY LICENSE AGREEMENT
+Meta Llama 3 Version Release Date: April 18, 2024
+
+“Agreement” means the terms and conditions for use, reproduction, distribution and modification of the
+Llama Materials set forth herein.
+
+“Documentation” means the specifications, manuals and documentation accompanying Meta Llama 3
+distributed by Meta at https://llama.meta.com/get-started/.
+
+“Licensee” or “you” means you, or your employer or any other person or entity (if you are entering into
+this Agreement on such person or entity’s behalf), of the age required under applicable laws, rules or
+regulations to provide legal consent and that has legal authority to bind your employer or such other
+person or entity if you are entering in this Agreement on their behalf.
+
+“Meta Llama 3” means the foundational large language models and software and algorithms, including
+machine-learning model code, trained model weights, inference-enabling code, training-enabling code,
+fine-tuning enabling code and other elements of the foregoing distributed by Meta at
+https://llama.meta.com/llama-downloads.
+
+“Llama Materials” means, collectively, Meta’s proprietary Meta Llama 3 and Documentation (and any
+portion thereof) made available under this Agreement.
+
+“Meta” or “we” means Meta Platforms Ireland Limited (if you are located in or, if you are an entity, your
+principal place of business is in the EEA or Switzerland) and Meta Platforms, Inc. (if you are located
+outside of the EEA or Switzerland).
+
+By clicking “I Accept” below or by using or distributing any portion or element of the Llama Materials,
+you agree to be bound by this Agreement.
+
+1. License Rights and Redistribution.
+
+  a. Grant of Rights. You are granted a non-exclusive, worldwide, non-transferable and royalty-free
+limited license under Meta’s intellectual property or other rights owned by Meta embodied in the Llama
+Materials to use, reproduce, distribute, copy, create derivative works of, and make modifications to the
+Llama Materials.
+
+  b. Redistribution and Use.
+
+      i. If you distribute or make available the Llama Materials (or any derivative works
+thereof), or a product or service that uses any of them, including another AI model, you shall (A) provide
+a copy of this Agreement with any such Llama Materials; and (B) prominently display “Built with Meta
+Llama 3” on a related website, user interface, blogpost, about page, or product documentation. If you
+use the Llama Materials to create, train, fine tune, or otherwise improve an AI model, which is
+distributed or made available, you shall also include “Llama 3” at the beginning of any such AI model
+name.
+
+      ii. If you receive Llama Materials, or any derivative works thereof, from a Licensee as part 
+of an integrated end user product, then Section 2 of this Agreement will not apply to you.
+
+      iii. You must retain in all copies of the Llama Materials that you distribute the following
+attribution notice within a “Notice” text file distributed as a part of such copies: “Meta Llama 3 is
+licensed under the Meta Llama 3 Community License, Copyright © Meta Platforms, Inc. All Rights
+Reserved.”
+
+      iv. Your use of the Llama Materials must comply with applicable laws and regulations
+(including trade compliance laws and regulations) and adhere to the Acceptable Use Policy for the Llama
+Materials (available at https://llama.meta.com/llama3/use-policy), which is hereby incorporated by
+reference into this Agreement.
+
+      v. You will not use the Llama Materials or any output or results of the Llama Materials to
+improve any other large language model (excluding Meta Llama 3 or derivative works thereof).
+
+2. Additional Commercial Terms. If, on the Meta Llama 3 version release date, the monthly active users
+of the products or services made available by or for Licensee, or Licensee’s affiliates, is greater than 700
+million monthly active users in the preceding calendar month, you must request a license from Meta,
+which Meta may grant to you in its sole discretion, and you are not authorized to exercise any of the
+rights under this Agreement unless or until Meta otherwise expressly grants you such rights.
+
+3. Disclaimer of Warranty. UNLESS REQUIRED BY APPLICABLE LAW, THE LLAMA MATERIALS AND ANY
+OUTPUT AND RESULTS THEREFROM ARE PROVIDED ON AN “AS IS” BASIS, WITHOUT WARRANTIES OF
+ANY KIND, AND META DISCLAIMS ALL WARRANTIES OF ANY KIND, BOTH EXPRESS AND IMPLIED,
+INCLUDING, WITHOUT LIMITATION, ANY WARRANTIES OF TITLE, NON-INFRINGEMENT,
+MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. YOU ARE SOLELY RESPONSIBLE FOR
+DETERMINING THE APPROPRIATENESS OF USING OR REDISTRIBUTING THE LLAMA MATERIALS AND
+ASSUME ANY RISKS ASSOCIATED WITH YOUR USE OF THE LLAMA MATERIALS AND ANY OUTPUT AND
+RESULTS.
+
+4. Limitation of Liability. IN NO EVENT WILL META OR ITS AFFILIATES BE LIABLE UNDER ANY THEORY OF
+LIABILITY, WHETHER IN CONTRACT, TORT, NEGLIGENCE, PRODUCTS LIABILITY, OR OTHERWISE, ARISING
+OUT OF THIS AGREEMENT, FOR ANY LOST PROFITS OR ANY INDIRECT, SPECIAL, CONSEQUENTIAL,
+INCIDENTAL, EXEMPLARY OR PUNITIVE DAMAGES, EVEN IF META OR ITS AFFILIATES HAVE BEEN ADVISED
+OF THE POSSIBILITY OF ANY OF THE FOREGOING.
+
+5. Intellectual Property.
+
+  a. No trademark licenses are granted under this Agreement, and in connection with the Llama
+Materials, neither Meta nor Licensee may use any name or mark owned by or associated with the other
+or any of its affiliates, except as required for reasonable and customary use in describing and
+redistributing the Llama Materials or as set forth in this Section 5(a). Meta hereby grants you a license to
+use “Llama 3” (the “Mark”) solely as required to comply with the last sentence of Section 1.b.i. You will
+comply with Meta’s brand guidelines (currently accessible at
+https://about.meta.com/brand/resources/meta/company-brand/ ). All goodwill arising out of your use
+of the Mark will inure to the benefit of Meta.
+
+  b. Subject to Meta’s ownership of Llama Materials and derivatives made by or for Meta, with
+respect to any derivative works and modifications of the Llama Materials that are made by you, as
+between you and Meta, you are and will be the owner of such derivative works and modifications.
+
+  c. If you institute litigation or other proceedings against Meta or any entity (including a
+cross-claim or counterclaim in a lawsuit) alleging that the Llama Materials or Meta Llama 3 outputs or
+results, or any portion of any of the foregoing, constitutes infringement of intellectual property or other
+rights owned or licensable by you, then any licenses granted to you under this Agreement shall
+terminate as of the date such litigation or claim is filed or instituted. You will indemnify and hold
+harmless Meta from and against any claim by any third party arising out of or related to your use or
+distribution of the Llama Materials.
+
+6. Term and Termination. The term of this Agreement will commence upon your acceptance of this
+Agreement or access to the Llama Materials and will continue in full force and effect until terminated in
+accordance with the terms and conditions herein. Meta may terminate this Agreement if you are in
+breach of any term or condition of this Agreement. Upon termination of this Agreement, you shall delete
+and cease use of the Llama Materials. Sections 3, 4 and 7 shall survive the termination of this
+Agreement.
+
+7. Governing Law and Jurisdiction. This Agreement will be governed and construed under the laws of
+the State of California without regard to choice of law principles, and the UN Convention on Contracts
+for the International Sale of Goods does not apply to this Agreement. The courts of California shall have
+exclusive jurisdiction of any dispute arising out of this Agreement.

config.json

@@ -0,0 +1,40 @@
+{
+  "architectures": [
+    "CogVLMVideoForCausalLM"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_cogvlm.CogVLMConfig",
+    "AutoModelForCausalLM": "modeling_cogvlm.CogVLMVideoForCausalLM"
+  },
+  "bos_token_id": 128000,
+  "eos_token_id": 128001,
+  "pad_token_id": 128002,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "initializer_range": 0.02,
+  "intermediate_size": 14336,
+  "max_position_embeddings": 2048,
+  "num_attention_heads": 32,
+  "num_hidden_layers": 32,
+  "num_multi_query_heads": 8,
+  "rms_norm_eps": 1e-05,
+  "template_version": "base",
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.41.0",
+  "use_cache": true,
+  "vision_config": {
+    "dropout_prob": 0.0,
+    "hidden_act": "gelu",
+    "hidden_size": 1792,
+    "image_size": 224,
+    "in_channels": 3,
+    "intermediate_size": 15360,
+    "layer_norm_eps": 1e-06,
+    "num_heads": 16,
+    "num_hidden_layers": 63,
+    "num_positions": 257,
+    "patch_size": 14
+  },
+  "vocab_size": 128256
+}

configuration.json

@@ -0,0 +1 @@
+{"framework":"Pytorch","task":"video-question-answering"}

configuration_cogvlm.py

@@ -0,0 +1,46 @@
+from typing import Literal
+from transformers import PretrainedConfig
+
+
+class CogVLMConfig(PretrainedConfig):
+    _auto_class = "AutoConfig"
+
+    def __init__(
+            self,
+            vocab_size=32000,
+            hidden_size=4096,
+            intermediate_size=11008,
+            num_hidden_layers=32,
+            num_attention_heads=32,
+            num_multi_query_heads=8,
+            hidden_act='silu',
+            max_position_embeddings=2048,
+            initializer_range=0.02,
+            rms_norm_eps=1e-06,
+            template_version: Literal["base", "chat"] = "chat",
+            pad_token_id=128002,
+            bos_token_id=128001,
+            eos_token_id=128002,
+            tie_word_embeddings=False,
+            use_cache=True,
+            **kwargs,
+    ):
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_attention_heads = num_attention_heads
+        self.num_multi_query_heads = num_multi_query_heads
+        self.max_position_embeddings = max_position_embeddings
+        self.rms_norm_eps = rms_norm_eps
+        self.initializer_range = initializer_range
+        self.vocab_size = vocab_size
+        self.num_hidden_layers = num_hidden_layers
+        self.hidden_act = hidden_act
+        self.template_version = template_version
+        self.use_cache = use_cache
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

generation_config.json

@@ -0,0 +1,11 @@
+{
+  "bos_token_id": 128000,
+  "eos_token_id": 128001,
+  "pad_token_id": 128002,
+  "do_sample": true,
+  "temperature": 0.1,
+  "max_length": 2048,
+  "top_p": 0.1,
+  "top_k": 1,
+  "transformers_version": "4.41.0"
+}

model-00001-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f53e34de951ee7205398dc1e03a8a0b797ddc4b7a6714181adb51b2a906a94c5
+size 4976699712

model-00002-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:555733b658ef1733b5afa979a6c999d2f3310ac8e2f851eb6c53b9b73a10c4a9
+size 4999803504

model-00003-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d90e87d1db7f29deb7a59817476da9bba04339cd3c7029d3fc4ca97c6aa4b776
+size 4915917160

model-00004-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf5581fa2fb9be9bf89ded6c82607fec36c809e200b00ce10893e80501036a25
+size 4956242104

model-00005-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:40c0ca93426ca37f04b66d9e1e6cb9687e0bcbaa02b4827f282b564577b2b6da
+size 4115863248

model-00006-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:73b20e2d92146829390d468bb37bb9bacf23df3e22f6000754cd94c1dfc83f23
+size 1050673280

modeling_cogvlm.py

@@ -0,0 +1,898 @@
+"""largely copy from llama and adapt for cogvlm"""
+import warnings
+from typing import TYPE_CHECKING, Optional, Tuple, List, Union, Literal, Dict, Any
+
+import math
+import torch
+from torch import nn
+from torch.nn import CrossEntropyLoss
+from torchvision import transforms
+from einops import rearrange
+
+from decord import VideoReader, cpu
+import decord
+import io
+import numpy as np
+
+from transformers import PreTrainedModel, PreTrainedTokenizer
+from transformers.utils.logging import get_logger
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from torchvision.transforms.functional import InterpolationMode
+from torchvision.transforms import Lambda
+from torchvision.transforms._transforms_video import NormalizeVideo, RandomCropVideo, RandomHorizontalFlipVideo, CenterCropVideo
+from pytorchvideo.transforms import ApplyTransformToKey, ShortSideScale
+from .configuration_cogvlm import CogVLMConfig
+from .util import FastRotaryEmbedding
+from .visual import EVA2CLIPModel
+
+
+
+if TYPE_CHECKING:
+    from transformers.utils import ModelOutput
+
+logger = get_logger(__name__)
+
+LANGUAGE_TOKEN_TYPE = 0
+VISION_TOKEN_TYPE = 1
+
+
+# Copied from transformers.models.bart.modeling_bart._make_causal_mask
+def _make_causal_mask(
+        input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
+):
+    """
+    Make causal mask used for bi-directional self-attention.
+    """
+    bsz, tgt_len = input_ids_shape
+    mask = torch.full((tgt_len, tgt_len), torch.finfo(dtype).min, device=device)
+    mask_cond = torch.arange(mask.size(-1), device=device)
+    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
+    mask = mask.to(dtype)
+
+    if past_key_values_length > 0:
+        mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
+    return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
+
+
+# Copied from transformers.models.bart.modeling_bart._expand_mask
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    """
+    bsz, src_len = mask.size()
+    tgt_len = tgt_len if tgt_len is not None else src_len
+
+    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
+
+    inverted_mask = 1.0 - expanded_mask
+
+    return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
+
+
+class RMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-5):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return (self.weight * hidden_states).to(input_dtype)
+
+
+class MLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+
+def get_expert_mask(token_type_ids: "torch.LongTensor(B, L)") -> "[torch.BoolTensor(B, L), torch.BoolTensor(B, L)]":
+    vision_token_mask = torch.zeros_like(token_type_ids, dtype=torch.bool)
+    vision_token_mask[:, :-1] = (token_type_ids[:, :-1] == VISION_TOKEN_TYPE) & (token_type_ids[:, 1:] == VISION_TOKEN_TYPE)
+    language_token_mask = ~vision_token_mask
+    return vision_token_mask, language_token_mask
+
+
+class VisionExpertMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.language_mlp = MLP(config)
+        # self.vision_mlp = MLP(config)
+
+    def forward(self, hidden_states: "torch.Tensor(B, L, D)", token_type_ids: "torch.LongTensor(B, L)"):
+        # output = torch.empty(hidden_states.shape, dtype=hidden_states.dtype, device=hidden_states.device)
+        # vision_token_mask, language_token_mask = get_expert_mask(token_type_ids)
+        # output[vision_token_mask] = self.vision_mlp(hidden_states[vision_token_mask])
+        # output[language_token_mask] = self.language_mlp(hidden_states[language_token_mask])
+        
+        output = self.language_mlp(hidden_states)
+        return output
+
+
+def attention_fn(
+        query_layer: "torch.tensor(B, H, L, HD)",
+        key_layer: "torch.tensor(B, H, L, HD)",
+        value_layer: "torch.tensor(B, H, L, HD)",
+        attention_mask: "torch.tensor(B, H, L, HD)",
+        *,
+        scaling_attention_score: bool = True,
+        attention_dropout: nn.Module = None
+):
+    attention_mask_bool = (attention_mask == 0)
+    is_low_triangle = (attention_mask_bool == torch.ones_like(attention_mask_bool, dtype=torch.float).tril()).all()
+    is_full = (attention_mask_bool > 0).all()
+    if not (int(torch.__version__.split('.')[0]) >= 2):
+        warnings.warn("It's recommended to use torch2.0 or higher.")
+    if int(torch.__version__.split('.')[0]) >= 2 and scaling_attention_score and (is_full or is_low_triangle):
+        dropout_p = 0. if attention_dropout is None or not attention_dropout.training else attention_dropout.p
+        return torch.nn.functional.scaled_dot_product_attention(
+            query_layer, key_layer, value_layer,
+            attn_mask=None,
+            dropout_p=dropout_p,
+            is_causal=not is_full
+        )
+    else:
+        if scaling_attention_score:
+            query_layer = query_layer / math.sqrt(query_layer.shape[-1])
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+        attention_scores = attention_scores + attention_mask
+        attention_scores = nn.functional.softmax(attention_scores, dim=-1, dtype=torch.float32).to(query_layer.dtype)
+        if attention_dropout is not None:
+            attention_scores = attention_dropout(attention_scores)
+        context_layer = torch.matmul(attention_scores, value_layer)
+        return context_layer
+
+
+class VisionExpertAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.num_attention_heads = config.num_attention_heads
+        self.num_multi_query_heads = config.num_multi_query_heads
+        self.hidden_size_per_attention_head = self.hidden_size // self.num_attention_heads
+        self.stride = [self.num_attention_heads, self.num_multi_query_heads, self.num_multi_query_heads]
+        self.qkv_size = self.hidden_size + self.hidden_size_per_attention_head * self.num_multi_query_heads * 2
+        self.head_dim = self.hidden_size // self.num_attention_heads
+        self.max_position_embeddings = config.max_position_embeddings
+        self.rotary_emb = FastRotaryEmbedding(dim=self.head_dim, pos_idx_in_fp32=False, base=500000)
+        # self.vision_expert_query_key_value = nn.Linear(self.hidden_size, self.qkv_size, bias=True)
+        # self.vision_expert_dense = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
+        self.language_expert_query_key_value = nn.Linear(self.hidden_size, self.qkv_size, bias=False)
+        self.language_expert_dense = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
+
+    def _transpose_for_scores(self, tensor):
+        """Transpose a 3D tensor [B, L, H*HD] into a 4D tensor with size [B H L HD]."""
+        new_tensor_shape = tensor.size()[:-1] + \
+                           (-1, # flexible for multi-query
+                            self.hidden_size_per_attention_head)
+        tensor = tensor.view(*new_tensor_shape)
+        return tensor.permute(0, 2, 1, 3)
+
+    def forward(
+            self,
+            hidden_states: torch.Tensor,
+            token_type_ids: torch.LongTensor,
+            position_ids: torch.LongTensor,
+            attention_mask: Optional[torch.Tensor] = None,
+            past_key_value: Optional[Tuple[torch.Tensor]] = None,
+            output_attentions: bool = False,
+            use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+        # vision_token_mask, language_token_mask = get_expert_mask(token_type_ids)
+
+        shape = list(hidden_states.shape)
+        shape[-1] = self.qkv_size
+        # mixed_raw_layer = torch.empty(shape, dtype=hidden_states.dtype, device=hidden_states.device)
+        # mixed_raw_layer[vision_token_mask] = self.vision_expert_query_key_value(hidden_states[vision_token_mask])
+        # mixed_raw_layer[language_token_mask] = self.language_expert_query_key_value(hidden_states[language_token_mask])
+        mixed_raw_layer = self.language_expert_query_key_value(hidden_states)
+
+        # query_states, key_states, value_states = torch.split(mixed_raw_layer, self.hidden_size, dim=-1)
+        factor = mixed_raw_layer.size()[-1] // sum(self.stride)
+        query_states, key_states, value_states = torch.split(mixed_raw_layer, [factor * x for x in self.stride], dim=-1)
+
+        query_states = self._transpose_for_scores(query_states)  # B, H, L, HD
+        key_states = self._transpose_for_scores(key_states)  # B, H, L, HD
+        value_states = self._transpose_for_scores(value_states)  # B, H, L, HD
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value[0].shape[-2]
+
+        query_states, key_states = self.rotary_emb(query_states, key_states, position_ids=position_ids, max_seqlen=position_ids.max() + 1)
+
+        if past_key_value is not None:
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+
+        past_key_value = (key_states, value_states) if use_cache else None
+
+        key_states = key_states.unsqueeze(2).expand(-1, -1, self.num_attention_heads // self.num_multi_query_heads, -1, -1).contiguous().view(
+            bsz, self.num_attention_heads, *key_states.shape[2:])
+        value_states = value_states.unsqueeze(2).expand(-1, -1, self.num_attention_heads // self.num_multi_query_heads, -1,
+                                                        -1).contiguous().view(bsz, self.num_attention_heads, *value_states.shape[2:])
+
+        context_layer = attention_fn(
+            query_layer=query_states, key_layer=key_states, value_layer=value_states, attention_mask=attention_mask,
+            scaling_attention_score=True, attention_dropout=None)
+        if context_layer.size() != (bsz, self.num_attention_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_attention_heads, q_len, self.head_dim)}, but is"
+                f" {context_layer.size()}"
+            )
+        context_layer = context_layer.transpose(1, 2).contiguous().reshape(bsz, q_len, self.hidden_size)
+
+        # attn_output = torch.empty(context_layer.shape, dtype=hidden_states.dtype, device=hidden_states.device)
+        # attn_output[vision_token_mask] = self.vision_expert_dense(context_layer[vision_token_mask])
+        # attn_output[language_token_mask] = self.language_expert_dense(context_layer[language_token_mask])
+        
+        attn_output = self.language_expert_dense(context_layer)
+
+        if output_attentions:
+            warnings.warn("output_attentions is not implemented.")
+
+        return attn_output, None, past_key_value
+
+
+class CogVLMDecoderLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = VisionExpertAttention(config=config)
+        self.mlp = VisionExpertMLP(config)
+        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+            self,
+            hidden_states: torch.Tensor,
+            token_type_ids: torch.LongTensor,
+            position_ids: torch.LongTensor,
+            attention_mask: Optional[torch.Tensor] = None,
+            past_key_value: Optional[Tuple[torch.Tensor]] = None,
+            output_attentions: Optional[bool] = False,
+            use_cache: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            attention_mask=attention_mask,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states, token_type_ids=token_type_ids)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs  # type: ignore
+
+
+class CogVLMPreTrainedModel(PreTrainedModel):
+    config_class = CogVLMConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = False
+    _no_split_modules = ["CogVLMDecoderLayer"]
+    _skip_keys_device_placement = "past_key_values"
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+
+def is_empty(images_list: Optional[List[List[torch.Tensor]]]):
+    if images_list is None or len(images_list) == 0:
+        return True
+    for image_list in images_list:
+        if len(image_list):
+            return False
+    return True
+
+
+def build_position_ids(x: "torch.BoolTensor(B, L)", attention_mask: Optional["torch.BoolTensor(B, L)"] = None) -> "torch.LongTensor(B, L)":
+    if attention_mask is not None:
+        tmp = x.clone()
+        tmp[~(attention_mask.bool())] = -1
+    else:
+        tmp = x.clone()
+    # image boi eoi token as LANGUAGE_TOKEN_TYPE
+    is_boi_eoi = torch.zeros_like(x, dtype=torch.bool)
+    is_boi_eoi[:, 1:] |= (tmp[:, 1:] == VISION_TOKEN_TYPE) & (tmp[:, :-1] == LANGUAGE_TOKEN_TYPE)
+    is_boi_eoi[:, 0] |= (tmp[:, 0] == VISION_TOKEN_TYPE)
+    is_boi_eoi[:, :-1] |= (tmp[:, :-1] == VISION_TOKEN_TYPE) & (tmp[:, 1:] == LANGUAGE_TOKEN_TYPE)
+    is_boi_eoi[:, -1] |= (tmp[:, -1] == VISION_TOKEN_TYPE)
+    tmp[is_boi_eoi] = LANGUAGE_TOKEN_TYPE
+    # final position ids
+    y = torch.zeros_like(x, dtype=torch.long)
+    y[:, 1:] = (tmp[:, 1:] == LANGUAGE_TOKEN_TYPE) | ((tmp[:, 1:] == VISION_TOKEN_TYPE) & (tmp[:, :-1] == LANGUAGE_TOKEN_TYPE))
+    y = y.cumsum(dim=-1)
+    return y
+
+
+class CogVLMVideoModel(CogVLMPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.padding_idx = 128002
+        self.vocab_size = config.vocab_size
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList([CogVLMDecoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        self.vision = EVA2CLIPModel(config)
+
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def encode_images(self, images: List[List[torch.Tensor]], ) -> torch.Tensor:
+        images_list, images = images, []
+
+        images = []
+        for image_list in images_list:
+            for image in image_list:
+                images.append(image)
+
+        # images = torch.stack(images) # video images is already stacked
+        images_features = self.vision(images[0])
+        return images_features
+
+    def forward(
+            self,
+            input_ids: torch.LongTensor = None,
+            images: List[List[torch.Tensor]] = None,
+            token_type_ids: Optional[torch.LongTensor] = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            position_ids: Optional[torch.LongTensor] = None,
+            past_key_values: Optional[List[torch.FloatTensor]] = None,
+            inputs_embeds: Optional[torch.FloatTensor] = None,
+            use_cache: Optional[bool] = None,
+            output_attentions: Optional[bool] = None,
+            output_hidden_states: Optional[bool] = None,
+            return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        """take care of image_encode, token_type_ids, position_ids and (attention_mask = None is fine)"""
+
+        if past_key_values is not None:
+            pass  # generate mode with past_key_values. the image features are already mapped
+        else:
+            # not allow for inputs_embeds, because we want to process image feature
+            assert input_ids is not None and inputs_embeds is None, f"{input_ids} {inputs_embeds}"
+            if not is_empty(images):  # multi-modality
+                assert token_type_ids is not None, f"multi-modality requires `token_type_ids`!"
+                assert len(input_ids) == len(images), f"{len(input_ids)} {len(images)}"
+                inputs_embeds = self.embed_tokens(input_ids)
+                images_features = self.encode_images(images)
+                images_features = rearrange(images_features, 'b n d -> (b n) d')
+                images_features = images_features.to(dtype=inputs_embeds.dtype, device=inputs_embeds.device)
+
+                inputs_embeds = inputs_embeds.index_put([token_type_ids == VISION_TOKEN_TYPE], images_features)
+            else:  # single-modality
+                if token_type_ids is None:
+                    token_type_ids = torch.ones_like(input_ids, dtype=torch.long, device=input_ids.device) * LANGUAGE_TOKEN_TYPE
+                assert not (token_type_ids == VISION_TOKEN_TYPE).any(), f"{(token_type_ids == VISION_TOKEN_TYPE).sum()}"
+                inputs_embeds = self.embed_tokens(input_ids)
+
+            if position_ids is None:
+                position_ids = build_position_ids(token_type_ids, attention_mask)
+            input_ids = None
+        return self.llm_forward(
+            input_ids=input_ids,
+            token_type_ids=token_type_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+    def llm_forward(
+            self,
+            input_ids: torch.LongTensor = None,
+            token_type_ids: torch.LongTensor = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            position_ids: Optional[torch.LongTensor] = None,
+            past_key_values: Optional[List[torch.FloatTensor]] = None,
+            inputs_embeds: Optional[torch.FloatTensor] = None,
+            use_cache: Optional[bool] = None,
+            output_attentions: Optional[bool] = None,
+            output_hidden_states: Optional[bool] = None,
+            return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        """largely copy from llama forward and adapt for cogvlm with `token_type_ids`"""
+        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
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+        elif input_ids is not None:
+            batch_size, seq_length = input_ids.shape
+        elif inputs_embeds is not None:
+            batch_size, seq_length, _ = inputs_embeds.shape
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        seq_length_with_past = seq_length
+        past_key_values_length = 0
+
+        if past_key_values is not None:
+            past_key_values_length = past_key_values[0][0].shape[2]
+            seq_length_with_past = seq_length_with_past + past_key_values_length
+
+        if position_ids is None:
+            device = input_ids.device if input_ids is not None else inputs_embeds.device
+            position_ids = torch.arange(
+                past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
+            )
+            position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
+        else:
+            position_ids = position_ids.view(-1, seq_length).long()
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+        # embed positions
+        if attention_mask is None:
+            attention_mask = torch.ones(
+                (batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device
+            )
+        attention_mask = self._prepare_decoder_attention_mask(
+            attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
+        )
+
+        hidden_states = inputs_embeds
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = () if use_cache else None
+
+        for idx, decoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+            layer_outputs = decoder_layer(
+                hidden_states,
+                token_type_ids=token_type_ids,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_value=past_key_value,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+            )
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    # noinspection PyMethodMayBeStatic
+    # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
+    def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
+        # create causal mask
+        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+        combined_attention_mask = None
+        if input_shape[-1] > 1:
+            combined_attention_mask = _make_causal_mask(
+                input_shape,
+                inputs_embeds.dtype,
+                device=inputs_embeds.device,
+                past_key_values_length=past_key_values_length,
+            )
+
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
+                inputs_embeds.device
+            )
+            combined_attention_mask = (
+                expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
+            )
+
+        return combined_attention_mask
+
+
+def _history_to_prompt(signal_type, history, query):
+    if signal_type == 'base':
+        return query
+    elif signal_type == 'vqa':
+        answer_format = 'Short answer:'
+    elif signal_type == 'chat':
+        answer_format = 'Answer:'
+    else:
+        assert False, f"Unknown signal type {signal_type}"
+
+    prompt = ''
+    for i, (old_query, response) in enumerate(history):
+        prompt += 'Question: ' + old_query + " {} ".format(answer_format) + response + "\n"
+    prompt += 'Question: {} {}'.format(query, answer_format)
+    return prompt
+
+def load_video(video_path):
+    mp4_stream = None
+    decord.bridge.set_bridge('torch')
+    with open(video_path, 'rb') as f:
+        mp4_stream = f.read()
+    clip_end_sec = 60 # clip video to <= 1 minute
+    clip_start_sec = 0
+    num_frames = 24
+    # decord.bridge.set_bridge('torch')
+    if mp4_stream is not None:
+        decord_vr = VideoReader(io.BytesIO(mp4_stream), ctx=cpu(0))
+    else:
+        decord_vr = VideoReader(video_path, ctx=cpu(0))
+    duration = len(decord_vr) # duration in terms of frames
+    start_frame = int(clip_start_sec * decord_vr.get_avg_fps())
+    end_frame = min(duration, int(clip_end_sec*decord_vr.get_avg_fps())) if \
+        clip_end_sec is not None else duration
+    frame_id_list = np.linspace(start_frame, end_frame-1, num_frames, dtype=int)
+    # frame_id_list = np.linspace(0, duration-1, num_frames, dtype=int)
+    video_data = decord_vr.get_batch(frame_id_list)
+    video_data = video_data.permute(3, 0, 1, 2)  # (T, H, W, C) -> (C, T, H, W)
+    # video_outputs = transform(video_data)
+    return video_data
+
+def load_video_1fps(video_path):
+    mp4_stream = None
+    decord.bridge.set_bridge('torch')
+    with open(video_path, 'rb') as f:
+        mp4_stream = f.read()
+
+    num_frames = 24
+    # decord.bridge.set_bridge('torch')
+    if mp4_stream is not None:
+        decord_vr = VideoReader(io.BytesIO(mp4_stream), ctx=cpu(0))
+    else:
+        decord_vr = VideoReader(video_path, ctx=cpu(0))
+    
+    total_frames = len(decord_vr)
+    timestamps = decord_vr.get_frame_timestamp(np.arange(total_frames)) 
+    timestamps = [i[0] for i in timestamps]
+
+    max_second = round(max(timestamps)) + 1
+    frame_id_list = []
+    for second in range(max_second):
+        closest_num = min(timestamps, key=lambda x: abs(x - second))
+        index = timestamps.index(closest_num)
+        frame_id_list.append(index)
+        if len(frame_id_list) > num_frames:
+            break
+        
+    video_data = decord_vr.get_batch(frame_id_list)
+    video_data = video_data.permute(3, 0, 1, 2)  # (T, H, W, C) -> (C, T, H, W)
+    # video_outputs = transform(video_data)
+    return video_data
+
+
+
+class CogVLMVideoForCausalLM(CogVLMPreTrainedModel):
+    _auto_class = "AutoModelForCausalLM"
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = CogVLMVideoModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.video_downsample = 1 # TODO: change this to config
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    def forward(
+            self,
+            input_ids: torch.LongTensor = None,
+            images: List[List[torch.Tensor]] = None,
+            token_type_ids: Optional[torch.LongTensor] = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            position_ids: Optional[torch.LongTensor] = None,
+            past_key_values: Optional[List[torch.FloatTensor]] = None,
+            inputs_embeds: Optional[torch.FloatTensor] = None,
+            use_cache: Optional[bool] = None,
+            output_attentions: Optional[bool] = None,
+            output_hidden_states: Optional[bool] = None,
+            return_dict: Optional[bool] = None,
+            labels: Optional[torch.LongTensor] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            images=images,
+            token_type_ids=token_type_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+        logits = self.lm_head(hidden_states)
+        logits = logits.float()
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def _prepare_attention_mask_for_generation(
+            self,
+            inputs: torch.Tensor,
+            pad_token_id: Optional[int],
+            eos_token_id: Optional[Union[int, List[int]]],
+    ) -> torch.LongTensor:
+        return torch.ones(inputs.shape[:2], dtype=torch.long, device=inputs.device)  # type: ignore
+
+    def prepare_inputs_for_generation(
+            self, input_ids, token_type_ids, images=None, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
+    ):
+        # build position_ids if needed
+        position_ids = kwargs.get("position_ids", None)
+        if position_ids is None:
+            position_ids = build_position_ids(token_type_ids, attention_mask)
+
+        if past_key_values:
+            input_ids = input_ids[:, -1:]
+            token_type_ids = token_type_ids[:, -1:]
+            position_ids = position_ids[:, -1:]
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "token_type_ids": token_type_ids,
+                "images": images,
+                "position_ids": position_ids,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+
+    def _update_model_kwargs_for_generation(
+            self,
+            outputs: "ModelOutput",
+            model_kwargs: Dict[str, Any],
+            is_encoder_decoder: bool = False,
+            standardize_cache_format: bool = False,
+    ) -> Dict[str, Any]:
+        # update past_key_values
+        model_kwargs["past_key_values"] = self._extract_past_from_model_output(
+            outputs, standardize_cache_format=standardize_cache_format
+        )
+        if getattr(outputs, "state", None) is not None:
+            model_kwargs["state"] = outputs.state
+
+        # update token_type_ids with last value
+        if "token_type_ids" in model_kwargs:
+            token_type_ids = model_kwargs["token_type_ids"]
+            new_token_type_ids = torch.ones(size=(token_type_ids.shape[0], 1), dtype=token_type_ids.dtype, device=token_type_ids.device) * LANGUAGE_TOKEN_TYPE
+            model_kwargs["token_type_ids"] = torch.cat([token_type_ids, new_token_type_ids], dim=-1)
+
+        if not is_encoder_decoder:
+            # update attention mask
+            if "attention_mask" in model_kwargs:
+                attention_mask = model_kwargs["attention_mask"]
+                model_kwargs["attention_mask"] = torch.cat(
+                    [attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1
+                )
+        else:
+            # update decoder attention mask
+            if "decoder_attention_mask" in model_kwargs:
+                decoder_attention_mask = model_kwargs["decoder_attention_mask"]
+                model_kwargs["decoder_attention_mask"] = torch.cat(
+                    [decoder_attention_mask, decoder_attention_mask.new_ones((decoder_attention_mask.shape[0], 1))],
+                    dim=-1,
+                )
+
+        return model_kwargs
+
+    def _reorder_cache(self, past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (
+                tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
+            )
+        return reordered_past
+
+
+
+    def build_conversation_input_ids(
+            self,
+            tokenizer: "PreTrainedTokenizer",
+            *,
+            query: str,
+            history: Optional[List[Tuple[str, str]]] = None,
+            images: Optional[List["PIL.Image"]] = None,
+            template_version: Optional[Literal["base", "chat", "vqa"]] = None,
+            answer: str = None,
+    ):
+        image_size: int = self.config.vision_config['image_size']
+        template_version = template_version or self.config.template_version
+        assert images is None or len(images) <= 1, f"not support multi images by now."
+        history = history or []
+        text = _history_to_prompt(template_version, history, query)
+        input_ids = [tokenizer.bos_token_id]
+        token_type_ids = [LANGUAGE_TOKEN_TYPE]
+        add_time_indices = False
+        if images is not None and len(images) == 1:
+            # vision
+            transform = transforms.Compose(
+                [
+                    # UniformTemporalSubsample(num_frames),
+                    Lambda(lambda x: x / 255.0),
+                    NormalizeVideo(mean=(0.48145466, 0.4578275, 0.40821073), std=(0.26862954, 0.26130258, 0.27577711)),
+                    ShortSideScale(size=image_size),
+                    CenterCropVideo(image_size),
+                    # RandomHorizontalFlipVideo(p=0.5),
+                ]
+            )
+            images = [transform(images[0]).transpose(0, 1)] # (T, C, H, W)
+            num_eois = len(images[0])
+            tokenizer.pad_token_id = 128002
+            vision_token_num = (64 + 2) * num_eois 
+            if not add_time_indices:
+                input_ids += [tokenizer.pad_token_id] * vision_token_num # add spetial token 
+                token_type_ids += [VISION_TOKEN_TYPE] * vision_token_num
+            else:
+                video_ids, video_type_ids = [], []
+                for _time_idx in range(num_eois):
+                    video_ids += [tokenizer.pad_token_id] * vision_token_num
+                    video_type_ids += [VISION_TOKEN_TYPE] * vision_token_num
+                    # add time indices
+                    time_indices = tokenizer.encode(str(_time_idx), add_special_tokens=False)
+                    video_ids += time_indices
+                    video_type_ids += [LANGUAGE_TOKEN_TYPE] * len(time_indices)
+                # llama3 adapt for cogvlm
+                input_ids += video_ids
+                token_type_ids += video_type_ids
+            
+        text_ids = tokenizer.encode(text, add_special_tokens=False)
+
+        if answer is not None:
+            answer_ids = tokenizer.encode(answer, add_special_tokens=False)
+            answer_ids += [tokenizer.eos_token_id]
+            text_ids += answer_ids
+
+
+        input_ids += text_ids
+        token_type_ids += [LANGUAGE_TOKEN_TYPE] * len(text_ids)
+        attention_mask = [1] * len(input_ids)
+        if answer is not None:
+            labels = [-100 for _ in range(len(input_ids) - len(answer_ids))] + answer_ids
+            labels = torch.tensor(labels, dtype=torch.long)
+        else:
+            labels = None
+
+        return {
+            'input_ids': torch.tensor(input_ids, dtype=torch.long),
+            'token_type_ids': torch.tensor(token_type_ids, dtype=torch.long),
+            'attention_mask': torch.tensor(attention_mask, dtype=torch.long),
+            'images': images,
+            'labels': labels,
+        }
+
+    

special_tokens_map.json

@@ -0,0 +1,4 @@
+{
+  "bos_token": "<|begin_of_text|>",
+  "eos_token": "<|end_of_text|>"
+}

tokenizer_config.json

@@ -0,0 +1,2064 @@
+{
+  "added_tokens_decoder": {
+    "128000": {
+      "content": "<|begin_of_text|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128001": {
+      "content": "<|end_of_text|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128002": {
+      "content": "<|reserved_special_token_0|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128003": {
+      "content": "<|reserved_special_token_1|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128004": {
+      "content": "<|reserved_special_token_2|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128005": {
+      "content": "<|reserved_special_token_3|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128006": {
+      "content": "<|start_header_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128007": {
+      "content": "<|end_header_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128008": {
+      "content": "<|reserved_special_token_4|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128009": {
+      "content": "<|eot_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128010": {
+      "content": "<|reserved_special_token_5|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128011": {
+      "content": "<|reserved_special_token_6|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128012": {
+      "content": "<|reserved_special_token_7|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128013": {
+      "content": "<|reserved_special_token_8|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128014": {
+      "content": "<|reserved_special_token_9|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128015": {
+      "content": "<|reserved_special_token_10|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128016": {
+      "content": "<|reserved_special_token_11|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128017": {
+      "content": "<|reserved_special_token_12|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128018": {
+      "content": "<|reserved_special_token_13|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128019": {
+      "content": "<|reserved_special_token_14|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128020": {
+      "content": "<|reserved_special_token_15|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128021": {
+      "content": "<|reserved_special_token_16|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128022": {
+      "content": "<|reserved_special_token_17|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128023": {
+      "content": "<|reserved_special_token_18|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128024": {
+      "content": "<|reserved_special_token_19|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128025": {
+      "content": "<|reserved_special_token_20|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128026": {
+      "content": "<|reserved_special_token_21|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128027": {
+      "content": "<|reserved_special_token_22|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128028": {
+      "content": "<|reserved_special_token_23|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128029": {
+      "content": "<|reserved_special_token_24|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128030": {
+      "content": "<|reserved_special_token_25|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128031": {
+      "content": "<|reserved_special_token_26|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128032": {
+      "content": "<|reserved_special_token_27|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128033": {
+      "content": "<|reserved_special_token_28|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128034": {
+      "content": "<|reserved_special_token_29|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128035": {
+      "content": "<|reserved_special_token_30|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128036": {
+      "content": "<|reserved_special_token_31|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128037": {
+      "content": "<|reserved_special_token_32|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128038": {
+      "content": "<|reserved_special_token_33|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128039": {
+      "content": "<|reserved_special_token_34|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128040": {
+      "content": "<|reserved_special_token_35|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128041": {
+      "content": "<|reserved_special_token_36|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128042": {
+      "content": "<|reserved_special_token_37|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128043": {
+      "content": "<|reserved_special_token_38|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128044": {
+      "content": "<|reserved_special_token_39|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128045": {
+      "content": "<|reserved_special_token_40|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128046": {
+      "content": "<|reserved_special_token_41|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128047": {
+      "content": "<|reserved_special_token_42|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128048": {
+      "content": "<|reserved_special_token_43|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128049": {
+      "content": "<|reserved_special_token_44|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128050": {
+      "content": "<|reserved_special_token_45|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128051": {
+      "content": "<|reserved_special_token_46|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128052": {
+      "content": "<|reserved_special_token_47|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128053": {
+      "content": "<|reserved_special_token_48|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128054": {
+      "content": "<|reserved_special_token_49|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128055": {
+      "content": "<|reserved_special_token_50|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128056": {
+      "content": "<|reserved_special_token_51|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128057": {
+      "content": "<|reserved_special_token_52|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128058": {
+      "content": "<|reserved_special_token_53|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128059": {
+      "content": "<|reserved_special_token_54|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128060": {
+      "content": "<|reserved_special_token_55|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128061": {
+      "content": "<|reserved_special_token_56|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128062": {
+      "content": "<|reserved_special_token_57|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128063": {
+      "content": "<|reserved_special_token_58|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128064": {
+      "content": "<|reserved_special_token_59|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128065": {
+      "content": "<|reserved_special_token_60|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128066": {
+      "content": "<|reserved_special_token_61|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128067": {
+      "content": "<|reserved_special_token_62|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128068": {
+      "content": "<|reserved_special_token_63|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128069": {
+      "content": "<|reserved_special_token_64|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128070": {
+      "content": "<|reserved_special_token_65|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128071": {
+      "content": "<|reserved_special_token_66|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128072": {
+      "content": "<|reserved_special_token_67|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128073": {
+      "content": "<|reserved_special_token_68|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128074": {
+      "content": "<|reserved_special_token_69|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128075": {
+      "content": "<|reserved_special_token_70|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128076": {
+      "content": "<|reserved_special_token_71|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128077": {
+      "content": "<|reserved_special_token_72|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128078": {
+      "content": "<|reserved_special_token_73|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128079": {
+      "content": "<|reserved_special_token_74|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128080": {
+      "content": "<|reserved_special_token_75|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128081": {
+      "content": "<|reserved_special_token_76|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128082": {
+      "content": "<|reserved_special_token_77|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128083": {
+      "content": "<|reserved_special_token_78|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128084": {
+      "content": "<|reserved_special_token_79|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128085": {
+      "content": "<|reserved_special_token_80|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128086": {
+      "content": "<|reserved_special_token_81|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128087": {
+      "content": "<|reserved_special_token_82|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128088": {
+      "content": "<|reserved_special_token_83|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128089": {
+      "content": "<|reserved_special_token_84|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128090": {
+      "content": "<|reserved_special_token_85|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128091": {
+      "content": "<|reserved_special_token_86|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128092": {
+      "content": "<|reserved_special_token_87|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128093": {
+      "content": "<|reserved_special_token_88|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128094": {
+      "content": "<|reserved_special_token_89|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128095": {
+      "content": "<|reserved_special_token_90|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128096": {
+      "content": "<|reserved_special_token_91|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128097": {
+      "content": "<|reserved_special_token_92|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128098": {
+      "content": "<|reserved_special_token_93|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128099": {
+      "content": "<|reserved_special_token_94|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128100": {
+      "content": "<|reserved_special_token_95|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128101": {
+      "content": "<|reserved_special_token_96|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128102": {
+      "content": "<|reserved_special_token_97|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128103": {
+      "content": "<|reserved_special_token_98|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128104": {
+      "content": "<|reserved_special_token_99|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128105": {
+      "content": "<|reserved_special_token_100|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128106": {
+      "content": "<|reserved_special_token_101|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128107": {
+      "content": "<|reserved_special_token_102|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128108": {
+      "content": "<|reserved_special_token_103|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128109": {
+      "content": "<|reserved_special_token_104|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128110": {
+      "content": "<|reserved_special_token_105|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128111": {
+      "content": "<|reserved_special_token_106|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128112": {
+      "content": "<|reserved_special_token_107|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128113": {
+      "content": "<|reserved_special_token_108|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128114": {
+      "content": "<|reserved_special_token_109|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128115": {
+      "content": "<|reserved_special_token_110|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128116": {
+      "content": "<|reserved_special_token_111|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128117": {
+      "content": "<|reserved_special_token_112|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128118": {
+      "content": "<|reserved_special_token_113|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128119": {
+      "content": "<|reserved_special_token_114|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128120": {
+      "content": "<|reserved_special_token_115|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128121": {
+      "content": "<|reserved_special_token_116|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128122": {
+      "content": "<|reserved_special_token_117|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128123": {
+      "content": "<|reserved_special_token_118|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128124": {
+      "content": "<|reserved_special_token_119|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128125": {
+      "content": "<|reserved_special_token_120|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128126": {
+      "content": "<|reserved_special_token_121|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128127": {
+      "content": "<|reserved_special_token_122|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128128": {
+      "content": "<|reserved_special_token_123|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128129": {
+      "content": "<|reserved_special_token_124|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128130": {
+      "content": "<|reserved_special_token_125|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128131": {
+      "content": "<|reserved_special_token_126|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128132": {
+      "content": "<|reserved_special_token_127|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128133": {
+      "content": "<|reserved_special_token_128|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128134": {
+      "content": "<|reserved_special_token_129|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128135": {
+      "content": "<|reserved_special_token_130|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128136": {
+      "content": "<|reserved_special_token_131|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128137": {
+      "content": "<|reserved_special_token_132|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128138": {
+      "content": "<|reserved_special_token_133|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128139": {
+      "content": "<|reserved_special_token_134|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128140": {
+      "content": "<|reserved_special_token_135|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128141": {
+      "content": "<|reserved_special_token_136|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128142": {
+      "content": "<|reserved_special_token_137|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128143": {
+      "content": "<|reserved_special_token_138|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128144": {
+      "content": "<|reserved_special_token_139|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128145": {
+      "content": "<|reserved_special_token_140|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128146": {
+      "content": "<|reserved_special_token_141|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128147": {
+      "content": "<|reserved_special_token_142|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128148": {
+      "content": "<|reserved_special_token_143|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128149": {
+      "content": "<|reserved_special_token_144|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128150": {
+      "content": "<|reserved_special_token_145|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128151": {
+      "content": "<|reserved_special_token_146|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128152": {
+      "content": "<|reserved_special_token_147|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128153": {
+      "content": "<|reserved_special_token_148|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128154": {
+      "content": "<|reserved_special_token_149|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128155": {
+      "content": "<|reserved_special_token_150|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128156": {
+      "content": "<|reserved_special_token_151|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128157": {
+      "content": "<|reserved_special_token_152|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128158": {
+      "content": "<|reserved_special_token_153|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128159": {
+      "content": "<|reserved_special_token_154|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128160": {
+      "content": "<|reserved_special_token_155|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128161": {
+      "content": "<|reserved_special_token_156|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128162": {
+      "content": "<|reserved_special_token_157|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128163": {
+      "content": "<|reserved_special_token_158|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128164": {
+      "content": "<|reserved_special_token_159|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128165": {
+      "content": "<|reserved_special_token_160|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128166": {
+      "content": "<|reserved_special_token_161|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128167": {
+      "content": "<|reserved_special_token_162|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128168": {
+      "content": "<|reserved_special_token_163|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128169": {
+      "content": "<|reserved_special_token_164|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128170": {
+      "content": "<|reserved_special_token_165|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128171": {
+      "content": "<|reserved_special_token_166|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128172": {
+      "content": "<|reserved_special_token_167|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128173": {
+      "content": "<|reserved_special_token_168|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128174": {
+      "content": "<|reserved_special_token_169|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128175": {
+      "content": "<|reserved_special_token_170|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128176": {
+      "content": "<|reserved_special_token_171|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128177": {
+      "content": "<|reserved_special_token_172|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128178": {
+      "content": "<|reserved_special_token_173|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128179": {
+      "content": "<|reserved_special_token_174|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128180": {
+      "content": "<|reserved_special_token_175|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128181": {
+      "content": "<|reserved_special_token_176|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128182": {
+      "content": "<|reserved_special_token_177|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128183": {
+      "content": "<|reserved_special_token_178|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128184": {
+      "content": "<|reserved_special_token_179|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128185": {
+      "content": "<|reserved_special_token_180|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128186": {
+      "content": "<|reserved_special_token_181|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128187": {
+      "content": "<|reserved_special_token_182|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128188": {
+      "content": "<|reserved_special_token_183|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128189": {
+      "content": "<|reserved_special_token_184|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128190": {
+      "content": "<|reserved_special_token_185|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128191": {
+      "content": "<|reserved_special_token_186|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128192": {
+      "content": "<|reserved_special_token_187|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128193": {
+      "content": "<|reserved_special_token_188|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128194": {
+      "content": "<|reserved_special_token_189|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128195": {
+      "content": "<|reserved_special_token_190|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128196": {
+      "content": "<|reserved_special_token_191|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128197": {
+      "content": "<|reserved_special_token_192|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128198": {
+      "content": "<|reserved_special_token_193|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128199": {
+      "content": "<|reserved_special_token_194|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128200": {
+      "content": "<|reserved_special_token_195|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128201": {
+      "content": "<|reserved_special_token_196|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128202": {
+      "content": "<|reserved_special_token_197|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128203": {
+      "content": "<|reserved_special_token_198|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128204": {
+      "content": "<|reserved_special_token_199|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128205": {
+      "content": "<|reserved_special_token_200|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128206": {
+      "content": "<|reserved_special_token_201|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128207": {
+      "content": "<|reserved_special_token_202|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128208": {
+      "content": "<|reserved_special_token_203|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128209": {
+      "content": "<|reserved_special_token_204|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128210": {
+      "content": "<|reserved_special_token_205|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128211": {
+      "content": "<|reserved_special_token_206|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128212": {
+      "content": "<|reserved_special_token_207|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128213": {
+      "content": "<|reserved_special_token_208|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128214": {
+      "content": "<|reserved_special_token_209|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128215": {
+      "content": "<|reserved_special_token_210|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128216": {
+      "content": "<|reserved_special_token_211|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128217": {
+      "content": "<|reserved_special_token_212|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128218": {
+      "content": "<|reserved_special_token_213|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128219": {
+      "content": "<|reserved_special_token_214|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128220": {
+      "content": "<|reserved_special_token_215|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128221": {
+      "content": "<|reserved_special_token_216|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128222": {
+      "content": "<|reserved_special_token_217|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128223": {
+      "content": "<|reserved_special_token_218|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128224": {
+      "content": "<|reserved_special_token_219|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128225": {
+      "content": "<|reserved_special_token_220|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128226": {
+      "content": "<|reserved_special_token_221|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128227": {
+      "content": "<|reserved_special_token_222|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128228": {
+      "content": "<|reserved_special_token_223|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128229": {
+      "content": "<|reserved_special_token_224|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128230": {
+      "content": "<|reserved_special_token_225|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128231": {
+      "content": "<|reserved_special_token_226|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128232": {
+      "content": "<|reserved_special_token_227|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128233": {
+      "content": "<|reserved_special_token_228|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128234": {
+      "content": "<|reserved_special_token_229|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128235": {
+      "content": "<|reserved_special_token_230|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128236": {
+      "content": "<|reserved_special_token_231|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128237": {
+      "content": "<|reserved_special_token_232|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128238": {
+      "content": "<|reserved_special_token_233|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128239": {
+      "content": "<|reserved_special_token_234|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128240": {
+      "content": "<|reserved_special_token_235|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128241": {
+      "content": "<|reserved_special_token_236|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128242": {
+      "content": "<|reserved_special_token_237|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128243": {
+      "content": "<|reserved_special_token_238|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128244": {
+      "content": "<|reserved_special_token_239|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128245": {
+      "content": "<|reserved_special_token_240|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128246": {
+      "content": "<|reserved_special_token_241|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128247": {
+      "content": "<|reserved_special_token_242|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128248": {
+      "content": "<|reserved_special_token_243|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128249": {
+      "content": "<|reserved_special_token_244|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128250": {
+      "content": "<|reserved_special_token_245|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128251": {
+      "content": "<|reserved_special_token_246|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128252": {
+      "content": "<|reserved_special_token_247|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128253": {
+      "content": "<|reserved_special_token_248|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128254": {
+      "content": "<|reserved_special_token_249|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128255": {
+      "content": "<|reserved_special_token_250|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<|begin_of_text|>",
+  "chat_template": "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{% if add_generation_prompt %}{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}{% else %}{{ eos_token }}{% endif %}",
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "<|end_of_text|>",
+  "model_input_names": [
+    "input_ids",
+    "token_type_ids",
+    "attention_mask",
+    "images"
+  ],
+  "model_max_length": 2048,
+  "tokenizer_class": "PreTrainedTokenizerFast"
+}

util.py

@@ -0,0 +1,472 @@
+from typing import Optional, Tuple, Union
+
+import torch
+from einops import rearrange
+import torch.nn.functional as F
+
+import triton
+import triton.language as tl
+
+
+@triton.jit
+def rotary_kernel(
+        OUT,
+        X,
+        COS,
+        SIN,
+        CU_SEQLENS,
+        SEQLEN_OFFSETS,
+        seqlen,
+        nheads,
+        rotary_dim,
+        seqlen_ro,
+        CACHE_KEY_SEQLEN,
+        # strides
+        stride_out_batch,
+        stride_out_nheads,
+        stride_out_seqlen,
+        stride_out_headdim,
+        stride_x_batch,
+        stride_x_nheads,
+        stride_x_seqlen,
+        stride_x_headdim,
+        BLOCK_K: tl.constexpr,
+        IS_SEQLEN_OFFSETS_TENSOR: tl.constexpr,
+        IS_VARLEN: tl.constexpr,
+        INTERLEAVED: tl.constexpr,
+        CONJUGATE: tl.constexpr,
+        BLOCK_M: tl.constexpr,
+):
+    pid_m = tl.program_id(axis=0)
+    pid_batch = tl.program_id(axis=1)
+    pid_head = tl.program_id(axis=2)
+    rotary_dim_half = rotary_dim // 2
+
+    if not IS_VARLEN:
+        X = X + pid_batch * stride_x_batch + pid_head * stride_x_nheads
+        OUT = OUT + pid_batch * stride_out_batch + pid_head * stride_out_nheads
+        COS = COS + pid_batch * seqlen_ro * rotary_dim_half
+        SIN = SIN + pid_batch * seqlen_ro * rotary_dim_half
+    else:
+        start_idx = tl.load(CU_SEQLENS + pid_batch)
+        seqlen = tl.load(CU_SEQLENS + pid_batch + 1) - start_idx
+        X = X + start_idx * stride_x_seqlen + pid_head * stride_x_nheads
+        OUT = OUT + start_idx * stride_out_seqlen + pid_head * stride_out_nheads
+
+    if pid_m * BLOCK_M >= seqlen:
+        return
+    rm = pid_m * BLOCK_M + tl.arange(0, BLOCK_M)
+    if not IS_SEQLEN_OFFSETS_TENSOR:
+        rm_cs = rm + SEQLEN_OFFSETS
+    else:
+        rm_cs = rm + tl.load(SEQLEN_OFFSETS + pid_batch)
+    rk = tl.arange(0, BLOCK_K)
+    rk_half = tl.arange(0, BLOCK_K // 2)
+
+    if not INTERLEAVED:
+        # Load the 1st and 2nd halves of X, do calculation, then store to 1st and 2nd halves of OUT
+        X = X + (rm[:, None] * stride_x_seqlen + rk_half[None, :] * stride_x_headdim)
+        COS = COS + (rm_cs[:, None] * rotary_dim_half + rk_half[None, :])
+        SIN = SIN + (rm_cs[:, None] * rotary_dim_half + rk_half[None, :])
+        cos = tl.load(
+            COS, mask=(rm_cs[:, None] < seqlen_ro) & (rk_half[None, :] < rotary_dim_half), other=1.0
+        )
+        sin = tl.load(
+            SIN, mask=(rm_cs[:, None] < seqlen_ro) & (rk_half[None, :] < rotary_dim_half), other=0.0
+        )
+        x0 = tl.load(
+            X, mask=(rm[:, None] < seqlen) & (rk_half[None, :] < rotary_dim_half), other=0.0
+        )
+        x1 = tl.load(
+            X + rotary_dim_half * stride_x_headdim,
+            mask=(rm[:, None] < seqlen) & (rk_half[None, :] < rotary_dim_half),
+            other=0.0,
+        )
+        if CONJUGATE:
+            sin = -sin
+        o0 = x0 * cos - x1 * sin
+        o1 = x0 * sin + x1 * cos
+        # write back result
+        OUT = OUT + (rm[:, None] * stride_out_seqlen + rk_half[None, :] * stride_out_headdim)
+        tl.store(OUT, o0, mask=(rm[:, None] < seqlen) & (rk_half[None, :] < rotary_dim_half))
+        tl.store(
+            OUT + rotary_dim_half * stride_out_headdim,
+            o1,
+            mask=(rm[:, None] < seqlen) & (rk_half[None, :] < rotary_dim_half),
+        )
+    else:
+        # We don't want to load X[0, 2, 4, ...] and X[1, 3, 5, ...] separately since both are slow.
+        # Instead, we load x0 = X[0, 1, 2, 3, ...] and x1 = X[1, 0, 3, 2, ...].
+        # Loading x0 will be fast but x1 will be slow.
+        # Then we load cos = COS[0, 0, 1, 1, ...] and sin = SIN[0, 0, 1, 1, ...].
+        # Then we do the calculation and use tl.where to pick put the right outputs for the even
+        # and for the odd indices.
+        rk_swap = rk + ((rk + 1) % 2) * 2 - 1  # 1, 0, 3, 2, 5, 4, ...
+        rk_repeat = tl.arange(0, BLOCK_K) // 2
+        X0 = X + (rm[:, None] * stride_x_seqlen + rk[None, :] * stride_x_headdim)
+        X1 = X + (rm[:, None] * stride_x_seqlen + rk_swap[None, :] * stride_x_headdim)
+        COS = COS + (rm_cs[:, None] * rotary_dim_half + rk_repeat[None, :])
+        SIN = SIN + (rm_cs[:, None] * rotary_dim_half + rk_repeat[None, :])
+        cos = tl.load(
+            COS,
+            mask=(rm_cs[:, None] < seqlen_ro) & (rk_repeat[None, :] < rotary_dim_half),
+            other=1.0,
+        ).to(tl.float32)
+        sin = tl.load(
+            SIN,
+            mask=(rm_cs[:, None] < seqlen_ro) & (rk_repeat[None, :] < rotary_dim_half),
+            other=0.0,
+        ).to(tl.float32)
+        x0 = tl.load(X0, mask=(rm[:, None] < seqlen) & (rk[None, :] < rotary_dim), other=0.0).to(
+            tl.float32
+        )
+        x1 = tl.load(
+            X1, mask=(rm[:, None] < seqlen) & (rk_swap[None, :] < rotary_dim), other=0.0
+        ).to(tl.float32)
+        if CONJUGATE:
+            sin = -sin
+        x0_cos = x0 * cos
+        x1_sin = x1 * sin
+        out = tl.where(rk[None, :] % 2 == 0, x0_cos - x1_sin, x0_cos + x1_sin)
+        OUT = OUT + (rm[:, None] * stride_out_seqlen + rk[None, :] * stride_out_headdim)
+        tl.store(OUT, out, mask=(rm[:, None] < seqlen) & (rk[None, :] < rotary_dim))
+
+
+def apply_rotary(
+        x: torch.Tensor,
+        cos: torch.Tensor,
+        sin: torch.Tensor,
+        seqlen_offsets: Union[int, torch.Tensor] = 0,
+        cu_seqlens: Optional[torch.Tensor] = None,
+        max_seqlen: Optional[int] = None,
+        interleaved=False,
+        inplace=False,
+        conjugate=False,
+) -> torch.Tensor:
+    """
+    Arguments:
+        x: (batch, seqlen, nheads, headdim) if cu_seqlens is None
+            else (total_seqlen, nheads, headdim).
+        cos: (seqlen_ro, rotary_dim / 2)
+        sin: (seqlen_ro, rotary_dim / 2)
+        seqlen_offsets: integer or integer tensor of size (batch,)
+        cu_seqlens: (batch + 1,) or None
+        max_seqlen: int
+    Returns:
+        y: (batch, seqlen, nheads, headdim)
+    """
+
+    batch, nheads, seqlen, headdim = x.shape
+
+    batch_ro, seqlen_ro, rotary_dim = cos.shape
+
+    assert batch == batch_ro
+    assert sin.shape == cos.shape
+    rotary_dim *= 2
+    assert rotary_dim <= headdim, "rotary_dim must be <= headdim"
+    assert headdim <= 256, "Only support headdim <= 256"
+
+    assert seqlen_ro >= seqlen, "seqlen_ro must be >= seqlen"
+
+    assert (
+            cos.dtype == sin.dtype
+    ), f"cos and sin must have the same dtype, got {cos.dtype} and {sin.dtype}"
+    assert (
+            x.dtype == cos.dtype
+    ), f"Input and cos/sin must have the same dtype, got {x.dtype} and {cos.dtype}"
+
+    cos, sin = cos.contiguous(), sin.contiguous()
+    if isinstance(seqlen_offsets, torch.Tensor):
+        assert seqlen_offsets.shape == (batch,)
+        assert seqlen_offsets.dtype in [torch.int32, torch.int64]
+        seqlen_offsets = seqlen_offsets.contiguous()
+    else:
+        assert seqlen_offsets + seqlen <= seqlen_ro
+
+    output = torch.empty_like(x) if not inplace else x
+    if rotary_dim < headdim and not inplace:
+        output[..., rotary_dim:].copy_(x[..., rotary_dim:])
+
+    BLOCK_K = (
+        32
+        if rotary_dim <= 32
+        else (64 if rotary_dim <= 64 else (128 if rotary_dim <= 128 else 256))
+    )
+    grid = lambda META: (triton.cdiv(seqlen, META["BLOCK_M"]), batch, nheads)  # noqa
+    BLOCK_M = 4 if interleaved else (8 if rotary_dim <= 64 else 4)
+
+    # Need this, otherwise Triton tries to launch from cuda:0 and we get
+    # ValueError: Pointer argument (at 0) cannot be accessed from Triton (cpu tensor?)
+    with torch.cuda.device(x.device.index):
+        rotary_kernel[grid](
+            output,  # data ptrs
+            x,
+            cos,
+            sin,
+            cu_seqlens,
+            seqlen_offsets,
+            seqlen,  # shapes
+            nheads,
+            rotary_dim,
+            seqlen_ro,
+            seqlen // 128,  # key for triton cache (limit number of compilations)
+            output.stride(0),  # batch_strides
+            output.stride(-3),  # nheads_stride
+            output.stride(-2),  # seqlen_stride
+            output.stride(-1),  # headdim_stride
+            x.stride(0),  # batch_strides
+            x.stride(-3),  # nheads stride
+            x.stride(-2),  # seqlen stride
+            x.stride(-1),  # headdim stride
+            BLOCK_K,
+            isinstance(seqlen_offsets, torch.Tensor),
+            False,
+            interleaved,
+            conjugate,
+            BLOCK_M,
+        )
+    return output
+
+
+class ApplyRotaryEmb(torch.autograd.Function):
+    @staticmethod
+    def forward(
+            ctx,
+            x,
+            cos,
+            sin,
+            interleaved=False,
+            inplace=False,
+            seqlen_offsets: Union[int, torch.Tensor] = 0,
+            cu_seqlens: Optional[torch.Tensor] = None,
+            max_seqlen: Optional[int] = None,
+    ):
+        out = apply_rotary(
+            x,
+            cos,
+            sin,
+            seqlen_offsets=seqlen_offsets,
+            cu_seqlens=cu_seqlens,
+            max_seqlen=max_seqlen,
+            interleaved=interleaved,
+            inplace=inplace,
+        )
+        if isinstance(seqlen_offsets, int):
+            ctx.save_for_backward(cos, sin, cu_seqlens)  # Can't save int with save_for_backward
+            ctx.seqlen_offsets = seqlen_offsets
+        else:
+            ctx.save_for_backward(cos, sin, cu_seqlens, seqlen_offsets)
+            ctx.seqlen_offsets = None
+        ctx.interleaved = interleaved
+        ctx.inplace = inplace
+        ctx.max_seqlen = max_seqlen
+        return out if not inplace else x
+
+    @staticmethod
+    def backward(ctx, do):
+        seqlen_offsets = ctx.seqlen_offsets
+        if seqlen_offsets is None:
+            cos, sin, cu_seqlens, seqlen_offsets = ctx.saved_tensors
+        else:
+            cos, sin, cu_seqlens = ctx.saved_tensors
+        # TD [2023-09-02]: For some reason Triton (2.0.0.post1) errors with
+        # "[CUDA]: invalid device context", and cloning makes it work. Idk why. Triton 2.1.0 works.
+        if not ctx.interleaved and not ctx.inplace:
+            do = do.clone()
+        dx = apply_rotary(
+            do,
+            cos,
+            sin,
+            seqlen_offsets=seqlen_offsets,
+            cu_seqlens=cu_seqlens,
+            max_seqlen=ctx.max_seqlen,
+            interleaved=ctx.interleaved,
+            inplace=ctx.inplace,
+            conjugate=True,
+        )
+        return dx, None, None, None, None, None, None, None
+
+
+def apply_rotary_emb(
+        x,
+        cos,
+        sin,
+        interleaved=False,
+        inplace=False,
+        seqlen_offsets: Union[int, torch.Tensor] = 0,
+        cu_seqlens: Optional[torch.Tensor] = None,
+        max_seqlen: Optional[int] = None,
+):
+    """
+    Arguments:
+        x: (batch_size, seqlen, nheads, headdim) if cu_seqlens is None
+            else (total_seqlen, nheads, headdim)
+        cos, sin: (seqlen_rotary, rotary_dim / 2)
+        interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead
+            of 1st half and 2nd half (GPT-NeoX style).
+        inplace: if True, apply rotary embedding in-place.
+        seqlen_offsets: (batch_size,) or int. Each sequence in x is shifted by this amount.
+            Most commonly used in inference when we have KV cache.
+        cu_seqlens: (batch + 1,) or None
+        max_seqlen: int
+    Return:
+        out: (batch_size, seqlen, nheads, headdim) if cu_seqlens is None
+            else (total_seqlen, nheads, headdim)
+    rotary_dim must be <= headdim
+    Apply rotary embedding to the first rotary_dim of x.
+    """
+    return ApplyRotaryEmb.apply(
+        x, cos, sin, interleaved, inplace, seqlen_offsets, cu_seqlens, max_seqlen
+    )
+
+
+# For backward compatibility
+apply_rotary_emb_func = apply_rotary_emb
+
+
+class FastRotaryEmbedding(torch.nn.Module):
+    """
+    The rotary position embeddings from RoFormer_ (Su et. al).
+    A crucial insight from the method is that the query and keys are
+    transformed by rotation matrices which depend on the relative positions.
+
+    Other implementations are available in the Rotary Transformer repo_ and in
+    GPT-NeoX_, GPT-NeoX was an inspiration
+
+    .. _RoFormer: https://arxiv.org/abs/2104.09864
+    .. _repo: https://github.com/ZhuiyiTechnology/roformer
+    .. _GPT-NeoX: https://github.com/EleutherAI/gpt-neox
+
+    If scale_base is not None, this implements XPos (Sun et al., https://arxiv.org/abs/2212.10554).
+    A recommended value for scale_base is 512: https://github.com/HazyResearch/flash-attention/issues/96
+    Reference: https://github.com/sunyt32/torchscale/blob/main/torchscale/component/xpos_relative_position.py
+    """
+
+    def __init__(
+            self,
+            dim: int,
+            base=10000,
+            interleaved=False,
+            scale_base=None,
+            pos_idx_in_fp32=True,
+            device=None,
+    ):
+        """
+        interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead
+            of 1st half and 2nd half (GPT-NeoX style).
+        pos_idx_in_fp32: if True, the position indices [0.0, ..., seqlen - 1] are in fp32,
+            otherwise they might be in lower precision.
+            This option was added because previously (before 2023-07-02), when we construct
+            the position indices, we use the dtype of self.inv_freq. In most cases this would
+            be fp32, but if the model is trained in pure bf16 (not mixed precision), then
+            self.inv_freq would be bf16, and the position indices are also in bf16.
+            Because of the limited precision of bf16 (e.g. 1995.0 is rounded to 2000.0), the
+            embeddings for some positions will coincide.
+            To maintain compatibility with models previously trained in pure bf16,
+            we add this option.
+        """
+        super().__init__()
+        self.dim = dim
+        self.base = base
+        self.pos_idx_in_fp32 = pos_idx_in_fp32
+        # Generate and save the inverse frequency buffer (non trainable)
+        inv_freq = self._compute_inv_freq(device)
+        self.register_buffer("inv_freq", inv_freq)
+        self.interleaved = interleaved
+        self.scale_base = scale_base
+        scale = (
+            (torch.arange(0, dim, 2, device=device, dtype=torch.float32) + 0.4 * dim) / (1.4 * dim)
+            if scale_base is not None
+            else None
+        )
+        self.register_buffer("scale", scale, persistent=False)
+
+        self._seq_len_cached = 0
+        self._cos_cached = None
+        self._sin_cached = None
+        self._cos_k_cached = None
+        self._sin_k_cached = None
+        self.cos = None
+        self.sin = None
+
+    def _compute_inv_freq(self, device=None):
+        return 1.0 / (
+                self.base
+                ** (torch.arange(0, self.dim, 2, device=device) / self.dim)
+            # ** (torch.arange(0, self.dim, 2, device=device).float() / self.dim)
+        )
+
+    def _update_cos_sin_cache(self, seqlen, position_id, device=None, dtype=None):
+
+        if (
+                seqlen > self._seq_len_cached
+        ):
+            self._seq_len_cached = seqlen
+            # We want fp32 here, not self.inv_freq.dtype, since the model could be loaded in bf16
+            # And the output of arange can be quite large, so bf16 would lose a lot of precision.
+            # However, for compatibility reason, we add an option to use the dtype of self.inv_freq.
+            if self.pos_idx_in_fp32:
+                t = torch.arange(seqlen, device=device, dtype=torch.float32)
+                # We want fp32 here as well since inv_freq will be multiplied with t, and the output
+                # will be large. Having it in bf16 will lose a lot of precision and cause the
+                # cos & sin output to change significantly.
+                # We want to recompute self.inv_freq if it was not loaded in fp32
+                if self.inv_freq.dtype != torch.float32:
+                    inv_freq = self._compute_inv_freq(device=device)
+                else:
+                    inv_freq = self.inv_freq
+            else:
+                t = torch.arange(seqlen, device=device, dtype=self.inv_freq.dtype)
+                inv_freq = self.inv_freq
+            freqs = torch.einsum("i,j->ij", t, inv_freq)
+            if self.scale is None:
+                self._cos_cached = torch.cos(freqs).to(dtype)
+                self._sin_cached = torch.sin(freqs).to(dtype)
+
+            else:
+                power = (
+                                torch.arange(seqlen, dtype=self.scale.dtype, device=self.scale.device)
+                                - seqlen // 2
+                        ) / self.scale_base
+                scale = self.scale.to(device=power.device) ** rearrange(power, "s -> s 1")
+                # We want the multiplication by scale to happen in fp32
+                self._cos_cached = (torch.cos(freqs) * scale).to(dtype)
+                self._sin_cached = (torch.sin(freqs) * scale).to(dtype)
+                self._cos_k_cached = (torch.cos(freqs) / scale).to(dtype)
+                self._sin_k_cached = (torch.sin(freqs) / scale).to(dtype)
+
+    def forward(
+            self,
+            q: torch.Tensor,
+            k: torch.Tensor,
+            position_ids: torch.Tensor,
+            max_seqlen,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        q: (batch, nheads, seqlen, headdim)
+        k: (batch, nheads, seqlen, headdim)
+        position_id: (batch, seqlen)
+        max_seqlen: int
+        layer_id: int
+            only if layer_id == 0, then update cons and sin
+        Apply rotary embedding *inplace* to q k.
+        """
+
+        self._update_cos_sin_cache(max_seqlen, position_ids, device=q.device, dtype=q.dtype)
+        cos, sin = F.embedding(position_ids, self._cos_cached), F.embedding(position_ids, self._sin_cached)
+
+        q = apply_rotary_emb_func(
+            q,
+            cos,
+            sin,
+            interleaved=self.interleaved,
+            inplace=True
+        )
+        k = apply_rotary_emb_func(
+            k,
+            cos,
+            sin,
+            interleaved=self.interleaved,
+            inplace=True
+        )
+        return q, k

visual.py

@@ -0,0 +1,177 @@
+import torch
+from torch import nn
+from argparse import Namespace
+import torch.nn.functional as F
+from transformers.activations import ACT2FN
+import math
+
+def standard_attention(query_layer, key_layer, value_layer, scaling_attention_score=True):
+    if scaling_attention_score:
+        query_layer = query_layer / math.sqrt(query_layer.shape[-1])
+    attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+    attention_probs = F.softmax(attention_scores, dim=-1)
+
+    context_layer = torch.matmul(attention_probs, value_layer)
+    return context_layer
+
+def attention_fn_default(query_layer, key_layer, value_layer, scaling_attention_score=True):
+    # expand head dim to query dim, if necessary
+    # only useful for multi-query attention
+    batch_size, num_query_heads = query_layer.shape[:2] # [b, np, s, hn]
+    num_kv_heads = key_layer.shape[1] # [b, np, s, hn]
+    key_layer = key_layer.unsqueeze(2).expand(-1, -1, num_query_heads//num_kv_heads, -1, -1).contiguous().view(batch_size, num_query_heads, *key_layer.shape[2:])
+    value_layer = value_layer.unsqueeze(2).expand(-1, -1, num_query_heads//num_kv_heads, -1, -1).contiguous().view(batch_size, num_query_heads, *value_layer.shape[2:])
+
+    if int(torch.__version__.split('.')[0]) >= 2 and scaling_attention_score:
+        # Pytorch 2.0 attention uses very much memory if attention_mask is float, and has NaN bug if attention_mask is None.
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_layer, key_layer, value_layer, 
+            attn_mask=None,
+            dropout_p=0.,
+            is_causal=False
+        )
+        return attn_output
+    else:
+        return standard_attention(
+            query_layer, key_layer, value_layer, scaling_attention_score=scaling_attention_score
+        )
+
+class PatchEmbedding(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.proj = nn.Conv2d(config.in_channels, config.hidden_size, kernel_size=config.patch_size, stride=config.patch_size)
+        self.cls_embedding = nn.Parameter(torch.zeros(1, config.hidden_size))
+        self.position_embedding = nn.Embedding(config.num_positions, config.hidden_size)
+
+    def forward(self, images: "tensor(B, C, H, W)") -> "tensor(B, L, D)":
+        x = self.proj(images)
+        x = x.flatten(2).transpose(1, 2)
+        cls_token = self.cls_embedding.expand(x.shape[0], -1, -1)
+        x = torch.cat((cls_token, x), dim=1)
+        x += self.position_embedding.weight.unsqueeze(0)
+        return x
+
+
+class Attention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.num_heads = config.num_heads
+        head_dim = config.hidden_size // config.num_heads
+        self.scale = head_dim ** -0.5
+        self.query_key_value = nn.Linear(config.hidden_size, config.hidden_size * 3)
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.output_dropout = torch.nn.Dropout(config.dropout_prob)
+
+    def forward(self, x: "tensor(B, L, D)") -> "tensor(B, L, D)":
+        B, L, _ = x.shape
+        qkv = self.query_key_value(x)
+        qkv = qkv.reshape(B, L, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)  # 3, B, H, L, D
+        q, k, v = qkv[0], qkv[1], qkv[2]
+
+        out = attention_fn_default(
+            q, k, v
+        ) #  24 x 3 x 
+        out = out.transpose(2, 1)
+        # breakpoint()
+        # output = self.dense(out.reshape(B, L, -1))
+        output = self.dense(out.view(B, L, -1))
+        output = self.output_dropout(output)
+        return output
+
+    def attention(self, q, k, v):
+        attn_weights = torch.matmul(q * self.scale, k.transpose(-2, -1))
+        attn_weights = attn_weights.softmax(dim=-1)
+        output = torch.matmul(attn_weights, v)
+        return output
+
+
+class MLP(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, x: torch.Tensor) -> torch.Tensor:
+        x = self.fc1(x)
+        x = self.activation_fn(x)
+        x = self.fc2(x)
+        return x
+
+
+class TransformerLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.attention = Attention(config)
+        self.mlp = MLP(config)
+        self.post_attention_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+    def forward(self, hidden_states):
+        attention_input = hidden_states
+        attention_output = self.input_layernorm(self.attention(attention_input))
+        hidden_states = attention_input + attention_output
+        mlp_input = hidden_states
+        mlp_output = self.post_attention_layernorm(self.mlp(mlp_input))
+        output = mlp_input + mlp_output
+        return output
+
+
+class Transformer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.layers = nn.ModuleList([TransformerLayer(config) for _ in range(config.num_hidden_layers)])
+
+    def forward(self, hidden_states):
+        for layer_module in self.layers:
+            hidden_states = layer_module(hidden_states)
+        return hidden_states
+
+
+class GLU(nn.Module):
+    def __init__(self, config, in_features):
+        super().__init__()
+        self.linear_proj = nn.Linear(in_features, config.hidden_size, bias=False)
+        self.norm1 = nn.LayerNorm(config.hidden_size)
+        self.act1 = nn.GELU()
+        self.act2 = nn.functional.silu
+        self.dense_h_to_4h = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
+        self.gate_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
+        self.dense_4h_to_h = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
+
+    def forward(self, x):
+        x = self.linear_proj(x)
+        x = self.act1(self.norm1(x))
+        x = self.act2(self.gate_proj(x)) * self.dense_h_to_4h(x)
+        x = self.dense_4h_to_h(x)
+        return x
+
+
+class EVA2CLIPModel(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        vision_config = Namespace(**config.vision_config)
+        self.patch_embedding = PatchEmbedding(vision_config)
+        self.transformer = Transformer(vision_config)
+        self.linear_proj = GLU(config, in_features=vision_config.hidden_size)
+        self.conv = nn.Conv2d(in_channels=vision_config.hidden_size, out_channels=vision_config.hidden_size, kernel_size=2, stride=2)
+        self.boi = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
+        self.eoi = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
+
+    def forward(self, images: "tensor(B, C, H, W)") -> "tensor(B, L, D)":
+        x = self.patch_embedding(images)
+        x = self.transformer(x)
+        x = x[:, 1:]
+        b, s, h = x.shape
+        grid_size = int(s**0.5)
+        x = x.view(b, grid_size, grid_size, h).permute(0, 3, 1, 2)
+        x = self.conv(x)
+
+        x = x.flatten(2).transpose(1, 2)
+        x = self.linear_proj(x)
+        boi = self.boi.expand(x.shape[0], -1, -1)
+        eoi = self.eoi.expand(x.shape[0], -1, -1)
+        x = torch.cat((boi, x, eoi), dim=1)
+        return x