update .py file

configuration_qwen2.py

@@ -0,0 +1,169 @@
+# coding=utf-8
+# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Qwen2 model configuration"""
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+QWEN2_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "Qwen/Qwen2-7B-beta": "https://huggingface.co/Qwen/Qwen2-7B-beta/resolve/main/config.json",
+}
+
+
+class Qwen2Config(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Qwen2Model`]. It is used to instantiate a
+    Qwen2 model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of
+    Qwen2-7B-beta [Qwen/Qwen2-7B-beta](https://huggingface.co/Qwen/Qwen2-7B-beta).
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 151936):
+            Vocabulary size of the Qwen2 model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`Qwen2Model`]
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 22016):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_key_value_heads (`int`, *optional*, defaults to 32):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `32`.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 32768):
+            The maximum sequence length that this model might ever be used with.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied.
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        use_sliding_window (`bool`, *optional*, defaults to `False`):
+            Whether to use sliding window attention.
+        sliding_window (`int`, *optional*, defaults to 4096):
+            Sliding window attention (SWA) window size. If not specified, will default to `4096`.
+        max_window_layers (`int`, *optional*, defaults to 28):
+            The number of layers that use SWA (Sliding Window Attention). The bottom layers use SWA while the top use full attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+
+    ```python
+    >>> from transformers import Qwen2Model, Qwen2Config
+
+    >>> # Initializing a Qwen2 style configuration
+    >>> configuration = Qwen2Config()
+
+    >>> # Initializing a model from the Qwen2-7B style configuration
+    >>> model = Qwen2Model(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "qwen2"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=151936,
+        hidden_size=4096,
+        intermediate_size=22016,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=32,
+        hidden_act="silu",
+        max_position_embeddings=32768,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        use_sliding_window=False,
+        sliding_window=4096,
+        rope_scaling=None,
+        max_window_layers=28,
+        attention_dropout=0.0,
+        beacon_window=1024,
+        beacon_stride=1024,
+        beacon_attn="full-coverage",
+        beacon_ratio=[2,4,8,16,32],
+        beacon_ratio_mix="step-random",
+        beacon_param=[],
+        beacon_embed_init="eos",
+        beacon_sink_size=0,
+        beacon_attend_prev=True,
+        beacon_pos="interleave",
+        beacon_parallel_window=1,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.use_sliding_window = use_sliding_window
+        self.sliding_window = sliding_window
+        self.max_window_layers = max_window_layers
+        self.rope_scaling = rope_scaling
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.attention_dropout = attention_dropout
+
+        self.beacon_window = beacon_window
+        self.beacon_stride = beacon_stride
+        self.beacon_attn = beacon_attn
+        self.beacon_ratio = beacon_ratio
+        self.beacon_ratio_mix = beacon_ratio_mix
+        self.beacon_param = beacon_param
+        self.beacon_embed_init = beacon_embed_init
+        self.beacon_sink_size = beacon_sink_size
+        self.beacon_attend_prev = beacon_attend_prev
+        self.beacon_pos = beacon_pos
+        self.beacon_parallel_window = beacon_parallel_window
+
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

constants.py

@@ -0,0 +1,12 @@
+CONTROLLER_HEART_BEAT_EXPIRATION = 30
+WORKER_HEART_BEAT_INTERVAL = 15
+
+LOGDIR = "."
+
+# Model Constants
+IGNORE_INDEX = -100
+IMAGE_TOKEN_INDEX = -200
+DEFAULT_IMAGE_TOKEN = "<image>"
+DEFAULT_IMAGE_PATCH_TOKEN = "<im_patch>"
+DEFAULT_IM_START_TOKEN = "<im_start>"
+DEFAULT_IM_END_TOKEN = "<im_end>"

conversation.py

@@ -0,0 +1,554 @@
+import dataclasses
+from enum import auto, Enum
+from typing import List, Any, Dict, Union, Tuple
+import re
+import base64
+from io import BytesIO
+from PIL import Image
+from transformers import AutoTokenizer
+
+
+class SeparatorStyle(Enum):
+    """Different separator style."""
+
+    SINGLE = auto()
+    TWO = auto()
+    MPT = auto()
+    PLAIN = auto()
+    CHATML = auto()
+    LLAMA_2 = auto()
+    LLAMA_3 = auto()
+    QWEN = auto()
+    GEMMA = auto()
+
+
+@dataclasses.dataclass
+class Conversation:
+    """A class that keeps all conversation history."""
+
+    system: str
+    roles: List[str]
+    messages: List[List[str]]
+    offset: int
+    sep_style: SeparatorStyle = SeparatorStyle.SINGLE
+    sep: str = "###"
+    sep2: str = None
+    version: str = "Unknown"
+
+    tokenizer_id: str = ""
+    tokenizer: Any = None
+    # Stop criteria (the default one is EOS token)
+    stop_str: Union[str, List[str]] = None
+    # Stops generation if meeting any token in this list
+    stop_token_ids: List[int] = None
+
+    skip_next: bool = False
+
+    def get_prompt(self):
+        messages = self.messages
+        if len(messages) > 0 and type(messages[0][1]) is tuple:
+            messages = self.messages.copy()
+            init_role, init_msg = messages[0].copy()
+            init_msg = init_msg[0]
+            if "mmtag" in self.version:
+                init_msg = init_msg.replace("<image>", "").strip()
+                messages[0] = (init_role, init_msg)
+                messages.insert(0, (self.roles[0], "<Image><image></Image>"))
+                messages.insert(1, (self.roles[1], "Received."))
+            elif not init_msg.startswith("<image>"):
+                init_msg = init_msg.replace("<image>", "").strip()
+                messages[0] = (init_role, "<image>\n" + init_msg)
+            else:
+                messages[0] = (init_role, init_msg)
+
+        if self.sep_style == SeparatorStyle.SINGLE:
+            ret = self.system + self.sep
+            for role, message in messages:
+                if message:
+                    if type(message) is tuple:
+                        message, _, _ = message
+                    ret += role + ": " + message + self.sep
+                else:
+                    ret += role + ":"
+
+        elif self.sep_style == SeparatorStyle.TWO:
+            seps = [self.sep, self.sep2]
+            ret = self.system + seps[0]
+            for i, (role, message) in enumerate(messages):
+                if message:
+                    if type(message) is tuple:
+                        message, _, _ = message
+                    ret += role + ": " + message + seps[i % 2]
+                else:
+                    ret += role + ":"
+
+        elif self.sep_style == SeparatorStyle.CHATML:
+            ret = "" if self.system == "" else self.system + self.sep + "\n"
+            for role, message in messages:
+                if message:
+                    if type(message) is tuple:
+                        message, images = message
+                        message = "<image>" * len(images) + message
+                    ret += role + "\n" + message + self.sep + "\n"
+                else:
+                    ret += role + "\n"
+            return ret
+
+        elif self.sep_style == SeparatorStyle.LLAMA_3:
+            chat_template_messages = [{"role": "system", "content": self.system}]
+            for role, message in messages:
+                if message:
+                    if type(message) is tuple:
+                        message, images = message
+                        message = "<image>" * len(images) + message
+                    chat_template_messages.append({"role": role, "content": message})
+
+            # print(chat_template_messages)
+            return self.tokenizer.apply_chat_template(chat_template_messages, tokenize=False, add_generation_prompt=True)
+            # ret = "" if self.system == "" else self.system + self.sep + "\n"
+            # for role, message in messages:
+            #     if message:
+            #         if type(message) is tuple:
+            #             message, images = message
+            #             message = "<image>" * len(images) + message
+            #         ret += role + "\n" + message + self.sep + "\n"
+            #     else:
+            #         ret += role + "\n"
+            # return ret
+
+        elif self.sep_style == SeparatorStyle.MPT:
+            ret = self.system + self.sep
+            for role, message in messages:
+                if message:
+                    if type(message) is tuple:
+                        message, _, _ = message
+                    ret += role + message + self.sep
+                else:
+                    ret += role
+
+        elif self.sep_style == SeparatorStyle.GEMMA:
+            ret = ""
+            for i, (role, message) in enumerate(messages):
+                assert role == self.roles[i % 2], "Conversation should alternate user/assistant/user/assistant/..."
+                if message:
+                    if type(message) is tuple:
+                        message, _, _ = message
+                    ret += role + message + self.sep
+                else:
+                    ret += role
+
+        elif self.sep_style == SeparatorStyle.LLAMA_2:
+            wrap_sys = lambda msg: f"<<SYS>>\n{msg}\n<</SYS>>\n\n" if len(msg) > 0 else msg
+            wrap_inst = lambda msg: f"[INST] {msg} [/INST]"
+            ret = ""
+
+            for i, (role, message) in enumerate(messages):
+                if i == 0:
+                    assert message, "first message should not be none"
+                    assert role == self.roles[0], "first message should come from user"
+                if message:
+                    if type(message) is tuple:
+                        message, _, _ = message
+                    if i == 0:
+                        message = wrap_sys(self.system) + message
+                    if i % 2 == 0:
+                        message = wrap_inst(message)
+                        ret += self.sep + message
+                    else:
+                        ret += " " + message + " " + self.sep2
+                else:
+                    ret += ""
+            ret = ret.lstrip(self.sep)
+
+        elif self.sep_style == SeparatorStyle.PLAIN:
+            seps = [self.sep, self.sep2]
+            ret = self.system
+            for i, (role, message) in enumerate(messages):
+                if message:
+                    if type(message) is tuple:
+                        message, _, _ = message
+                    ret += message + seps[i % 2]
+                else:
+                    ret += ""
+        else:
+            raise ValueError(f"Invalid style: {self.sep_style}")
+
+        return ret
+
+    def append_message(self, role, message):
+        self.messages.append([role, message])
+
+    def process_image(self, image, image_process_mode, return_pil=False, image_format="PNG"):
+        if image_process_mode == "Pad":
+
+            def expand2square(pil_img, background_color=(122, 116, 104)):
+                width, height = pil_img.size
+                if width == height:
+                    return pil_img
+                elif width > height:
+                    result = Image.new(pil_img.mode, (width, width), background_color)
+                    result.paste(pil_img, (0, (width - height) // 2))
+                    return result
+                else:
+                    result = Image.new(pil_img.mode, (height, height), background_color)
+                    result.paste(pil_img, ((height - width) // 2, 0))
+                    return result
+
+            image = expand2square(image)
+        elif image_process_mode in ["Default", "Crop"]:
+            pass
+        elif image_process_mode == "Resize":
+            image = image.resize((336, 336))
+        else:
+            raise ValueError(f"Invalid image_process_mode: {image_process_mode}")
+
+        if type(image) is not Image.Image:
+            image = Image.open(image).convert("RGB")
+
+        max_hw, min_hw = max(image.size), min(image.size)
+        aspect_ratio = max_hw / min_hw
+        max_len, min_len = 672, 448
+        shortest_edge = int(min(max_len / aspect_ratio, min_len, min_hw))
+        longest_edge = int(shortest_edge * aspect_ratio)
+        W, H = image.size
+        if H > W:
+            H, W = longest_edge, shortest_edge
+        else:
+            H, W = shortest_edge, longest_edge
+        image = image.resize((W, H))
+        if return_pil:
+            return image
+        else:
+            buffered = BytesIO()
+            image.save(buffered, format=image_format)
+            img_b64_str = base64.b64encode(buffered.getvalue()).decode()
+            return img_b64_str
+
+    def get_images(self, return_pil=False, return_path=False):
+        images = []
+        for i, (role, msg) in enumerate(self.messages[self.offset :]):
+            if i % 2 == 0:
+                if type(msg) is tuple:
+                    msg, image, image_process_mode = msg
+                    if type(image) != list:
+                        image = [image]
+                    for img in image:
+                        if not return_path:
+                            img = self.process_image(img, image_process_mode, return_pil=return_pil)
+                        else:
+                            images.append(img)
+        return images
+
+    def to_gradio_chatbot(self):
+        ret = []
+        for i, (role, msg) in enumerate(self.messages[self.offset :]):
+            if i % 2 == 0:
+                if type(msg) is tuple:
+                    msg, image, image_process_mode = msg
+                    if type(image) != list:
+                        image = [image]
+                    if len(image) == 1:
+                        msg = "<image>\n" + msg.replace("<image>", "").strip()
+                    else:
+                        msg = re.sub(r"(<image>)\n(?=<image>)", r"\1 ", msg)
+                    for img in image:
+                        img_b64_str = self.process_image(img, "Default", return_pil=False, image_format="JPEG")
+                        img_str = f'<img src="data:image/jpeg;base64,{img_b64_str}"/>'
+                        msg = msg.replace("<image>", img_str, 1).strip()
+                    if len(msg) > 0:
+                        ret.append([msg, None])
+                else:
+                    ret.append([msg, None])
+            else:
+                ret[-1][-1] = msg
+        return ret
+
+    def copy(self):
+        return Conversation(system=self.system, roles=self.roles, messages=[[x, y] for x, y in self.messages], offset=self.offset, sep_style=self.sep_style, sep=self.sep, sep2=self.sep2, version=self.version)
+
+    def dict(self):
+        if len(self.get_images()) > 0:
+            return {
+                "system": self.system,
+                "roles": self.roles,
+                "messages": [[x, y[0] if type(y) is tuple else y] for x, y in self.messages],
+                "offset": self.offset,
+                "sep": self.sep,
+                "sep2": self.sep2,
+            }
+        return {
+            "system": self.system,
+            "roles": self.roles,
+            "messages": self.messages,
+            "offset": self.offset,
+            "sep": self.sep,
+            "sep2": self.sep2,
+        }
+
+
+conv_vicuna_v0 = Conversation(
+    system="A chat between a curious human and an artificial intelligence assistant. " "The assistant gives helpful, detailed, and polite answers to the human's questions.",
+    roles=("Human", "Assistant"),
+    messages=[
+        ["Human", "What are the key differences between renewable and non-renewable energy sources?"],
+        [
+            "Assistant",
+            "Renewable energy sources are those that can be replenished naturally in a relatively "
+            "short amount of time, such as solar, wind, hydro, geothermal, and biomass. "
+            "Non-renewable energy sources, on the other hand, are finite and will eventually be "
+            "depleted, such as coal, oil, and natural gas. Here are some key differences between "
+            "renewable and non-renewable energy sources:\n"
+            "1. Availability: Renewable energy sources are virtually inexhaustible, while non-renewable "
+            "energy sources are finite and will eventually run out.\n"
+            "2. Environmental impact: Renewable energy sources have a much lower environmental impact "
+            "than non-renewable sources, which can lead to air and water pollution, greenhouse gas emissions, "
+            "and other negative effects.\n"
+            "3. Cost: Renewable energy sources can be more expensive to initially set up, but they typically "
+            "have lower operational costs than non-renewable sources.\n"
+            "4. Reliability: Renewable energy sources are often more reliable and can be used in more remote "
+            "locations than non-renewable sources.\n"
+            "5. Flexibility: Renewable energy sources are often more flexible and can be adapted to different "
+            "situations and needs, while non-renewable sources are more rigid and inflexible.\n"
+            "6. Sustainability: Renewable energy sources are more sustainable over the long term, while "
+            "non-renewable sources are not, and their depletion can lead to economic and social instability.\n",
+        ],
+    ],
+    offset=2,
+    sep_style=SeparatorStyle.SINGLE,
+    sep="###",
+)
+
+conv_vicuna_v1 = Conversation(
+    system="A chat between a curious user and an artificial intelligence assistant. " "The assistant gives helpful, detailed, and polite answers to the user's questions.",
+    roles=("USER", "ASSISTANT"),
+    version="v1",
+    messages=[],
+    offset=0,
+    sep_style=SeparatorStyle.TWO,
+    sep=" ",
+    sep2="</s>",
+)
+
+conv_llama_2 = Conversation(
+    system="""You are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe.  Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature.
+
+If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information.""",
+    roles=("USER", "ASSISTANT"),
+    version="llama_v2",
+    messages=[],
+    offset=0,
+    sep_style=SeparatorStyle.LLAMA_2,
+    sep="<s>",
+    sep2="</s>",
+)
+
+conv_llava_llama_2 = Conversation(
+    system="You are a helpful language and vision assistant. " "You are able to understand the visual content that the user provides, " "and assist the user with a variety of tasks using natural language.",
+    roles=("USER", "ASSISTANT"),
+    version="llama_v2",
+    messages=[],
+    offset=0,
+    sep_style=SeparatorStyle.LLAMA_2,
+    sep="<s>",
+    sep2="</s>",
+)
+
+# This will lead to a bug when user can not access Meta-Llama-3-8B-Instruct
+# conv_llava_llama_3 = Conversation(
+#     system="You are a helpful language and vision assistant. " "You are able to understand the visual content that the user provides, " "and assist the user with a variety of tasks using natural language.",
+#     roles=("user", "assistant"),
+#     version="llama_v3",
+#     messages=[],
+#     offset=0,
+#     sep="<|eot_id|>",
+#     sep_style=SeparatorStyle.LLAMA_3,
+#     tokenizer_id="meta-llama/Meta-Llama-3-8B-Instruct",
+#     tokenizer=AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3-8B-Instruct"),
+#     stop_token_ids=[128009],
+# )
+
+conv_mistral_instruct = Conversation(
+    system="",
+    roles=("USER", "ASSISTANT"),
+    version="llama_v2",
+    messages=[],
+    offset=0,
+    sep_style=SeparatorStyle.LLAMA_2,
+    sep="",
+    sep2="</s>",
+)
+
+conv_llava_llama_2_simple = Conversation(
+    system="Answer the questions about the visual content that the user provides.",
+    roles=("USER", "ASSISTANT"),
+    version="llama_v2",
+    messages=[],
+    offset=0,
+    sep_style=SeparatorStyle.LLAMA_2,
+    sep="<s>",
+    sep2="</s>",
+)
+
+conv_llava_llama_2_mmtag = Conversation(
+    system="Answer the questions about the visual content that the user provides." "The visual content will be provided with the following format: <Image>visual content</Image>.",
+    roles=("USER", "ASSISTANT"),
+    version="llama_v2_mmtag",
+    messages=[],
+    offset=0,
+    sep_style=SeparatorStyle.LLAMA_2,
+    sep="<s>",
+    sep2="</s>",
+)
+
+conv_mpt = Conversation(
+    system="""<|im_start|>system
+A conversation between a user and an LLM-based AI assistant. The assistant gives helpful and honest answers.""",
+    roles=("<|im_start|>user\n", "<|im_start|>assistant\n"),
+    version="mpt",
+    messages=[],
+    offset=0,
+    sep_style=SeparatorStyle.MPT,
+    sep="<|im_end|>",
+)
+
+conv_qwen = Conversation(
+    system="""<|im_start|>system
+You are a helpful assistant.""",
+    roles=("<|im_start|>user", "<|im_start|>assistant"),
+    version="qwen",
+    messages=[],
+    offset=0,
+    sep_style=SeparatorStyle.CHATML,
+    sep="<|im_end|>",
+)
+
+conv_gemma_instruct = Conversation(system="", roles=("<start_of_turn>user\n", "<start_of_turn>model\n"), version="gemma", messages=[], offset=0, sep_style=SeparatorStyle.GEMMA, sep="<end_of_turn>\n")
+
+conv_llava_plain = Conversation(
+    system="",
+    roles=("", ""),
+    messages=[],
+    offset=0,
+    sep_style=SeparatorStyle.PLAIN,
+    sep="\n",
+)
+
+conv_llava_v0 = Conversation(
+    system="A chat between a curious human and an artificial intelligence assistant. " "The assistant gives helpful, detailed, and polite answers to the human's questions.",
+    roles=("Human", "Assistant"),
+    messages=[],
+    offset=0,
+    sep_style=SeparatorStyle.SINGLE,
+    sep="###",
+)
+
+conv_llava_v0_mmtag = Conversation(
+    system="A chat between a curious user and an artificial intelligence assistant. "
+    "The assistant is able to understand the visual content that the user provides, and assist the user with a variety of tasks using natural language."
+    "The visual content will be provided with the following format: <Image>visual content</Image>.",
+    roles=("Human", "Assistant"),
+    messages=[],
+    offset=0,
+    sep_style=SeparatorStyle.SINGLE,
+    sep="###",
+    version="v0_mmtag",
+)
+
+conv_llava_v1 = Conversation(
+    system="A chat between a curious human and an artificial intelligence assistant. " "The assistant gives helpful, detailed, and polite answers to the human's questions.",
+    roles=("USER", "ASSISTANT"),
+    version="v1",
+    messages=[],
+    offset=0,
+    sep_style=SeparatorStyle.TWO,
+    sep=" ",
+    sep2="</s>",
+)
+
+conv_llava_v1_mmtag = Conversation(
+    system="A chat between a curious user and an artificial intelligence assistant. "
+    "The assistant is able to understand the visual content that the user provides, and assist the user with a variety of tasks using natural language."
+    "The visual content will be provided with the following format: <Image>visual content</Image>.",
+    roles=("USER", "ASSISTANT"),
+    messages=[],
+    offset=0,
+    sep_style=SeparatorStyle.TWO,
+    sep=" ",
+    sep2="</s>",
+    version="v1_mmtag",
+)
+
+conv_mistral_orca = Conversation(
+    system="""<|im_start|>system
+You are MistralOrca, a large language model trained by Alignment Lab AI. Write out your reasoning step-by-step to be sure you get the right answers!""",
+    roles=("<|im_start|>user\n", "<|im_start|>assistant\n"),
+    version="mpt",
+    messages=[],
+    offset=0,
+    sep_style=SeparatorStyle.MPT,
+    sep="<|im_end|>",
+)
+
+conv_mistral_zephyr = Conversation(
+    system="""<|system|>
+You are a helpful AI assistant.""",
+    roles=("<|user|>\n", "<|assistant|>\n"),
+    version="mpt",
+    messages=[],
+    offset=0,
+    sep_style=SeparatorStyle.MPT,
+    sep="</s>",
+)
+
+conv_mistral_direct = Conversation(
+    system="""<|im_start|>system
+Answer the questions.""",
+    roles=("<|im_start|>user\n", "<|im_start|>assistant\n"),
+    version="mpt",
+    messages=[],
+    offset=0,
+    sep_style=SeparatorStyle.MPT,
+    sep="<|im_end|>",
+)
+
+conv_chatml_direct = Conversation(
+    system="""<|im_start|>system
+Answer the questions.""",
+    roles=("<|im_start|>user\n", "<|im_start|>assistant\n"),
+    version="mpt",
+    messages=[],
+    offset=0,
+    sep_style=SeparatorStyle.MPT,
+    sep="<|im_end|>",
+)
+
+default_conversation = conv_vicuna_v0
+conv_templates = {
+    "default": conv_vicuna_v0,
+    "v0": conv_vicuna_v0,
+    "v1": conv_vicuna_v1,
+    "vicuna_v1": conv_vicuna_v1,
+    "llama_2": conv_llama_2,
+    "mistral_instruct": conv_mistral_instruct,
+    "mistral_orca": conv_mistral_orca,
+    "mistral_zephyr": conv_mistral_zephyr,
+    "mistral_direct": conv_mistral_direct,
+    "plain": conv_llava_plain,
+    "v0_plain": conv_llava_plain,
+    "chatml_direct": conv_chatml_direct,
+    "llava_v0": conv_llava_v0,
+    "llava_v0_mmtag": conv_llava_v0_mmtag,
+    "llava_v1": conv_llava_v1,
+    "llava_v1_mmtag": conv_llava_v1_mmtag,
+    "llava_llama_2": conv_llava_llama_2,
+    "llava_llama_2_simple": conv_llava_llama_2_simple,
+    "llava_llama_2_mmtag": conv_llava_llama_2_mmtag,
+    "llava_mistral_instruct": conv_mistral_instruct,
+    "mpt": conv_mpt,
+    "qwen_1_5": conv_qwen,
+    "gemma_instruct": conv_gemma_instruct,
+}
+
+
+if __name__ == "__main__":
+    print(default_conversation.get_prompt())

llava_arch.py

@@ -0,0 +1,742 @@
+#    Copyright 2023 Haotian Liu
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+
+
+from abc import ABC, abstractmethod
+
+import math
+import re
+import time
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from .multimodal_encoder.builder import build_vision_tower
+from .multimodal_resampler.builder import build_vision_resampler
+from .multimodal_projector.builder import build_vision_projector
+from transformers import AutoTokenizer
+
+from longva.longva.constants import IGNORE_INDEX, IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_PATCH_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+
+from longva.longva.mm_utils import get_anyres_image_grid_shape
+from longva.longva.utils import rank0_print
+import random
+from .sae import SiglipAE
+from .WindowTimeToTokenAttention import WindowTimeToTokenAttention
+import numpy as np
+import torch.nn.functional as F
+import pdb
+class LlavaMetaModel:
+
+    def __init__(self, config):
+        super(LlavaMetaModel, self).__init__(config)
+
+        if hasattr(config, "mm_vision_tower"):
+            delay_load = getattr(config, "delay_load", False)
+            self.vision_tower = build_vision_tower(config, delay_load=delay_load)
+            self.vision_resampler = build_vision_resampler(config, vision_tower=self.vision_tower)
+            self.mm_projector = build_vision_projector(config, vision_cfg=self.vision_tower.config)
+
+            if "unpad" in getattr(config, "mm_patch_merge_type", ""):
+                self.image_newline = nn.Parameter(torch.empty(config.hidden_size, dtype=self.dtype))
+    
+        # self.llm_tokenizer = AutoTokenizer.from_pretrained(config._name_or_path)
+        self.hidden_size=config.hidden_size
+        # print(config)
+        # exit(0)
+        
+#         self.text_tokenizer = T5Tokenizer.from_pretrained('google-t5/t5-small')
+##############################################################################
+#         self.text_select_model = T5EncoderModel.from_pretrained('google-t5/t5-small')
+        
+        # self.text_gamma=0.75
+
+###############################################################################
+        self.text_mlp=nn.Sequential(
+            nn.Linear(config.hidden_size,config.hidden_size),
+            nn.GELU(),
+        )
+        self.sae=SiglipAE()
+        #self.sae.load_state_dict(torch.load('/share/LXRlxr0_0/code/videoxl2/videoxl2/longva/longva/model/encoder.pth'),strict=False)
+        
+###############################################################################
+        # self.vision_select=nn.Parameter(
+        #         torch.randn((4, self.config.hidden_size), dtype=self.dtype)
+        # )
+##############################################################################
+        
+    def get_vision_tower(self):
+        vision_tower = getattr(self, "vision_tower", None)
+        if type(vision_tower) is list:
+            vision_tower = vision_tower[0]
+        return vision_tower
+
+    def initialize_vision_modules(self, model_args, fsdp=None):
+        vision_tower = model_args.vision_tower
+        mm_vision_select_layer = model_args.mm_vision_select_layer
+        mm_vision_select_feature = model_args.mm_vision_select_feature
+        pretrain_mm_mlp_adapter = model_args.pretrain_mm_mlp_adapter
+        mm_patch_merge_type = model_args.mm_patch_merge_type
+
+        self.config.mm_vision_tower = vision_tower
+        self.config.vision_tower_pretrained = getattr(model_args, "vision_tower_pretrained", "")
+
+        if self.get_vision_tower() is None:
+            vision_tower = build_vision_tower(model_args)
+            vision_resampler = build_vision_resampler(model_args, vision_tower=vision_tower)
+            for k, v in vision_resampler.config.items():
+                setattr(self.config, k, v)
+
+            if fsdp is not None and len(fsdp) > 0:
+                self.vision_tower = [vision_tower]
+                self.vision_resampler = [vision_resampler]
+            else:
+                self.vision_tower = vision_tower
+                self.vision_resampler = vision_resampler
+        else:
+            if fsdp is not None and len(fsdp) > 0:
+                vision_resampler = self.vision_resampler[0]
+                vision_tower = self.vision_tower[0]
+            else:
+                vision_resampler = self.vision_resampler
+                vision_tower = self.vision_tower
+            vision_tower.load_model()
+
+            # In case it is frozen by LoRA
+            for p in self.vision_resampler.parameters():
+                p.requires_grad = True
+
+        self.config.use_mm_proj = True
+        self.config.mm_projector_type = getattr(model_args, "mm_projector_type", "linear")
+        self.config.mm_hidden_size = getattr(vision_resampler, "hidden_size", vision_tower.hidden_size)
+        self.config.mm_vision_select_layer = mm_vision_select_layer
+        self.config.mm_vision_select_feature = mm_vision_select_feature
+        self.config.mm_patch_merge_type = mm_patch_merge_type
+        
+        self.sae=SiglipAE()
+        self.sae.load_state_dict(torch.load('/share/LXRlxr0_0/code/videoxl2/videoxl2/longva/longva/model/encoder.pth'),strict=False)
+        ##############################################################################
+#         self.vision_select=nn.Parameter(
+#                 torch.randn((30, self.config.hidden_size), dtype=self.dtype)
+#         )
+        
+#         #self.text_tokenizer = T5Tokenizer.from_pretrained('google-t5/t5-small')
+#         self.text_select_model = T5EncoderModel.from_pretrained('google-t5/t5-small')
+        
+#         self.text_mlp=nn.Sequential(
+#             nn.Linear(512,self.config.hidden_size),
+#             nn.GELU(),
+#             # nn.Linear(config.hidden_size,config.hidden_size),
+#             # nn.GELU(),
+#         )
+        ##############################################################################
+        
+        
+        if getattr(self, "mm_projector", None) is None:
+            self.mm_projector = build_vision_projector(self.config, vision_cfg=vision_tower.config)
+
+            if "unpad" in mm_patch_merge_type:
+                embed_std = 1 / torch.sqrt(torch.tensor(self.config.hidden_size, dtype=self.dtype))
+                self.image_newline = nn.Parameter(torch.randn(self.config.hidden_size, dtype=self.dtype) * embed_std)
+        else:
+            # In case it is frozen by LoRA
+            for p in self.mm_projector.parameters():
+                p.requires_grad = True
+
+        if pretrain_mm_mlp_adapter is not None:
+            mm_projector_weights = torch.load(pretrain_mm_mlp_adapter, map_location="cpu")
+
+            def get_w(weights, keyword):
+                return {k.split(keyword + ".")[1]: v for k, v in weights.items() if keyword in k}
+
+            incompatible_keys = self.mm_projector.load_state_dict(get_w(mm_projector_weights, "mm_projector"))
+            rank0_print(f"Loaded mm projector weights from {pretrain_mm_mlp_adapter}. Incompatible keys: {incompatible_keys}")
+            incompatible_keys = self.vision_resampler.load_state_dict(get_w(mm_projector_weights, "vision_resampler"), strict=False)
+            rank0_print(f"Loaded vision resampler weights from {pretrain_mm_mlp_adapter}. Incompatible keys: {incompatible_keys}")
+            
+            
+#             self.vision_select.data = mm_projector_weights["model.vision_select"]
+            
+#             self.text_mlp.load_state_dict(get_w(mm_projector_weights, "text_mlp"))
+            
+#             self.text_select_model.load_state_dict(get_w(mm_projector_weights, "text_select_model"),strict=False)
+            #self.vision_tower.load_state_dict(get_w(mm_projector_weights, "vision_tower"),strict=False)
+
+def unpad_image(tensor, original_size):
+    """
+    Unpads a PyTorch tensor of a padded and resized image.
+
+    Args:
+    tensor (torch.Tensor): The image tensor, assumed to be in CxHxW format.
+    original_size (tuple): The original size of the image (height, width).
+
+    Returns:
+    torch.Tensor: The unpadded image tensor.
+    """
+    original_width, original_height = original_size
+    current_height, current_width = tensor.shape[1:]
+
+    # Compute aspect ratios
+    original_aspect_ratio = original_width / original_height
+    current_aspect_ratio = current_width / current_height
+
+    # Determine padding size and direction
+    if original_aspect_ratio > current_aspect_ratio:
+        # Padding was added to the height
+        scale_factor = current_width / original_width
+        new_height = int(original_height * scale_factor)
+        padding = (current_height - new_height) // 2
+        unpadded_tensor = tensor[:, padding : current_height - padding, :]
+    else:
+        # Padding was added to the width
+        scale_factor = current_height / original_height
+        new_width = int(original_width * scale_factor)
+        padding = (current_width - new_width) // 2
+        unpadded_tensor = tensor[:, :, padding : current_width - padding]
+
+    return unpadded_tensor
+
+def rotary_position_embedding(q):
+    seq_len, dim = q.shape
+
+    position = torch.arange(seq_len, dtype=torch.float).unsqueeze(-1).to(q.device)
+
+    div_term = torch.exp(torch.arange(0, dim, 2, dtype=torch.float) * -(math.log(1000000.0) / dim)).to(q.device)
+    
+    pos_emb = position * div_term
+    pos_emb = torch.stack([torch.sin(pos_emb), torch.cos(pos_emb)], dim=-1).flatten(-2, -1)
+    
+    cos_emb = pos_emb[..., 1::2].repeat_interleave(2, dim=-1)
+    sin_emb = pos_emb[..., ::2].repeat_interleave(2, dim=-1)
+    
+    q_alternate = torch.stack([-q[..., 1::2], q[..., ::2]], dim=-1).reshape(q.size())
+    
+    q_rotated = q * cos_emb + q_alternate * sin_emb
+
+    return q_rotated
+
+class LlavaMetaForCausalLM(ABC):
+
+    @abstractmethod
+    def get_model(self):
+        pass
+
+    def get_vision_tower(self):
+        return self.get_model().get_vision_tower()
+
+    def get_2dPool(self, image_feature):
+        height = width = self.get_vision_tower().num_patches_per_side
+        num_frames, num_tokens, num_dim = image_feature.shape
+        image_feature = image_feature.view(num_frames, height, width, -1)
+        image_feature = image_feature.permute(0, 3, 1, 2).contiguous()
+        # image_feature = nn.functional.max_pool2d(image_feature, self.config.mm_spatial_pool_stride)
+        if self.config.mm_spatial_pool_mode == "average":
+            image_feature = nn.functional.avg_pool2d(image_feature, self.config.mm_spatial_pool_stride)
+        elif self.config.mm_spatial_pool_mode == "max":
+            image_feature = nn.functional.max_pool2d(image_feature, self.config.mm_spatial_pool_stride)
+        else:
+            raise ValueError(f"Unexpected mm_spatial_pool_mode: {self.config.mm_spatial_pool_mode}")
+        image_feature = image_feature.permute(0, 2, 3, 1)
+        image_feature = image_feature.view(num_frames, -1, num_dim)
+        return image_feature
+
+    def encode_images(self, images):
+        image_features = self.get_model().get_vision_tower()(images)
+        #image_features = self.get_model().vision_resampler(image_features, images=images)
+        image_features = self.get_model().mm_projector(image_features)
+        image_features = self.get_model().vision_resampler(image_features, images=images)
+        return image_features
+
+    def add_image(self, image_features):
+        return torch.repeat_interleave(image_features, repeats=4, dim=0)
+    
+    def add_video(self, video_features):
+        if video_features.size(0)<4:
+            last_feature = video_features[-1:]
+
+            repeated_features = last_feature.repeat(4 - video_features.size(0), 1,1,1)
+            expanded_x = torch.cat([video_features, repeated_features], dim=0)
+            return expanded_x
+        
+        repeat_counts = torch.ones(video_features.size(0), dtype=torch.long, device=video_features.device)
+
+        sum_counts=torch.sum(repeat_counts)
+        if sum_counts % 4!=0:
+            padding_size = 4 - (sum_counts % 4)
+            random_indices = torch.randperm(repeat_counts.size(0))[:padding_size].to(video_features.device)
+            repeat_counts[random_indices] += 1 
+            
+        expanded_x = torch.repeat_interleave(video_features, repeat_counts, dim=0)
+        
+        return expanded_x
+
+    def encode_multimodals(self, videos_or_images, video_idx_in_batch, split_sizes=None):
+        #################################################################################
+        # if videos_or_images.shape[0] > 360:
+        #     random_indices = np.random.choice(videos_or_images.shape[0], size=360, replace=False)
+        #     videos_or_images = videos_or_images[random_indices]
+        #     split_sizes=videos_or_images.shape[0]
+            
+        #################################################################################
+        # Define the maximum batch size (1024 frames)
+        max_batch_size = 60
+        num_frames = videos_or_images.shape[0]
+        # Initialize a list to store the features from each batch
+        videos_or_images_features = []
+
+        # Split videos_or_images into smaller batches if num_frames > max_batch_size
+        if num_frames > max_batch_size:
+            # Calculate the number of batches needed
+            num_batches = (num_frames + max_batch_size - 1) // max_batch_size
+            for i in range(num_batches):
+                start_idx = i * max_batch_size
+                end_idx = min((i + 1) * max_batch_size, num_frames)
+
+                # Process each batch separately
+                batch_videos_or_images = videos_or_images[start_idx:end_idx]
+                batch_features = self.get_model().get_vision_tower()(batch_videos_or_images)
+                videos_or_images_features.append(batch_features)
+
+            # Concatenate the features of all batches
+            videos_or_images_features = torch.cat(videos_or_images_features, dim=0)
+        else:
+            videos_or_images_features = self.get_model().get_vision_tower()(videos_or_images)
+
+        per_videos_or_images_features = torch.split(videos_or_images_features, split_sizes, dim=0)  # tuple, (dim_1, 576, 4096)
+        all_videos_or_images_features = []
+        
+        
+        for idx, feat in enumerate(per_videos_or_images_features):
+            #print(feat.shape,end='1\n')
+            feat=self.interpolate(feat)
+            #######################################################
+            if idx in video_idx_in_batch:
+                feat=self.add_video(feat)
+            else:
+                feat=self.add_image(feat)
+
+            bc,ch,h,w=feat.shape
+            
+            feat = feat.view(bc//4,ch,4,h,w)
+            if bc//4>24:
+                chunk_size = 24
+                chunks = torch.split(feat, chunk_size, dim=0)
+                interpolated_chunks = []
+                for chunk in chunks:
+                    interpolated_chunk=self.get_model().sae(chunk).squeeze(2)
+                    interpolated_chunks.append(interpolated_chunk)
+                feat = torch.cat(interpolated_chunks, dim=0)
+                del interpolated_chunks
+                del chunks
+            else:
+                feat=self.get_model().sae(feat).squeeze(2)
+            feat = feat.permute(0, 2, 3, 1).contiguous().flatten(1, 2)
+            #print(feat.shape,end='3\n')
+            feat = self.get_model().mm_projector(feat)
+            #print(feat.shape,end='4\n')
+            # Post pooling
+            if idx in video_idx_in_batch:
+                #print('************************',idx,video_idx_in_batch)
+                feat = self.get_2dPool(feat)
+            all_videos_or_images_features.append(feat)
+            
+        del per_videos_or_images_features
+        return all_videos_or_images_features
+    ########################################################
+    def interpolate(self,image_features):
+        b, num_tokens, dim = image_features.shape
+        
+        #print(str(image_features.shape)+' i\n')
+        
+        target_h = target_w = int(576**0.5)
+        h = w = int(num_tokens**0.5)
+
+        image_features = image_features.view(b, h, w, dim)
+        image_features = image_features.permute(0, 3, 1, 2).contiguous()
+
+        chunk_size = 24
+        chunks = torch.split(image_features, chunk_size, dim=0)
+        interpolated_chunks = []
+        for chunk in chunks:
+            interpolated_chunk = F.interpolate(
+                chunk.to(torch.float32),
+                size=(target_h, target_w),
+                mode="bilinear",
+                align_corners=False,
+            ).to(chunk.dtype)
+            interpolated_chunks.append(interpolated_chunk)
+        image_features = torch.cat(interpolated_chunks, dim=0)
+        del interpolated_chunks
+        
+        del chunks
+
+        return image_features
+
+    def prepare_inputs_labels_for_multimodal(self, input_ids, position_ids, attention_mask, past_key_values, labels, images, modalities=["image"], image_sizes=None,time_embedding=None):
+        vision_tower = self.get_vision_tower()
+        if vision_tower is None or images is None or input_ids.shape[1] == 1:
+            return input_ids, position_ids, attention_mask, past_key_values, None, labels
+
+        if type(images) is list or images.ndim == 5:
+            if type(images) is list:
+                images = [x.unsqueeze(0) if x.ndim == 3 else x for x in images]
+
+            video_idx_in_batch = []
+            for _ in range(len(modalities)):
+                if modalities[_] == "video":
+                    video_idx_in_batch.append(_)
+
+            images_list = []
+            for image in images:
+                if image.ndim == 4:
+                    images_list.append(image)
+                else:
+                    images_list.append(image.unsqueeze(0))
+            #print(len(images_list),images_list[0].shape)
+
+            concat_images = torch.cat([image for image in images_list], dim=0)
+            split_sizes = [image.shape[0] for image in images_list]
+
+            image_features = self.encode_multimodals(concat_images, video_idx_in_batch, split_sizes)    #16,144,3584
+
+            mm_patch_merge_type = getattr(self.config, "mm_patch_merge_type", "flat")
+            image_aspect_ratio = getattr(self.config, "image_aspect_ratio", "square")
+
+            visual_drop_score=[]
+            new_image_features=[]
+            
+            if mm_patch_merge_type == "flat":
+                
+                if image_features[0].ndim>2:
+                    image_features = [x.flatten(0, 1) for x in image_features]
+            elif mm_patch_merge_type== "unires":
+                #print('unires')
+                for image_idx, image_feature in enumerate(image_features):
+                    # rank0_print(f"Initial feature size : {image_feature.shape}")
+                    if image_idx in video_idx_in_batch:  # video operations
+                        #print(image_feature.shape)
+                        image_feature = image_feature.flatten(0, 1)
+                        
+                    elif image_feature.shape[0] > 1:
+                        # base image feature is never used in unires
+                        base_image_feature = image_feature[0]
+                        image_feature = image_feature[1:]
+
+                        height = width = self.get_vision_tower().num_patches_per_side
+                        assert height * width == base_image_feature.shape[0]
+
+                        kernel_size = mm_patch_merge_type.split("avgpool")[-1].split("x")[-1]
+                        kernel_size = 2
+                        image_feature = image_feature.view(image_feature.shape[0], height, width, -1) # [4, 24, 24, 4096]
+                        image_feature = image_feature.permute(0, 3, 1, 2).contiguous() # [4, 4096, 24, 24]
+                        image_feature = nn.functional.avg_pool2d(image_feature,kernel_size) # [4, 4096, 12, 12]
+                        image_feature = image_feature.flatten(2, 3) # [4, 4096, 144]
+                        image_feature = image_feature.permute(0, 2, 1).contiguous() # [4, 144, 4096]
+                        
+                        #print(image_feature.shape)
+                        image_feature = image_feature.flatten(0, 1)
+                        
+                    else:
+                        
+                        image_feature = image_feature[0]
+                        
+                    new_image_features.append(image_feature)
+                
+                image_features = new_image_features
+                
+            elif mm_patch_merge_type.startswith("spatial"):
+                new_image_features = []
+                for image_idx, image_feature in enumerate(image_features):
+                    # FIXME: now assume the image is square, and split to 2x2 patches
+                    # num_patches = h * w, where h = w = sqrt(num_patches)
+                    # currently image_feature is a tensor of shape (4, num_patches, hidden_size)
+                    # we want to first unflatten it to (2, 2, h, w, hidden_size)
+                    if image_idx in video_idx_in_batch:  # video operations
+                        if "unpad" in mm_patch_merge_type:
+                            # image_feature = image_feature.permute(2, 0, 1).contiguous()
+                            # image_feature =  torch.cat((image_feature, self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1).to(image_feature.device)), dim=-1)
+                            # image_feature = image_feature.permute(1, 2, 0).contiguous()
+                            image_feature = image_feature.flatten(0, 1)
+                            image_feature = torch.cat((image_feature, self.model.image_newline[None].to(image_feature.device)), dim=0)
+
+                    elif image_feature.shape[0] > 1:  # multi patches and multi images operations
+                        base_image_feature = image_feature[0]
+                        image_feature = image_feature[1:]
+                        height = width = self.get_vision_tower().num_patches_per_side
+                        assert height * width == base_image_feature.shape[0]
+
+                        if "anyres_max" in image_aspect_ratio:
+                            matched_anyres_max_num_patches = re.match(r"anyres_max_(\d+)", image_aspect_ratio)
+                            if matched_anyres_max_num_patches:
+                                max_num_patches = int(matched_anyres_max_num_patches.group(1))
+
+                        if image_aspect_ratio == "anyres" or "anyres_max" in image_aspect_ratio:
+                            if hasattr(self.get_vision_tower(), "image_size"):
+                                vision_tower_image_size = self.get_vision_tower().image_size
+                            else:
+                                raise ValueError("vision_tower_image_size is not found in the vision tower.")
+                            num_patch_width, num_patch_height = get_anyres_image_grid_shape(image_sizes[image_idx], self.config.image_grid_pinpoints, vision_tower_image_size)
+                            image_feature = image_feature.view(num_patch_height, num_patch_width, height, width, -1)
+                        else:
+                            image_feature = image_feature.view(2, 2, height, width, -1)
+
+                        if "maxpool2x2" in mm_patch_merge_type:
+                            image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
+                            image_feature = image_feature.flatten(1, 2).flatten(2, 3)
+                            image_feature = nn.functional.max_pool2d(image_feature, 2)
+                            image_feature = image_feature.flatten(1, 2).transpose(0, 1)
+                        elif "unpad" in mm_patch_merge_type and "anyres_max" in image_aspect_ratio and matched_anyres_max_num_patches:
+                            unit = image_feature.shape[2]
+                            image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
+                            image_feature = image_feature.flatten(1, 2).flatten(2, 3)
+                            image_feature = unpad_image(image_feature, image_sizes[image_idx])
+                            c, h, w = image_feature.shape
+                            times = math.sqrt(h * w / (max_num_patches * unit**2))
+                            if times > 1.1:
+                                image_feature = image_feature[None]
+                                image_feature = nn.functional.interpolate(image_feature, [int(h // times), int(w // times)], mode="bilinear")[0]
+                            image_feature = torch.cat((image_feature, self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1).to(image_feature.device)), dim=-1)
+                            image_feature = image_feature.flatten(1, 2).transpose(0, 1)
+                        elif "unpad" in mm_patch_merge_type:
+                            image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
+                            image_feature = image_feature.flatten(1, 2).flatten(2, 3)
+                            image_feature = unpad_image(image_feature, image_sizes[image_idx])
+                            image_feature = torch.cat((image_feature, self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1).to(image_feature.device)), dim=-1)
+                            image_feature = image_feature.flatten(1, 2).transpose(0, 1)
+                        else:
+                            image_feature = image_feature.permute(0, 2, 1, 3, 4).contiguous()
+                            image_feature = image_feature.flatten(0, 3)
+                        if "nobase" in mm_patch_merge_type:
+                            pass
+                        else:
+                            image_feature = torch.cat((base_image_feature, image_feature), dim=0)
+                    else:  # single image operations
+                        image_feature = image_feature[0]
+                        if "unpad" in mm_patch_merge_type:
+                            image_feature = torch.cat((image_feature, self.model.image_newline[None]), dim=0)
+
+                    new_image_features.append(image_feature)
+                image_features = new_image_features
+            else:
+                raise ValueError(f"Unexpected mm_patch_merge_type: {self.config.mm_patch_merge_type}")
+        else:
+            error_message = """
+            Something is wrong with the input shape. Most likely, you did not wrap the image or video input in a list:
+            This is correct:
+                model.generate(input_ids, images=[video_tensor],  modalities=["video"], **gen_kwargs)
+                model.generate(input_ids, images=[image_tensor],  modalities=["image"], **gen_kwargs)
+            This is wrong:
+                model.generate(input_ids, images=video_tensor,  modalities=["video"], **gen_kwargs)
+                model.generate(input_ids, images=image_tensor,  modalities=["image"], **gen_kwargs)
+            """
+            raise ValueError(error_message)
+
+        #print(time_embedding[0].shape)
+        #video_token_indices=[]
+        for image_idx, image_feature in enumerate(image_features):
+            if time_embedding[image_idx] is not None:
+                mask = (time_embedding[image_idx] == 151654)
+                indices = torch.nonzero(mask).squeeze()
+
+                embed_token=self.get_model().embed_tokens(time_embedding[image_idx])
+                embed_token[indices]=image_features[image_idx]
+                
+                #video_token_indices.append(indices)
+                
+                image_features[image_idx]=embed_token
+    
+        if getattr(self.config, "tune_mm_mlp_adapter", False) and getattr(self.config, "mm_use_im_start_end", False):
+            raise NotImplementedError
+
+        # Let's just add dummy tensors if they do not exist,
+        # it is a headache to deal with None all the time.
+        # But it is not ideal, and if you have a better idea,
+        # please open an issue / submit a PR, thanks.
+        _labels = labels
+        _position_ids = position_ids
+        _attention_mask = attention_mask
+        if attention_mask is None:
+            attention_mask = torch.ones_like(input_ids, dtype=torch.bool)
+        else:
+            attention_mask = attention_mask.bool()
+        if position_ids is None:
+            position_ids = torch.arange(0, input_ids.shape[1], dtype=torch.long, device=input_ids.device)
+        if labels is None:
+            labels = torch.full_like(input_ids, IGNORE_INDEX)
+
+        # remove the padding using attention_mask -- FIXME
+        _input_ids = input_ids
+        input_ids = [cur_input_ids[cur_attention_mask] for cur_input_ids, cur_attention_mask in zip(input_ids, attention_mask)]
+        labels = [cur_labels[cur_attention_mask] for cur_labels, cur_attention_mask in zip(labels, attention_mask)]
+
+        new_input_embeds = []
+        new_labels = []
+        cur_image_idx = 0
+
+        for batch_idx, cur_input_ids in enumerate(input_ids):
+            num_images = (cur_input_ids == IMAGE_TOKEN_INDEX).sum()
+            #print(num_images)
+            if num_images>=2:
+                print(num_images,input_ids)
+            if num_images == 0:
+                cur_image_features = image_features[cur_image_idx]
+                cur_input_embeds_1 = self.get_model().embed_tokens(cur_input_ids)
+                cur_input_embeds = torch.cat([cur_input_embeds_1, cur_image_features[0:0]], dim=0)
+                new_input_embeds.append(cur_input_embeds)
+                new_labels.append(labels[batch_idx])
+                cur_image_idx += 1
+                continue
+
+            image_token_indices = [-1] + torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0].tolist() + [cur_input_ids.shape[0]]
+            #print(image_token_indices) #[-1, 14, 236]
+            cur_input_ids_noim = []
+            cur_labels = labels[batch_idx]
+            
+            # print(cur_input_ids)
+            # print(labels[batch_idx])
+            
+            cur_labels_noim = []
+            for i in range(len(image_token_indices) - 1):
+                cur_input_ids_noim.append(cur_input_ids[image_token_indices[i] + 1 : image_token_indices[i + 1]])
+                cur_labels_noim.append(cur_labels[image_token_indices[i] + 1 : image_token_indices[i + 1]])
+            split_sizes = [x.shape[0] for x in cur_labels_noim]
+            
+           #print(torch.cat(cur_input_ids_noim).shape,torch.cat(cur_input_ids_noim))
+            
+            cur_input_embeds = self.get_model().embed_tokens(torch.cat(cur_input_ids_noim))
+            cur_input_embeds_no_im = torch.split(cur_input_embeds, split_sizes, dim=0)
+            cur_new_input_embeds = []
+            cur_new_labels = []
+
+            for i in range(num_images + 1):
+                cur_new_input_embeds.append(cur_input_embeds_no_im[i])
+                cur_new_labels.append(cur_labels_noim[i])
+                if i < num_images:
+                    ##############
+                    cur_image_features = image_features[cur_image_idx]
+                    cur_image_idx += 1
+                    cur_new_input_embeds.append(cur_image_features)
+                    cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=cur_labels.device, dtype=cur_labels.dtype))
+
+            cur_new_input_embeds = [x.to(self.device) for x in cur_new_input_embeds]
+
+            # import pdb; pdb.set_trace()
+            cur_new_input_embeds = torch.cat(cur_new_input_embeds)
+
+            cur_new_labels = torch.cat(cur_new_labels)
+
+            new_input_embeds.append(cur_new_input_embeds)
+            new_labels.append(cur_new_labels)
+
+        # Truncate sequences to max length as image embeddings can make the sequence longer
+        tokenizer_model_max_length = getattr(self.config, "tokenizer_model_max_length", None)
+        # NOTE: qmh 注释
+        # new_input_embeds = [x[:tokenizer_model_max_length] for x, modality in zip(new_input_embeds, modalities)]
+        # new_labels = [x[:tokenizer_model_max_length] for x, modality in zip(new_labels, modalities)]
+
+        # TODO: Hard code for control loss spike
+        # if tokenizer_model_max_length is not None:
+        #     new_input_embeds = [x[:4096] if modality != "video" else x[:tokenizer_model_max_length] for x, modality in zip(new_input_embeds, modalities)]
+        #     new_labels = [x[:4096] if modality != "video" else x[:tokenizer_model_max_length] for x, modality in zip(new_labels, modalities)]
+
+        # Combine them
+        max_len = max(x.shape[0] for x in new_input_embeds)
+        batch_size = len(new_input_embeds)
+
+        new_input_embeds_padded = []
+        new_labels_padded = torch.full((batch_size, max_len), IGNORE_INDEX, dtype=new_labels[0].dtype, device=new_labels[0].device)
+        attention_mask = torch.zeros((batch_size, max_len), dtype=attention_mask.dtype, device=attention_mask.device)
+        position_ids = torch.zeros((batch_size, max_len), dtype=position_ids.dtype, device=position_ids.device)
+
+        for i, (cur_new_embed, cur_new_labels) in enumerate(zip(new_input_embeds, new_labels)):
+            cur_len = cur_new_embed.shape[0]
+            if getattr(self.config, "tokenizer_padding_side", "right") == "left":
+                new_input_embeds_padded.append(torch.cat((torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device), cur_new_embed), dim=0))
+                if cur_len > 0:
+                    new_labels_padded[i, -cur_len:] = cur_new_labels
+                    attention_mask[i, -cur_len:] = True
+                    position_ids[i, -cur_len:] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
+            else:
+                new_input_embeds_padded.append(torch.cat((cur_new_embed, torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device)), dim=0))
+                if cur_len > 0:
+                    new_labels_padded[i, :cur_len] = cur_new_labels
+                    attention_mask[i, :cur_len] = True
+                    position_ids[i, :cur_len] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
+
+        new_input_embeds = torch.stack(new_input_embeds_padded, dim=0)
+
+        if _labels is None:
+            new_labels = None
+        else:
+            new_labels = new_labels_padded
+
+        if _attention_mask is None:
+            attention_mask = None
+        else:
+            attention_mask = attention_mask.to(dtype=_attention_mask.dtype)
+
+        if _position_ids is None:
+            position_ids = None
+        if getattr(self.config, "use_pos_skipping", False) and self.training:
+            position_ids = torch.arange(new_input_embeds.size(1), device=new_input_embeds.device).unsqueeze(0).to(new_input_embeds.device)
+            split_position = random.randint(0, new_input_embeds.size(1))
+            left_add = random.randint(0, self.config.pos_skipping_range)
+            right_add = random.randint(left_add, self.config.pos_skipping_range)
+            position_ids[:, :split_position] += left_add
+            position_ids[:, split_position:] += right_add
+        # import pdb; pdb.set_trace()
+        return None, position_ids, attention_mask, past_key_values, new_input_embeds, new_labels
+
+    def initialize_vision_tokenizer(self, model_args, tokenizer):
+        if model_args.mm_use_im_patch_token:
+            tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
+            self.resize_token_embeddings(len(tokenizer))
+
+        if model_args.mm_use_im_start_end:
+            num_new_tokens = tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True)
+            self.resize_token_embeddings(len(tokenizer))
+
+            if num_new_tokens > 0:
+                input_embeddings = self.get_input_embeddings().weight.data
+                output_embeddings = self.get_output_embeddings().weight.data
+
+                input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(dim=0, keepdim=True)
+                output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(dim=0, keepdim=True)
+
+                input_embeddings[-num_new_tokens:] = input_embeddings_avg
+                output_embeddings[-num_new_tokens:] = output_embeddings_avg
+
+            if model_args.tune_mm_mlp_adapter:
+                for p in self.get_input_embeddings().parameters():
+                    p.requires_grad = True
+                for p in self.get_output_embeddings().parameters():
+                    p.requires_grad = False
+
+            if model_args.pretrain_mm_mlp_adapter:
+                mm_projector_weights = torch.load(model_args.pretrain_mm_mlp_adapter, map_location="cpu")
+                embed_tokens_weight = mm_projector_weights["model.embed_tokens.weight"]
+                assert num_new_tokens == 2
+                if input_embeddings.shape == embed_tokens_weight.shape:
+                    input_embeddings[-num_new_tokens:] = embed_tokens_weight[-num_new_tokens:]
+                elif embed_tokens_weight.shape[0] == num_new_tokens:
+                    input_embeddings[-num_new_tokens:] = embed_tokens_weight
+                else:
+                    raise ValueError(f"Unexpected embed_tokens_weight shape. Pretrained: {embed_tokens_weight.shape}. Current: {input_embeddings.shape}. Numer of new tokens: {num_new_tokens}.")
+
+        elif model_args.mm_use_im_patch_token:
+            if model_args.tune_mm_mlp_adapter:
+                for p in self.get_input_embeddings().parameters():
+                    p.requires_grad = False
+                for p in self.get_output_embeddings().parameters():
+                    p.requires_grad = False
+    

llava_qwen.py

@@ -0,0 +1,709 @@
+#    Copyright 2024 Hao Zhang
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+
+
+from typing import List, Optional, Tuple, Union, Dict
+import torch
+import torch.nn as nn
+from torch.nn import CrossEntropyLoss
+import transformers
+from transformers import AutoConfig, AutoModelForCausalLM, LlamaConfig, LlamaModel, LlamaForCausalLM
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.generation.utils import GenerateOutput
+from longva.longva.model.llava_arch import LlavaMetaModel, LlavaMetaForCausalLM
+from .modeling_qwen2 import Qwen2Config, Qwen2Model, Qwen2ForCausalLM
+import pdb
+import time
+import random
+random.seed(42)
+import torch
+from statistics import mean
+import torch.nn.functional as F
+import PIL
+from decord import VideoReader, cpu
+from .conversation import conv_templates, SeparatorStyle
+from .constants import IGNORE_INDEX, IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_PATCH_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN, DEFAULT_IMAGE_TOKEN
+from .mm_utils import tokenizer_image_token, load_video
+
+
+class LlavaQwenConfig(Qwen2Config):
+    model_type = "llava_qwen"
+
+
+class LlavaQwenModel(LlavaMetaModel, Qwen2Model):
+    config_class = LlavaQwenConfig
+
+    def __init__(self, config: Qwen2Config):
+        super(LlavaQwenModel, self).__init__(config)
+
+
+class LlavaQwenForCausalLM(Qwen2ForCausalLM, LlavaMetaForCausalLM):
+    config_class = LlavaQwenConfig
+
+    def __init__(self, config):
+        # super(Qwen2ForCausalLM, self).__init__(config)
+        Qwen2ForCausalLM.__init__(self, config)
+        config.model_type = "llava_qwen"
+        config.rope_scaling = None
+
+        self.model = LlavaQwenModel(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_model(self):
+        return self.model
+
+    def uniform_sampling(self, embeds, start_idx, end_idx, step):
+        indices = torch.arange(start_idx, end_idx, step).to(device=embeds.device)
+        return embeds.index_select(1, indices), indices
+    def pooling_sampling(self, embeds, start_idx, end_idx, step, pool_type='avg'):
+        selected = embeds[:, start_idx:end_idx, :]
+        B, D, L = selected.shape
+        kernel_size = step
+        stride = step
+        
+        selected_transposed = selected.transpose(1, 2)  # shape: (1, 12, 4)
+
+        if pool_type == 'avg_pool':
+            pooled = F.avg_pool1d(selected_transposed, kernel_size=kernel_size, stride=stride)
+        elif pool_type == 'max_pool':
+            pooled = F.max_pool1d(selected_transposed, kernel_size=kernel_size, stride=stride)
+        else:
+            raise ValueError(f"Unsupported pooling type: {pool_type}")
+
+        pooled = pooled.transpose(1, 2)  # shape: (1, 2, 12)
+        return pooled, torch.arange(start_idx, start_idx + pooled.shape[1] * step, step).to(device=embeds.device)
+
+    def process_block(self, block_embeds, current_past_key_values=None, bsz=1, device=None, position_ids=None, key_position_ids=None):
+        if current_past_key_values is None:
+            seq_len = block_embeds.size(1)
+            position_ids = torch.arange(0, seq_len, device=device).expand(bsz, -1)
+            attention_mask = torch.ones((bsz, seq_len), device=device, dtype=torch.long)
+        else:
+            seq_len = block_embeds.size(1)
+            prefix_len = current_past_key_values[0][0].size(2)
+            attention_mask = torch.ones((bsz, prefix_len + seq_len), device=device, dtype=torch.long)
+
+
+        outputs = self.model(
+            inputs_embeds=block_embeds,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            key_position_ids=key_position_ids,
+            past_key_values=current_past_key_values,
+            use_cache=True,
+            return_dict=True,
+        )
+        return outputs.past_key_values
+
+    def pooling_kvs(self, kvs, step):
+        # kvs shape: (bsz, 4, seq_len, head_dim)
+        kernel_size = step
+        stride = step
+        # kvs = kvs.transpose(2, 3)
+        # pooled_kvs = F.avg_pool1d(kvs, kernel_size=kernel_size, stride=stride)
+        kvs_permuted = kvs.permute(0, 1, 3, 2) # (batch_size, num_heads, feature_dim, sequence_length)
+        N_flat = kvs_permuted.shape[0] * kvs_permuted.shape[1]
+        C = kvs_permuted.shape[2]
+        L = kvs_permuted.shape[3]
+        kvs_for_pool = kvs_permuted.reshape(N_flat, C, L)
+        pooled_kvs = F.avg_pool1d(kvs_for_pool, kernel_size=kernel_size, stride=stride)
+        pooled_kvs_restored = pooled_kvs.view(kvs.shape[0], kvs.shape[1], pooled_kvs.shape[1], pooled_kvs.shape[2]).permute(0, 1, 3, 2)
+        return pooled_kvs_restored
+
+
+    def get_sparse_attention_mask(self, total_len, num_blocks, block_size, time_token_start_indices, time_token_end_indices, time_token_indices, visual_token_start_pos, visual_token_end_pos, attention_mask, inputs_embeds, prev_blocks_num=None):
+
+        causal_mask = torch.tril(torch.ones((total_len, total_len), dtype=torch.bool)).unsqueeze(0).repeat(1, 1, 1)
+        mask = torch.zeros(total_len, total_len, dtype=torch.bool)
+        start = visual_token_start_pos
+
+        record_block_start = []
+        for i in range(num_blocks):
+            next_time_token_pos = (i + 1)*block_size
+            if next_time_token_pos >= len(time_token_start_indices):
+                end = visual_token_end_pos
+            else:
+                end =  time_token_start_indices[ next_time_token_pos ]
+
+            mask[start:end, start:end] = True  
+
+            if len(record_block_start) >= prev_blocks_num:
+                prev_start = record_block_start[-prev_blocks_num]
+            else:
+                prev_start = visual_token_start_pos
+
+            mask[start:end, prev_start:start] = True   
+            record_block_start.append(start)
+            start = end
+
+            
+        mask[:, :visual_token_start_pos] = True
+        mask[visual_token_end_pos:, :] = True        
+     
+        for idx in time_token_indices:
+            mask[idx, :] = True   
+            mask[:, idx] = True   
+
+        causal_mask = torch.tril(torch.ones(total_len, total_len, dtype=torch.bool))
+        final_mask = (mask & causal_mask).unsqueeze(0).unsqueeze(0).to(dtype=attention_mask.dtype, device=attention_mask.device)
+
+        num_allowed = final_mask.sum().item()
+        upper_triangle_num = total_len * (total_len + 1) // 2
+        ratio = num_allowed / upper_triangle_num
+        
+        invert_mask = 1.0 - final_mask
+        final_mask = ((1.0 - final_mask) * -1e9).to(dtype=inputs_embeds.dtype)
+        return final_mask, ratio
+
+
+    def cat_history_kvs(self, prefix_kvs, kvs_part2, kvs_part3):
+        prefix_kvs = [[kvs] for kvs in prefix_kvs]
+        cat_kvs = []
+        for prefix_kvs_this_layer, kvs_part2_this_layer, kvs_part3_this_layer in zip(prefix_kvs, kvs_part2, kvs_part3):
+            prefix_key_this_layer = [tmp[0] for tmp in prefix_kvs_this_layer]
+            prefix_val_this_layer = [tmp[1] for tmp in prefix_kvs_this_layer]
+
+            key_part2_this_layer = [tmp[0] for tmp in kvs_part2_this_layer]
+            val_part2_this_layer = [tmp[1] for tmp in kvs_part2_this_layer]
+
+            key_part3_this_layer = [tmp[0] for tmp in kvs_part3_this_layer]
+            val_part3_this_layer = [tmp[1] for tmp in kvs_part3_this_layer]
+
+            key_this_layer = torch.cat(prefix_key_this_layer + key_part2_this_layer + key_part3_this_layer, dim=-2)
+            val_this_layer = torch.cat(prefix_val_this_layer + val_part2_this_layer + val_part3_this_layer, dim=-2)
+
+            cat_kvs.append((key_this_layer, val_this_layer))
+        return cat_kvs
+
+    def forward_streaming(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        key_position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        dpo_forward: Optional[bool] = False,
+        cache_position=None,
+        visual_token_start_pos=None,
+        visual_token_end_pos=None,
+        time_token_start_indices=None,
+        frames_num=None,
+        time_token_indices=None,
+        time_token_end_indices=None,
+        block_size_chosed=None,
+        prev_blocks_num=None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        
+        block_size = block_size_chosed 
+        visual_token_start_pos = visual_token_start_pos
+        visual_token_end_pos = visual_token_end_pos
+        visual_len = visual_token_end_pos - visual_token_start_pos
+        num_blocks = (frames_num  + block_size * 4 - 1) // (block_size * 4)
+        # print(f'block_size: {block_size}, num_blocks: {num_blocks}')
+        
+        # streaming inps
+        blocks_positions = [[(0, 0, visual_token_start_pos)]]
+        frames_groups = [(0, visual_token_start_pos)]
+        for idx, (time_start, time_end) in enumerate(zip(time_token_start_indices, time_token_end_indices)):
+            if idx + 1 < len(time_token_start_indices):
+                frames_group_end = time_token_start_indices[idx + 1]
+            else:
+                frames_group_end = visual_token_end_pos
+            frames_groups.append(
+                (time_start, time_end, frames_group_end)
+            )
+
+        single_block = []
+        for group in frames_groups[1:]: 
+            single_block.append(group)
+            if len(single_block) == block_size:
+                blocks_positions.append(single_block)
+                single_block = []
+        if len(single_block) != 0:
+           blocks_positions.append(single_block)
+        num_blocks = len(blocks_positions)
+
+        start = time.time()
+        record_prefill_time = 0
+
+        full_inputs_embeds = inputs_embeds 
+        bsz, total_len, embed_dim = full_inputs_embeds.size()
+        device = full_inputs_embeds.device
+
+        prefix_embeds = full_inputs_embeds[:, :visual_token_start_pos, :]
+        visual_embeds = full_inputs_embeds[:, visual_token_start_pos:visual_token_end_pos, :]
+        suffix_embeds = full_inputs_embeds[:, visual_token_end_pos:, :]
+        num_visual_tokens = visual_embeds.size(1)
+
+        all_past_key_values = [[] for _ in range(len(self.model.layers))]  # 假设 model 有 layers 属性
+        prefix_past_key_values = []  
+
+        torch.cuda.reset_peak_memory_stats()
+        
+        if prefix_embeds.size(1) > 0:
+            pkv = self.process_block(prefix_embeds, bsz=bsz, device=device)
+            for i in range(len(pkv)):
+                all_past_key_values[i].append(pkv[i])
+                prefix_past_key_values.append(pkv[i])
+
+        prev_blocks = blocks_positions[1:1+prev_blocks_num]
+        prev_the_first_block = prev_blocks[0]
+        prev_b_start = prev_the_first_block[0][0]
+        prev_the_last_block = prev_blocks[-1]
+        prev_b_end = prev_the_last_block[-1][-1]
+
+        block_streaming_past_key_values = prefix_past_key_values
+
+        query_position_ids = torch.arange(prev_b_start, prev_b_end, dtype=torch.long, device=device)
+        past_key_position_ids = torch.arange(0, block_streaming_past_key_values[0][0].size(2), dtype=torch.long, device=device)
+        key_position_ids = torch.cat([past_key_position_ids, query_position_ids], dim=0)
+
+        visual_embeds_this_block = full_inputs_embeds[:,prev_b_start:prev_b_end,:]
+        pkv = self.process_block(visual_embeds_this_block, current_past_key_values=block_streaming_past_key_values, bsz=bsz, device=device, position_ids=query_position_ids.unsqueeze(0), key_position_ids=key_position_ids.unsqueeze(0))
+
+        for i in range(len(pkv)):
+            for block in prev_blocks:
+                block_start, _, _ = block[0]
+                _, _, block_end = block[-1]
+                all_past_key_values[i].append( (pkv[i][0][:,:,block_start:block_end], pkv[i][1][:,:,block_start:block_end]) )
+
+        block_streaming_past_key_values_part1 = prefix_past_key_values
+        position_ids_part1 = torch.arange(0, prefix_past_key_values[0][0].size(2), dtype=torch.long, device=device)
+        block_streaming_past_key_values_part2 = [[] for _ in range(len(self.model.layers))]   # 存
+        position_ids_part2 = torch.tensor([], dtype=torch.long, device=device)
+        block_streaming_past_key_values_part3=None
+        position_ids_part3 = None
+        
+        query_position_ids = None
+        for idx, single_block in enumerate(blocks_positions[:]):
+            if idx == 0:
+                continue
+            if idx <= prev_blocks_num:
+                continue
+            
+            b_start, _, _ = single_block[0]
+            _, _, b_end = single_block[-1]
+            visual_embeds_this_block = full_inputs_embeds[:,b_start:b_end,:]
+            prev_blocks = blocks_positions[max(idx - prev_blocks_num, 1):idx]
+            prev_the_first_block = prev_blocks[0]
+            prev_b_start = prev_the_first_block[0][0]
+            
+            this_block_length = b_end - prev_b_start
+            prev_block_length = b_start - prev_b_start
+            true_block_length = b_end - b_start
+
+            block_streaming_past_key_values_part3 = [tmp[-prev_blocks_num:] for tmp in all_past_key_values]
+            # block_streaming_past_key_values_part3 = [
+            #     [
+            #         (t[0].to(device=device), t[1].to(device=device))
+            #         for t in sublist
+            #     ]
+            #     for sublist in block_streaming_past_key_values_part3
+            # ]
+
+            block_streaming_past_key_values = self.cat_history_kvs(block_streaming_past_key_values_part1, block_streaming_past_key_values_part2, block_streaming_past_key_values_part3)
+
+            query_position_ids = torch.arange(b_start, b_end, dtype=torch.long, device=device)
+            position_ids_part3 = torch.arange(prev_b_start, b_start, dtype=torch.long, device=device)
+            key_position_ids = torch.cat([position_ids_part1, position_ids_part2, position_ids_part3, query_position_ids], dim=0)
+
+            start_1 = time.time() 
+            pkv = self.process_block(visual_embeds_this_block, current_past_key_values=block_streaming_past_key_values, bsz=bsz, device=device, position_ids=query_position_ids.unsqueeze(0), key_position_ids=key_position_ids.unsqueeze(0))
+            end_1 = time.time()
+
+            record_prefill_time += end_1-start_1
+
+            for i in range(len(pkv)):
+                length_before_chunk = block_streaming_past_key_values[i][0].size(2)
+                key_this_block, val_this_block = pkv[i]
+                key_this_block = key_this_block[:,:,length_before_chunk:,:]
+                val_this_block = val_this_block[:,:,length_before_chunk:,:]
+                all_past_key_values[i].append( (key_this_block, val_this_block) )
+                # all_past_key_values[i].append( (key_this_block.to('cpu'), val_this_block.to('cpu')) )
+
+                time_keys_list = []
+                time_vals_list = []
+
+                extract_timestamps_position_ids_list = []
+                for group in prev_the_first_block:
+                    time_start, time_end, _ = group
+                    extract_timestamps_position_ids_list.append(torch.arange(time_start, time_end, dtype=torch.long, device=device))
+
+                    time_start = time_start - prev_b_start
+                    time_end = time_end - prev_b_start
+
+                    time_keys_list.append(block_streaming_past_key_values_part3[i][0][0][:,:,time_start:time_end,:])
+                    time_vals_list.append(block_streaming_past_key_values_part3[i][0][1][:,:,time_start:time_end,:])
+                    
+                time_keys = torch.cat(time_keys_list, dim=2)
+                time_vals = torch.cat(time_vals_list, dim=2)
+
+                block_streaming_past_key_values_part2[i].append( (time_keys, time_vals) )
+
+                if i == 0: 
+                    position_ids_part2 = torch.cat([position_ids_part2] + extract_timestamps_position_ids_list, dim=0)
+
+
+        merged_pkv = []
+        for layer_pkvs in all_past_key_values:
+            if not layer_pkvs:
+                continue
+            keys = torch.cat([pkv[0].to(device=device) for pkv in layer_pkvs], dim=2)  # dim=2 是 sequence 维度
+            values = torch.cat([pkv[1].to(device=device) for pkv in layer_pkvs], dim=2)
+            merged_pkv.append((keys, values))
+
+    
+        pkv = merged_pkv
+        del block_streaming_past_key_values
+        del all_past_key_values
+        del block_streaming_past_key_values_part1
+        del block_streaming_past_key_values_part2
+        del block_streaming_past_key_values_part3
+        torch.cuda.empty_cache()
+
+        # TODO: bi-decoding acceleration
+        mixed_prefill_past_key_values = pkv
+        prefill_len = visual_token_end_pos
+        
+        # Process suffix
+        if suffix_embeds.size(1) > 0:
+            seq_len = suffix_embeds.size(1)
+            total_len = prefill_len + seq_len
+            position_ids = torch.arange(prefill_len, total_len, device=device, dtype=torch.long).expand(bsz, -1)
+            key_position_ids = torch.arange(0, total_len, device=device, dtype=torch.long).expand(bsz, -1)
+            attention_mask = torch.ones((bsz, total_len), device=device, dtype=torch.long)
+
+            outputs = super().forward(
+                inputs_embeds=suffix_embeds,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                key_position_ids=key_position_ids,
+                past_key_values=mixed_prefill_past_key_values,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                use_cache=True,
+                return_dict=return_dict,
+                # blocks_positions=None,
+            )
+            del mixed_prefill_past_key_values
+            torch.cuda.empty_cache()
+
+        return outputs
+    def forward_mask(
+        self,
+        input_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,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        dpo_forward: Optional[bool] = False,
+        cache_position=None,
+        visual_token_start_pos=None,
+        visual_token_end_pos=None,
+        time_token_start_indices=None,
+        time_token_end_indices=None,
+        frames_num=None,
+        time_token_indices=None,
+        prev_blocks_num=None,
+        block_size_chosed=None
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        bsz, total_len, embed_dim = inputs_embeds.size()
+        visual_token_start_pos = visual_token_start_pos
+        visual_token_end_pos = visual_token_end_pos
+        visual_len = visual_token_end_pos - visual_token_start_pos
+        
+        block_size_list = [2,4,8,16,32]
+        best_block_size = None
+        min_diff = float('inf')
+       
+        block_size = block_size_chosed
+        num_blocks = (frames_num  + block_size * 4 - 1) // (block_size * 4)
+        final_mask, ratio = self.get_sparse_attention_mask(total_len, num_blocks, block_size, time_token_start_indices, time_token_end_indices, time_token_indices, visual_token_start_pos, visual_token_end_pos, attention_mask, inputs_embeds, prev_blocks_num)
+
+        # print(f'frames:{frames_num}, block_num:{num_blocks}, bsz:{block_size}, prev_blocks_num:{prev_blocks_num}, ratio:{ratio}') 
+
+        return super().forward(
+            input_ids=input_ids,
+            attention_mask=final_mask, # final_mask
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            labels=labels,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        key_position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        images: Optional[torch.FloatTensor] = None,
+        image_sizes: Optional[List[List[int]]] = None,
+        return_dict: Optional[bool] = None,
+        modalities: Optional[List[str]] = ["image"],
+        dpo_forward: Optional[bool] = False,
+        cache_position=None,
+        time_embedding=None,
+        visual_token_start_pos=None,
+        visual_token_end_pos=None,
+        time_token_start_indices=None,
+        frames_num=None,
+        time_token_indices=None,
+        time_token_end_indices=None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        
+        if input_ids is not None and input_ids.size(1) == 1:
+            past_key_len = past_key_values[0][0].size(-2)
+            key_position_ids = torch.arange(0, past_key_len+1, device=position_ids.device,dtype=torch.long).expand(1, -1)
+            if position_ids[0][0] != past_key_len:
+                position_ids = torch.tensor([[past_key_len]]).to(device=position_ids.device, dtype=position_ids.dtype)
+                key_position_ids = torch.arange(0, past_key_len+1, device=position_ids.device,dtype=torch.long).expand(1, -1)
+
+            return super().forward(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                key_position_ids=key_position_ids,
+                past_key_values=past_key_values,
+                inputs_embeds=inputs_embeds,
+                labels=labels,
+                use_cache=use_cache,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+
+        if inputs_embeds is None:
+            (input_ids, position_ids, attention_mask, past_key_values, inputs_embeds, labels) = self.prepare_inputs_labels_for_multimodal(input_ids, position_ids, attention_mask, past_key_values, labels, images, modalities, image_sizes, time_embedding)
+
+        if self.config.enable_sparse:
+            block_size_chosed = self.config.sparse_config['block_size_chosed']
+            prev_blocks_num = self.config.sparse_config['prev_blocks_num']
+            if self.config.sparse_mode=='streaming':
+                return self.forward_streaming(
+                    input_ids=input_ids,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    key_position_ids=key_position_ids,
+                    past_key_values=past_key_values,
+                    inputs_embeds=inputs_embeds,
+                    labels=labels,
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                    output_hidden_states=output_hidden_states,
+                    return_dict=return_dict,
+                    cache_position=cache_position,
+                    visual_token_start_pos=visual_token_start_pos,
+                    visual_token_end_pos=visual_token_end_pos,
+                    time_token_start_indices=time_token_start_indices,
+                    frames_num=frames_num,
+                    time_token_indices=time_token_indices,
+                    time_token_end_indices=time_token_end_indices,
+                    block_size_chosed=block_size_chosed,
+                    prev_blocks_num=prev_blocks_num,
+                )
+            elif self.config.sparse_mode=='mask':
+                return self.forward_mask(
+                    input_ids=input_ids,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    past_key_values=past_key_values,
+                    inputs_embeds=inputs_embeds,
+                    labels=labels,
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                    output_hidden_states=output_hidden_states,
+                    return_dict=return_dict,
+                    cache_position=cache_position,
+                    visual_token_start_pos=visual_token_start_pos,
+                    visual_token_end_pos=visual_token_end_pos,
+                    time_token_start_indices=time_token_start_indices,
+                    frames_num=frames_num,
+                    time_token_indices=time_token_indices,
+                    time_token_end_indices=time_token_end_indices,
+                    block_size_chosed=block_size_chosed,
+                    prev_blocks_num=prev_blocks_num,
+                )
+        else:
+            return super().forward(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_values=past_key_values,
+                inputs_embeds=inputs_embeds,
+                labels=labels,
+                use_cache=use_cache,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+
+        
+    @torch.no_grad()
+    def generate(
+        self,
+        inputs: Optional[torch.Tensor] = None,
+        images: Optional[torch.Tensor] = None,
+        image_sizes: Optional[torch.Tensor] = None,
+        modalities: Optional[List[str]] = ["image"],
+        time_embedding=None,
+        **kwargs,
+    ) -> Union[GenerateOutput, torch.LongTensor]:
+
+        position_ids = kwargs.pop("position_ids", None)
+        attention_mask = kwargs.pop("attention_mask", None)
+
+        if "inputs_embeds" in kwargs:
+            raise NotImplementedError("`inputs_embeds` is not supported")
+
+        if images is not None and images[0].size(0) > 0:
+            IMAGE_TOKEN_INDEX = -200
+            TOKEN_PERFRAME = 36
+            frames_num = images[0].size(0)
+            visual_token_start_pos = (inputs == IMAGE_TOKEN_INDEX).nonzero(as_tuple=True)[1].item()
+            num_tokens = time_embedding[0].size(0)
+            visual_token_end_pos = visual_token_start_pos + num_tokens
+            kwargs['visual_token_start_pos'] = visual_token_start_pos
+            kwargs['visual_token_end_pos'] = visual_token_end_pos
+            # time_token_start_indices = (time_embedding[0] == 1462).nonzero(as_tuple=True)
+            time_token_start_indices = (time_embedding[0] == 1462).nonzero(as_tuple=True)[0].cpu().tolist()
+            kwargs['time_token_start_indices'] = [idx + visual_token_start_pos for idx in time_token_start_indices]
+            # kwargs['time_token_start_indices'] = time_token_start_indices + visual_token_start_pos
+            kwargs['frames_num'] = frames_num
+            time_token_indices = (time_embedding[0] != 151654).nonzero(as_tuple=True)[0].cpu().tolist()
+            kwargs['time_token_indices'] = [idx + visual_token_start_pos for idx in time_token_indices]
+            time_token_end_indices = (time_embedding[0] == 25).nonzero(as_tuple=True)[0].cpu().tolist()
+            kwargs['time_token_end_indices'] = [idx + visual_token_start_pos + 1 for idx in time_token_end_indices]
+            # kwargs['time_token_end_indices'] = time_token_end_indices + visual_token_start_pos
+
+        #print(images[0].shape)
+        if images is not None:
+            (inputs, position_ids, attention_mask, _, inputs_embeds, _) = self.prepare_inputs_labels_for_multimodal(inputs, position_ids, attention_mask, None, None, images, modalities, image_sizes=image_sizes,time_embedding=time_embedding)
+        
+        else:
+            inputs_embeds = self.get_model().embed_tokens(inputs)
+        
+        #print(inputs_embeds.shape)
+        return super().generate(position_ids=position_ids, attention_mask=attention_mask, inputs_embeds=inputs_embeds, **kwargs)
+
+    @torch.no_grad()
+    def chat(self,
+        video_path,
+        tokenizer,
+        user_prompt,
+        chat_history=None,
+        return_history=True,
+        max_num_frames=512,
+        sample_fps=1,
+        max_sample_fps=4,
+        generation_config={}):
+
+        # prepare text input
+        conv = conv_templates["qwen_1_5"].copy()
+        if chat_history is None or len(chat_history) == 0:
+            user_prompt = f'{DEFAULT_IMAGE_TOKEN}\n{user_prompt}'
+        else:
+            assert DEFAULT_IMAGE_TOKEN in chat_history[0]['content'], chat_history
+            for msg in chat_history:
+                conv.append_message(msg['role'], msg['content'])
+        
+        conv.append_message(conv.roles[0], user_prompt)
+        conv.append_message(conv.roles[1], None)
+        prompt = conv.get_prompt()
+        input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt").unsqueeze(0).to(self.model.device)
+
+        # prepare video input
+        frames, timestamps = load_video(video_path, max_num_frames, fps=sample_fps, max_fps=max_sample_fps)
+
+        time_stamps=[]
+        token_frames_sum=(len(timestamps)+3)//4
+        compress_frame = timestamps[::4]
+        time_embedding = []
+        for time in compress_frame:
+            item = f"Time {time}s:"
+            time_embedding.append(tokenizer(item).input_ids)
+            time_embedding.append([151654]*144)
+
+        time_embedding = [item for sublist in time_embedding for item in sublist]
+        time_embedding = torch.tensor(time_embedding, dtype=torch.long).to(self.model.device)
+        time_stamps.append(time_embedding)
+
+        video_tensor = self.get_vision_tower().image_processor.preprocess(frames, return_tensors="pt")["pixel_values"].to(self.model.device, dtype=torch.float16)
+        
+        with torch.inference_mode():
+            output_ids = self.generate(input_ids, images=[video_tensor],time_embedding=time_stamps, modalities=["video"], **generation_config)
+
+        outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
+        
+        if chat_history is None:
+            chat_history = []
+
+        chat_history.append({"role":conv.roles[0], "content":user_prompt})
+        chat_history.append({"role":conv.roles[1], "content":outputs})
+        if return_history:
+            return outputs, chat_history
+        else:
+            return outputs
+
+    def prepare_inputs_for_generation(self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs):
+        images = kwargs.pop("images", None)
+        image_sizes = kwargs.pop("image_sizes", None)
+        visual_token_start_pos = kwargs.get("visual_token_start_pos", None)
+        visual_token_end_pos = kwargs.get("visual_token_end_pos", None)
+        time_token_start_indices = kwargs.get("time_token_start_indices", None)
+        frames_num = kwargs.get("frames_num", None)
+        time_token_indices = kwargs.get("time_token_indices", None)
+        time_token_end_indices = kwargs.get("time_token_end_indices", None)
+
+        inputs = super().prepare_inputs_for_generation(input_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, **kwargs)
+
+        inputs["visual_token_start_pos"] = visual_token_start_pos
+        inputs["visual_token_end_pos"] = visual_token_end_pos
+        inputs["time_token_start_indices"] = time_token_start_indices
+        inputs["frames_num"] = frames_num
+        inputs["time_token_indices"] = time_token_indices
+        inputs["time_token_end_indices"] = time_token_end_indices
+
+        if images is not None:
+            inputs["images"] = images
+        if image_sizes is not None:
+            inputs["image_sizes"] = image_sizes
+        return inputs
+
+
+AutoConfig.register("llava_qwen", LlavaQwenConfig)
+AutoModelForCausalLM.register(LlavaQwenConfig, LlavaQwenConfig)

mm_utils.py

@@ -0,0 +1,454 @@
+from PIL import Image
+from io import BytesIO
+import base64
+import math
+import ast
+import re
+import torch
+from transformers import StoppingCriteria
+from .constants import IMAGE_TOKEN_INDEX
+import pdb
+import numpy as np
+
+def transform_input_id(input_ids,num_tokens,special_token):
+    start_value = -200
+    insert_index = (input_ids == start_value).nonzero(as_tuple=True)[1][0].item()
+    negative_tokens = torch.arange(start_value, start_value - num_tokens, -1, device=input_ids.device)
+    before_input_ids = input_ids[:, :insert_index]
+    after_input_ids = input_ids[:, insert_index + 1:]
+    input_ids = torch.cat((before_input_ids, negative_tokens.unsqueeze(0), after_input_ids), dim=1)
+    input_ids[input_ids < 0] = special_token
+    return input_ids
+
+def resize_and_center_crop(image, shortest_edge_length):
+    # Calculate new dimensions and resize
+    aspect_ratio = float(image.width) / float(image.height)
+    if aspect_ratio > 1:
+        new_width = int(shortest_edge_length * aspect_ratio)
+        new_height = shortest_edge_length
+    else:
+        new_width = shortest_edge_length
+        new_height = int(shortest_edge_length / aspect_ratio)
+    resized_image = image.resize((new_width, new_height), Image.ANTIALIAS)
+
+    # Calculate the position and perform the center crop
+    left = (new_width - shortest_edge_length) / 2
+    top = (new_height - shortest_edge_length) / 2
+    right = (new_width + shortest_edge_length) / 2
+    bottom = (new_height + shortest_edge_length) / 2
+    cropped_image = resized_image.crop((left, top, right, bottom))
+
+    return cropped_image
+
+
+def auto_pad_images(image, grid_params):
+    assert isinstance(image, Image.Image), "Input should be a Pillow Image"
+    assert len(grid_params) > 0, "Grid parameters should not be empty"
+
+    # Step 1: Calculate and find the closest aspect ratio
+    input_width, input_height = image.size
+    input_aspect_ratio = input_width / input_height
+    candidate_resolutions = [(w / h, w, h) for w in grid_params for h in grid_params]
+    closest_aspect_ratio = min(candidate_resolutions, key=lambda x: abs(input_aspect_ratio - x[0]))
+
+    candidate_resolutions = [(x[1], x[2]) for x in candidate_resolutions if abs(x[0] - closest_aspect_ratio[0]) < 1e-3]
+
+    target_resolution = min(candidate_resolutions, key=lambda res: abs(max(input_width, input_height) / max(res) - 1))
+
+    resize_width, resize_height = target_resolution
+    if input_width > input_height:
+        resize_height = int(resize_width / input_aspect_ratio)
+    else:
+        resize_width = int(resize_height * input_aspect_ratio)
+    resized_image = image.resize((resize_width, resize_height), Image.ANTIALIAS)
+
+    # Step 5: Pad the resized image if necessary to match the target resolution
+    pad_width = target_resolution[0] - resize_width
+    pad_height = target_resolution[1] - resize_height
+    padded_image = Image.new("RGB", target_resolution, color=(0, 0, 0))
+    padded_image.paste(resized_image, (pad_width // 2, pad_height // 2))
+
+    return padded_image
+
+
+def extract_patches(image, patch_size, overlap_ratio):
+    assert isinstance(image, Image.Image), "Input should be a Pillow Image"
+    assert patch_size > 0, "Patch size should be greater than 0"
+    assert 0 <= overlap_ratio < 1, "Overlap ratio should be between 0 and 1"
+
+    W, H = image.size
+    patches = []
+
+    stride = int(patch_size * (1 - overlap_ratio))
+
+    num_patches_y = (H - patch_size) // stride + 1
+    num_patches_x = (W - patch_size) // stride + 1
+
+    y_start = (H - (num_patches_y - 1) * stride - patch_size) // 2
+    x_start = (W - (num_patches_x - 1) * stride - patch_size) // 2
+
+    for y in range(y_start, y_start + num_patches_y * stride, stride):
+        for x in range(x_start, x_start + num_patches_x * stride, stride):
+            patch = image.crop((x, y, x + patch_size, y + patch_size))
+            patches.append(patch)
+
+    return patches
+
+
+def process_highres_image_crop_split(image, data_args, processor=None):
+    crop_resolution = data_args.image_crop_resolution
+    split_resolution = data_args.image_split_resolution
+    if processor is None:
+        processor = data_args.image_processor
+    image_crop = resize_and_center_crop(image, crop_resolution)
+    image_patches = extract_patches(image_crop, patch_size=split_resolution, overlap_ratio=0)
+    image_patches = [processor.preprocess(image_patch, return_tensors="pt")["pixel_values"][0] for image_patch in image_patches]
+    return torch.stack(image_patches, dim=0)
+
+
+def process_highres_image(image, processor, grid_pinpoints):
+    grid_params = [int(x) for x in grid_pinpoints.split(",")]
+    width_height = max(image.size)
+    fit_grid_params = [x for x in grid_params if x >= width_height]
+    if len(fit_grid_params) == 0:
+        select_size = max(grid_params)
+    else:
+        select_size = min(fit_grid_params)
+    # FIXME: always select the 448
+    select_size = max(grid_params)
+    image_padded = expand2square(image, tuple(int(x * 255) for x in processor.image_mean))
+
+    # FIXME: this seems to be a bug that it always resizes instead of padding
+    image_original_resize = image.resize((processor.size["shortest_edge"], processor.size["shortest_edge"]))
+    image_padded = image_padded.resize((select_size, select_size))
+    image_patches = extract_patches(image_padded, patch_size=processor.size["shortest_edge"], overlap_ratio=0)
+    image_patches = [image_original_resize] + image_patches
+    image_patches = [processor.preprocess(image_patch, return_tensors="pt")["pixel_values"][0] for image_patch in image_patches]
+    return torch.stack(image_patches, dim=0)
+
+
+def select_best_resolution(original_size, possible_resolutions):
+    """
+    Selects the best resolution from a list of possible resolutions based on the original size.
+
+    Args:
+        original_size (tuple): The original size of the image in the format (width, height).
+        possible_resolutions (list): A list of possible resolutions in the format [(width1, height1), (width2, height2), ...].
+
+    Returns:
+        tuple: The best fit resolution in the format (width, height).
+    """
+    original_width, original_height = original_size
+    best_fit = None
+    max_effective_resolution = 0
+    min_wasted_resolution = float("inf")
+
+    for width, height in possible_resolutions:
+        # Calculate the downscaled size to keep the aspect ratio
+        scale = min(width / original_width, height / original_height)
+        downscaled_width, downscaled_height = int(original_width * scale), int(original_height * scale)
+
+        # Calculate effective and wasted resolutions
+        effective_resolution = min(downscaled_width * downscaled_height, original_width * original_height)
+        wasted_resolution = (width * height) - effective_resolution
+
+        if effective_resolution > max_effective_resolution or (effective_resolution == max_effective_resolution and wasted_resolution < min_wasted_resolution):
+            max_effective_resolution = effective_resolution
+            min_wasted_resolution = wasted_resolution
+            best_fit = (width, height)
+
+    return best_fit
+
+
+def resize_and_pad_image(image, target_resolution):
+    """
+    Resize and pad an image to a target resolution while maintaining aspect ratio.
+
+    Args:
+        image (PIL.Image.Image): The input image.
+        target_resolution (tuple): The target resolution (width, height) of the image.
+
+    Returns:
+        PIL.Image.Image: The resized and padded image.
+    """
+    original_width, original_height = image.size
+    target_width, target_height = target_resolution
+
+    # Determine which dimension (width or height) to fill
+    scale_w = target_width / original_width
+    scale_h = target_height / original_height
+
+    if scale_w < scale_h:
+        # Width will be filled completely
+        new_width = target_width
+        new_height = min(math.ceil(original_height * scale_w), target_height)
+    else:
+        # Height will be filled completely
+        new_height = target_height
+        new_width = min(math.ceil(original_width * scale_h), target_width)
+
+    # Resize the image
+    resized_image = image.resize((new_width, new_height))
+
+    # Create a new image with the target size and paste the resized image onto it
+    new_image = Image.new("RGB", (target_width, target_height), (0, 0, 0))
+    paste_x = (target_width - new_width) // 2
+    paste_y = (target_height - new_height) // 2
+    new_image.paste(resized_image, (paste_x, paste_y))
+
+    return new_image
+
+
+def divide_to_patches(image, patch_size):
+    """
+    Divides an image into patches of a specified size.
+
+    Args:
+        image (PIL.Image.Image): The input image.
+        patch_size (int): The size of each patch.
+
+    Returns:
+        list: A list of PIL.Image.Image objects representing the patches.
+    """
+    patches = []
+    width, height = image.size
+    for i in range(0, height, patch_size):
+        for j in range(0, width, patch_size):
+            box = (j, i, j + patch_size, i + patch_size)
+            patch = image.crop(box)
+            patches.append(patch)
+
+    return patches
+
+
+def get_anyres_image_grid_shape(image_size, grid_pinpoints, patch_size):
+    """
+    Calculate the shape of the image patch grid after the preprocessing for images of any resolution.
+
+    Args:
+        image_size (tuple): The size of the input image in the format (width, height).
+        grid_pinpoints (str): A string representation of a list of possible resolutions.
+        patch_size (int): The size of each image patch.
+
+    Returns:
+        tuple: The shape of the image patch grid in the format (width, height).
+    """
+    if isinstance(grid_pinpoints, str) and "x" in grid_pinpoints:
+        assert patch_size in [224, 336, 384, 448, 512], "patch_size should be in [224, 336, 384, 448, 512]"
+        # Use regex to extract the range from the input string
+        matches = re.findall(r"\((\d+)x(\d+)\)", grid_pinpoints)
+        range_start = tuple(map(int, matches[0]))
+        range_end = tuple(map(int, matches[-1]))
+        # Generate a matrix of tuples from (range_start[0], range_start[1]) to (range_end[0], range_end[1])
+        grid_pinpoints = [(i, j) for i in range(range_start[0], range_end[0] + 1) for j in range(range_start[1], range_end[1] + 1)]
+        # Multiply all elements by patch_size
+        grid_pinpoints = [[dim * patch_size for dim in pair] for pair in grid_pinpoints]
+    if type(grid_pinpoints) is list:
+        possible_resolutions = grid_pinpoints
+    else:
+        possible_resolutions = ast.literal_eval(grid_pinpoints)
+    width, height = select_best_resolution(image_size, possible_resolutions)
+    return width // patch_size, height // patch_size
+
+
+def process_anyres_image(image, processor, grid_pinpoints):
+    """
+    Process an image with variable resolutions.
+
+    Args:
+        image (PIL.Image.Image): The input image to be processed.
+        processor: The image processor object.
+        grid_pinpoints (str): A string representation of a list of possible resolutions.
+
+    Returns:
+        torch.Tensor: A tensor containing the processed image patches.
+    """
+    # Convert grid_pinpoints from string to list
+    if isinstance(grid_pinpoints, str) and "x" in grid_pinpoints:
+        try:
+            patch_size = processor.size[0]
+        except Exception as e:
+            patch_size = processor.size["shortest_edge"]
+        assert patch_size in [224, 336, 384, 448, 512], "patch_size should be in [224, 336, 384, 448, 512]"
+        #print(patch_size)
+        # Use regex to extract the range from the input string
+        matches = re.findall(r"\((\d+)x(\d+)\)", grid_pinpoints)
+        range_start = tuple(map(int, matches[0]))
+        range_end = tuple(map(int, matches[-1]))
+        # Generate a matrix of tuples from (range_start[0], range_start[1]) to (range_end[0], range_end[1])
+        grid_pinpoints = [(i, j) for i in range(range_start[0], range_end[0] + 1) for j in range(range_start[1], range_end[1] + 1)]
+        # Multiply all elements by patch_size
+        grid_pinpoints = [[dim * patch_size for dim in pair] for pair in grid_pinpoints]
+
+    if type(grid_pinpoints) is list:
+        possible_resolutions = grid_pinpoints
+    else:
+        possible_resolutions = ast.literal_eval(grid_pinpoints)
+    best_resolution = select_best_resolution(image.size, possible_resolutions)
+    image_padded = resize_and_pad_image(image, best_resolution)
+
+    patches = divide_to_patches(image_padded, processor.crop_size["height"])
+    
+    # FIXME: this seems to be a bug that it resizes instead of pad.
+    # but to keep it consistent with previous, i will keep it as it is
+    # TODO: uncomment below to ablate with the padding
+    if isinstance(processor.size, dict):
+        shortest_edge = processor.size["shortest_edge"]
+    else:
+        shortest_edge = min(processor.size)
+    image_original_resize = image.resize((shortest_edge, shortest_edge))
+    # image_padded_square = expand2square(image, tuple(int(x*255) for x in processor.image_mean))
+    # image_original_resize = image_padded_square.resize((processor.size['shortest_edge'], processor.size['shortest_edge']))
+
+    image_patches = [image_original_resize] + patches
+    image_patches = [processor.preprocess(image_patch, return_tensors="pt")["pixel_values"][0] for image_patch in image_patches]
+    return torch.stack(image_patches, dim=0)
+
+
+def load_image_from_base64(image):
+    return Image.open(BytesIO(base64.b64decode(image)))
+
+
+def expand2square(pil_img, background_color):
+    width, height = pil_img.size
+    if width == height:
+        return pil_img
+    elif width > height:
+        result = Image.new(pil_img.mode, (width, width), background_color)
+        result.paste(pil_img, (0, (width - height) // 2))
+        return result
+    else:
+        result = Image.new(pil_img.mode, (height, height), background_color)
+        result.paste(pil_img, ((height - width) // 2, 0))
+        return result
+
+def process_images_mvbench(images, image_processor, model_cfg):
+    image_aspect_ratio = getattr(model_cfg, "image_aspect_ratio", None)
+    new_images = []
+
+    for image in images:
+            image = expand2square(image, tuple(int(x * 255) for x in image_processor.image_mean))
+            image = image_processor.preprocess(image, return_tensors="pt")["pixel_values"][0]
+            new_images.append(image)
+
+    if all(x.shape == new_images[0].shape for x in new_images):
+        new_images = torch.stack(new_images, dim=0)
+    return new_images
+def process_images(images, image_processor, model_cfg):
+    image_aspect_ratio = getattr(model_cfg, "image_aspect_ratio", None)
+    new_images = []
+    if image_aspect_ratio == "highres":
+        for image in images:
+            image = process_highres_image(image, image_processor, model_cfg.image_grid_pinpoints)
+            new_images.append(image)
+    elif image_aspect_ratio == "anyres" or "anyres_max" in image_aspect_ratio:
+        for image in images:
+            image = process_anyres_image(image, image_processor, model_cfg.image_grid_pinpoints)
+            new_images.append(image)
+    elif image_aspect_ratio == "crop_split":
+        for image in images:
+            image = process_highres_image_crop_split(image, model_cfg, image_processor)
+            new_images.append(image)
+    elif image_aspect_ratio == "pad":
+        for image in images:
+            image = expand2square(image, tuple(int(x * 255) for x in image_processor.image_mean))
+            image = image_processor.preprocess(image, return_tensors="pt")["pixel_values"][0]
+            new_images.append(image)
+    else:
+        return image_processor(images, return_tensors="pt")["pixel_values"]
+    if all(x.shape == new_images[0].shape for x in new_images):
+        new_images = torch.stack(new_images, dim=0)
+    return new_images
+
+
+def tokenizer_image_token(prompt, tokenizer, image_token_index=IMAGE_TOKEN_INDEX, return_tensors=None):
+    prompt_chunks = [tokenizer(chunk).input_ids for chunk in prompt.split("<image>")]
+
+    def insert_separator(X, sep):
+        return [ele for sublist in zip(X, [sep] * len(X)) for ele in sublist][:-1]
+
+    input_ids = []
+    offset = 0
+    if len(prompt_chunks) > 0 and len(prompt_chunks[0]) > 0 and prompt_chunks[0][0] == tokenizer.bos_token_id:
+        offset = 1
+        input_ids.append(prompt_chunks[0][0])
+
+    for x in insert_separator(prompt_chunks, [image_token_index] * (offset + 1)):
+        input_ids.extend(x[offset:])
+
+    if return_tensors is not None:
+        if return_tensors == "pt":
+            return torch.tensor(input_ids, dtype=torch.long)
+        raise ValueError(f"Unsupported tensor type: {return_tensors}")
+    return input_ids
+
+
+def get_model_name_from_path(model_path):
+    model_path = model_path.strip("/")
+    model_paths = model_path.split("/")
+    if model_paths[-1].startswith("checkpoint-"):
+        return model_paths[-2] + "_" + model_paths[-1]
+    else:
+        return model_paths[-1]
+
+
+class KeywordsStoppingCriteria(StoppingCriteria):
+    def __init__(self, keywords, tokenizer, input_ids):
+        self.keywords = keywords
+        self.keyword_ids = []
+        for keyword in keywords:
+            cur_keyword_ids = tokenizer(keyword).input_ids
+            if len(cur_keyword_ids) > 1 and cur_keyword_ids[0] == tokenizer.bos_token_id:
+                cur_keyword_ids = cur_keyword_ids[1:]
+            self.keyword_ids.append(torch.tensor(cur_keyword_ids))
+        self.tokenizer = tokenizer
+        self.start_len = input_ids.shape[1]
+
+    def __call__(self, output_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
+        assert output_ids.shape[0] == 1, "Only support batch size 1 (yet)"  # TODO
+        offset = min(output_ids.shape[1] - self.start_len, 3)
+        self.keyword_ids = [keyword_id.to(output_ids.device) for keyword_id in self.keyword_ids]
+        for keyword_id in self.keyword_ids:
+            if output_ids[0, -keyword_id.shape[0] :] == keyword_id:
+                return True
+        outputs = self.tokenizer.batch_decode(output_ids[:, -offset:], skip_special_tokens=True)[0]
+        for keyword in self.keywords:
+            if keyword in outputs:
+                return True
+        return False
+        
+from decord import VideoReader, cpu
+def load_video(video_path, max_frames_num, fps=1, max_fps=4):
+    if isinstance(video_path, str):
+        vr = VideoReader(video_path, ctx=cpu(0))
+    else:
+        vr = VideoReader(video_path[0], ctx=cpu(0))
+    total_frame_num = len(vr)
+    avg_fps_from_decord = vr.get_avg_fps()
+
+    if avg_fps_from_decord <= 0:
+        print("Warning: Effective FPS is 0, cannot estimate timestamps.")
+        return None, None, []
+
+    video_fps = fps
+    step = round(avg_fps_from_decord / video_fps) if video_fps > 0 and avg_fps_from_decord > 0 else 1
+    frame_idx = [i for i in range(0, total_frame_num, step)]
+
+    fps_upbound = max_fps
+    frames_upbound = max_frames_num
+
+    if fps_upbound is not None:
+        higher_fps = min(frames_upbound//len(frame_idx), fps_upbound)
+        if higher_fps > video_fps:
+            higher_steps = round(avg_fps_from_decord / higher_fps)
+            frame_idx = [i for i in range(0, total_frame_num, higher_steps)]
+
+    if frames_upbound > 0:
+        if len(frame_idx) > frames_upbound:
+            uniform_sampled_frames = np.linspace(0, total_frame_num - 1, frames_upbound, dtype=int)
+            frame_idx = uniform_sampled_frames.tolist()
+
+    timestamps = [round(idx / avg_fps_from_decord, 1) for idx in frame_idx]
+    video = vr.get_batch(frame_idx).asnumpy()
+    vr.seek(0)
+    return video, timestamps

modeling_qwen2.py

@@ -0,0 +1,1549 @@
+# coding=utf-8
+# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch Qwen2 model."""
+
+import math
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache, StaticCache
+from transformers.modeling_attn_mask_utils import (
+    AttentionMaskConverter,
+)
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+    SequenceClassifierOutputWithPast,
+    TokenClassifierOutput,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import (
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    is_flash_attn_2_available,
+    is_flash_attn_greater_or_equal_2_10,
+    logging,
+    replace_return_docstrings,
+)
+from transformers.models.qwen2.configuration_qwen2 import Qwen2Config
+
+
+if is_flash_attn_2_available():
+    from transformers.modeling_flash_attention_utils import _flash_attention_forward
+
+import pdb
+
+logger = logging.get_logger(__name__)
+
+
+_CHECKPOINT_FOR_DOC = "Qwen/Qwen2-7B-beta"
+_CONFIG_FOR_DOC = "Qwen2Config"
+
+
+# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Qwen2
+class Qwen2RMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        Qwen2RMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+
+# Copied from transformers.models.mixtral.modeling_mixtral.MixtralRotaryEmbedding with Mixtral->Qwen2
+class Qwen2RotaryEmbedding(nn.Module):
+    def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
+        super().__init__()
+
+        self.dim = dim
+        self.max_position_embeddings = max_position_embeddings
+        self.base = base
+        inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+
+        # Build here to make `torch.jit.trace` work.
+        self._set_cos_sin_cache(
+            seq_len=max_position_embeddings, device=self.inv_freq.device, dtype=torch.get_default_dtype()
+        )
+
+    def _set_cos_sin_cache(self, seq_len, device, dtype):
+        self.max_seq_len_cached = seq_len
+        t = torch.arange(self.max_seq_len_cached, device=device, dtype=torch.int64).type_as(self.inv_freq)
+
+        freqs = torch.outer(t, self.inv_freq)
+        # Different from paper, but it uses a different permutation in order to obtain the same calculation
+        emb = torch.cat((freqs, freqs), dim=-1)
+        self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
+        self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)
+
+    def forward(self, x, seq_len=None):
+        # x: [bs, num_attention_heads, seq_len, head_size]
+        if seq_len > self.max_seq_len_cached:
+            self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)
+
+        return (
+            self.cos_cached[:seq_len].to(dtype=x.dtype),
+            self.sin_cached[:seq_len].to(dtype=x.dtype),
+        )
+
+
+# Copied from transformers.models.llama.modeling_llama.rotate_half
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+# Copied from transformers.models.mixtral.modeling_mixtral.apply_rotary_pos_emb
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids, key_position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`):
+            The position indices of the tokens corresponding to the query and key tensors. For example, this can be
+            used to pass offsetted position ids when working with a KV-cache.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    q_cos = cos[position_ids].unsqueeze(unsqueeze_dim)
+    q_sin = sin[position_ids].unsqueeze(unsqueeze_dim)
+    q_embed = (q * q_cos) + (rotate_half(q) * q_sin)
+    if key_position_ids is None:
+        k_embed = (k * q_cos) + (rotate_half(k) * q_sin)
+    else:
+        k_cos = cos[key_position_ids].unsqueeze(unsqueeze_dim)
+        k_sin = sin[key_position_ids].unsqueeze(unsqueeze_dim)
+        k_embed = (k * k_cos) + (rotate_half(k) * k_sin)
+
+    return q_embed, k_embed
+
+
+# Copied from transformers.models.mistral.modeling_mistral.MistralMLP with Mistral->Qwen2
+class Qwen2MLP(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, hidden_state):
+        return self.down_proj(self.act_fn(self.gate_proj(hidden_state)) * self.up_proj(hidden_state))
+
+
+# Copied from transformers.models.llama.modeling_llama.repeat_kv
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+class Qwen2Attention(nn.Module):
+    """
+    Multi-headed attention from 'Attention Is All You Need' paper. Modified to use sliding window attention: Longformer
+    and "Generating Long Sequences with Sparse Transformers".
+    """
+
+    def __init__(self, config: Qwen2Config, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        if layer_idx is None:
+            logger.warning_once(
+                f"Instantiating {self.__class__.__name__} without passing `layer_idx` is not recommended and will "
+                "to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
+                "when creating this class."
+            )
+
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.max_position_embeddings = config.max_position_embeddings
+        self.rope_theta = config.rope_theta
+        self.is_causal = True
+        self.attention_dropout = config.attention_dropout
+
+        if (self.head_dim * self.num_heads) != self.hidden_size:
+            raise ValueError(
+                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
+                f" and `num_heads`: {self.num_heads})."
+            )
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=True)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+
+        self.rotary_emb = Qwen2RotaryEmbedding(
+            self.head_dim,
+            max_position_embeddings=self.max_position_embeddings,
+            base=self.rope_theta,
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        blocks_positions=None,
+        key_position_ids=None,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            if self.layer_idx is None:
+                raise ValueError(
+                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
+                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
+                    "with a layer index."
+                )
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+        
+        # cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+        # query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+
+        if past_key_value is not None:
+            # cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
+            cache_kwargs = None
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+       # kv_seq_len = key_states.shape[-2]
+        max_seq_len = position_ids[0][-1] + 1
+        cos, sin = self.rotary_emb(value_states, seq_len=max_seq_len)
+
+        try:
+            query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids, key_position_ids)
+        except:
+            pdb.set_trace()
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        if attention_mask is not None:  # no matter the length, we just slice it
+            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+            attn_weights = attn_weights + causal_mask
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+# class Qwen2FlashAttention2(Qwen2Attention):
+#     """
+#     Qwen2 flash attention module, following Qwen2 attention module. This module inherits from `Qwen2Attention`
+#     as the weights of the module stays untouched. The only required change would be on the forward pass
+#     where it needs to correctly call the public API of flash attention and deal with padding tokens
+#     in case the input contains any of them. Additionally, for sliding window attention, we apply SWA only to the bottom
+#     config.max_window_layers layers.
+#     """
+
+#     # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
+#     def __init__(self, *args, **kwargs):
+#         super().__init__(*args, **kwargs)
+
+#         # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+#         # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+#         # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+#         self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+#     def forward(
+#         self,
+#         hidden_states: torch.Tensor,
+#         attention_mask: Optional[torch.Tensor] = None,
+#         position_ids: Optional[torch.LongTensor] = None,
+#         past_key_value: Optional[Cache] = None,
+#         output_attentions: bool = False,
+#         use_cache: bool = False,
+#         cache_position: Optional[torch.LongTensor] = None,
+#     ):
+#         import pdb
+
+#         bsz, q_len, _ = hidden_states.size()
+
+#         query_states = self.q_proj(hidden_states)
+#         key_states = self.k_proj(hidden_states)
+#         value_states = self.v_proj(hidden_states)
+
+#         query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+#         key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+#         value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+#         kv_seq_len = key_states.shape[-2]
+#         if past_key_value is not None:
+#             if self.layer_idx is None:
+#                 raise ValueError(
+#                     f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
+#                     "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
+#                     "with a layer index."
+#                 )
+#             kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+
+#         # Because the input can be padded, the absolute sequence length depends on the max position id.
+#         rotary_seq_len = max(kv_seq_len, position_ids[:, -1].max().item()) + 1
+#         cos, sin = self.rotary_emb(value_states, seq_len=rotary_seq_len)
+        
+#         # if self.layer_idx == 0:
+#         #     pdb.set_trace()
+
+#         if past_key_value is not None:
+#             # Activate slicing cache only if the config has a value `sliding_windows` attribute
+#             cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0
+#             if (
+#                 getattr(self.config, "sliding_window", None) is not None
+#                 and kv_seq_len > self.config.sliding_window
+#                 and cache_has_contents
+#             ):
+#                 slicing_tokens = 1 - self.config.sliding_window
+
+#                 past_key = past_key_value[self.layer_idx][0]
+#                 past_value = past_key_value[self.layer_idx][1]
+
+#                 past_key = past_key[:, :, slicing_tokens:, :].contiguous()
+#                 past_value = past_value[:, :, slicing_tokens:, :].contiguous()
+
+#                 if past_key.shape[-2] != self.config.sliding_window - 1:
+#                     raise ValueError(
+#                         f"past key must have a shape of (`batch_size, num_heads, self.config.sliding_window-1, head_dim`), got"
+#                         f" {past_key.shape}"
+#                     )
+
+#                 if attention_mask is not None:
+#                     attention_mask = attention_mask[:, slicing_tokens:]
+#                     attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1)
+
+#             cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
+#             key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+            
+#             key_position_ids = torch.arange(0, key_states.size(2), device=key_states.device).expand(1, -1)
+#         else:
+#             key_position_ids = None
+        
+#         # if self.layer_idx == 0:
+#         #     if not torch.equal(key_position_ids, position_ids):
+#         #         print(f'key_position_ids: {key_position_ids[:3]}')
+#         #         print(f'position_ids: {position_ids[:3]}')
+#         #         print(f'='*50)
+
+#         query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids, key_position_ids)
+
+#         # repeat k/v heads if n_kv_heads < n_heads
+#         key_states = repeat_kv(key_states, self.num_key_value_groups)
+#         value_states = repeat_kv(value_states, self.num_key_value_groups)
+#         dropout_rate = 0.0 if not self.training else self.attention_dropout
+
+#         # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+#         # therefore the input hidden states gets silently casted in float32. Hence, we need
+#         # cast them back in float16 just to be sure everything works as expected.
+#         input_dtype = query_states.dtype
+#         if input_dtype == torch.float32:
+#             if torch.is_autocast_enabled():
+#                 target_dtype = torch.get_autocast_gpu_dtype()
+#             # Handle the case where the model is quantized
+#             elif hasattr(self.config, "_pre_quantization_dtype"):
+#                 target_dtype = self.config._pre_quantization_dtype
+#             else:
+#                 target_dtype = self.q_proj.weight.dtype
+
+#             logger.warning_once(
+#                 f"The input hidden states seems to be silently casted in float32, this might be related to"
+#                 f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+#                 f" {target_dtype}."
+#             )
+
+#             query_states = query_states.to(target_dtype)
+#             key_states = key_states.to(target_dtype)
+#             value_states = value_states.to(target_dtype)
+
+#         # Reashape to the expected shape for Flash Attention
+#         query_states = query_states.transpose(1, 2)
+#         key_states = key_states.transpose(1, 2)
+#         value_states = value_states.transpose(1, 2)
+
+#         if (
+#             self.config.use_sliding_window
+#             and getattr(self.config, "sliding_window", None) is not None
+#             and self.layer_idx >= self.config.max_window_layers
+#         ):
+#             sliding_window = self.config.sliding_window
+#         else:
+#             sliding_window = None
+
+#         attn_output = _flash_attention_forward(
+#             query_states,
+#             key_states,
+#             value_states,
+#             attention_mask,
+#             q_len,
+#             dropout=dropout_rate,
+#             sliding_window=sliding_window,
+#             is_causal=self.is_causal,
+#             use_top_left_mask=self._flash_attn_uses_top_left_mask,
+#         )
+
+#         attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
+#         attn_output = self.o_proj(attn_output)
+
+#         if not output_attentions:
+#             attn_weights = None
+
+#         return attn_output, attn_weights, past_key_value
+
+
+
+class Qwen2FlashAttention2(Qwen2Attention):
+    """
+    Qwen2 flash attention module, following Qwen2 attention module. This module inherits from `Qwen2Attention`
+    as the weights of the module stays untouched. The only required change would be on the forward pass
+    where it needs to correctly call the public API of flash attention and deal with padding tokens
+    in case the input contains any of them. Additionally, for sliding window attention, we apply SWA only to the bottom
+    config.max_window_layers layers.
+    """
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+    ):
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            if self.layer_idx is None:
+                raise ValueError(
+                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
+                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
+                    "with a layer index."
+                )
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+
+        # Because the input can be padded, the absolute sequence length depends on the max position id.
+        rotary_seq_len = max(kv_seq_len, position_ids[:, -1].max().item()) + 1
+        cos, sin = self.rotary_emb(value_states, seq_len=rotary_seq_len)
+
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+
+        if past_key_value is not None:
+            # Activate slicing cache only if the config has a value `sliding_windows` attribute
+            cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0
+            if (
+                getattr(self.config, "sliding_window", None) is not None
+                and kv_seq_len > self.config.sliding_window
+                and cache_has_contents
+            ):
+                slicing_tokens = 1 - self.config.sliding_window
+
+                past_key = past_key_value[self.layer_idx][0]
+                past_value = past_key_value[self.layer_idx][1]
+
+                past_key = past_key[:, :, slicing_tokens:, :].contiguous()
+                past_value = past_value[:, :, slicing_tokens:, :].contiguous()
+
+                if past_key.shape[-2] != self.config.sliding_window - 1:
+                    raise ValueError(
+                        f"past key must have a shape of (`batch_size, num_heads, self.config.sliding_window-1, head_dim`), got"
+                        f" {past_key.shape}"
+                    )
+
+                if attention_mask is not None:
+                    attention_mask = attention_mask[:, slicing_tokens:]
+                    attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1)
+
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+        dropout_rate = 0.0 if not self.training else self.attention_dropout
+
+        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+        # therefore the input hidden states gets silently casted in float32. Hence, we need
+        # cast them back in float16 just to be sure everything works as expected.
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.q_proj.weight.dtype
+
+            logger.warning_once(
+                f"The input hidden states seems to be silently casted in float32, this might be related to"
+                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+                f" {target_dtype}."
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        # Reashape to the expected shape for Flash Attention
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        if (
+            self.config.use_sliding_window
+            and getattr(self.config, "sliding_window", None) is not None
+            and self.layer_idx >= self.config.max_window_layers
+        ):
+            sliding_window = self.config.sliding_window
+        else:
+            sliding_window = None
+
+        attn_output = _flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            q_len,
+            dropout=dropout_rate,
+            sliding_window=sliding_window,
+            is_causal=self.is_causal,
+            use_top_left_mask=self._flash_attn_uses_top_left_mask,
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+# Copied from transformers.models.mixtral.modeling_mixtral.MixtralSdpaAttention with Mixtral->Qwen2
+class Qwen2SdpaAttention(Qwen2Attention):
+    """
+    Qwen2 attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
+    `Qwen2Attention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
+    SDPA API.
+    """
+    
+    def compute_similarity(self, q_reps, p_reps):
+        if len(p_reps.size()) == 2:
+            return torch.matmul(q_reps, p_reps.transpose(0, 1))
+        return torch.matmul(q_reps, p_reps.transpose(-2, -1))
+
+    # Adapted from Qwen2Attention.forward
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        key_position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        blocks_positions=None,
+        layer_idx=None,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        if output_attentions:
+            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
+            logger.warning_once(
+                "Qwen2Model is using Qwen2SdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
+                'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+            )
+            return super().forward(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_value=past_key_value,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+            )
+
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        if past_key_value is not None:
+            # cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
+            cache_kwargs = None
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+        
+        max_seq_len = position_ids[0][-1] + 1
+        cos, sin = self.rotary_emb(value_states, seq_len=max_seq_len)
+
+        try:
+            query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids, key_position_ids)
+        except:
+            pdb.set_trace()
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+ 
+        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
+        # Reference: https://github.com/pytorch/pytorch/issues/112577.
+        if query_states.device.type == "cuda" and attention_mask is not None:
+            query_states = query_states.contiguous()
+            key_states = key_states.contiguous()
+            value_states = value_states.contiguous()
+ 
+        causal_mask = attention_mask
+        if attention_mask is not None:  # no matter the length, we just slice it
+            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
+        # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
+        # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
+        is_causal = True if causal_mask is None and q_len > 1 else False
+
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=causal_mask,
+            dropout_p=self.attention_dropout if self.training else 0.0,
+            is_causal=is_causal,
+        )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.view(bsz, q_len, self.hidden_size)
+
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output, None, past_key_value
+
+
+QWEN2_ATTENTION_CLASSES = {
+    "eager": Qwen2Attention,
+    "flash_attention_2": Qwen2FlashAttention2,
+    "sdpa": Qwen2SdpaAttention,
+}
+
+
+class Qwen2DecoderLayer(nn.Module):
+    def __init__(self, config: Qwen2Config, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        if config.sliding_window and config._attn_implementation != "flash_attention_2":
+            logger.warning_once(
+                f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
+                "unexpected results may be encountered."
+            )
+        self.self_attn = QWEN2_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)
+
+        self.mlp = Qwen2MLP(config)
+        self.input_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        key_position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        blocks_positions=None,
+        **kwargs,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+                `(batch, sequence_length)` where padding elements are indicated by 0.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+                Indices depicting the position of the input sequence tokens in the sequence.
+            kwargs (`dict`, *optional*):
+                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
+                into the model
+        """
+
+        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,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            key_position_ids=key_position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            blocks_positions=blocks_positions,
+        )
+        hidden_states = residual + hidden_states
+        # self.self_attn.record_unselected_block
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+QWEN2_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`Qwen2Config`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+@add_start_docstrings(
+    "The bare Qwen2 Model outputting raw hidden-states without any specific head on top.",
+    QWEN2_START_DOCSTRING,
+)
+class Qwen2PreTrainedModel(PreTrainedModel):
+    config_class = Qwen2Config
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["Qwen2DecoderLayer"]
+    _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_cache_class = True
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+
+QWEN2_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
+            `past_key_values`).
+
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
+            Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
+            returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
+
+            Two formats are allowed:
+            - a [`~cache_utils.Cache`] instance;
+            - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
+            shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
+            cache format.
+
+            The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
+            legacy cache format will be returned.
+
+            If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
+            have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
+            of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+            Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
+            this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
+            the complete sequence length.
+"""
+
+
+@add_start_docstrings(
+    "The bare Qwen2 Model outputting raw hidden-states without any specific head on top.",
+    QWEN2_START_DOCSTRING,
+)
+class Qwen2Model(Qwen2PreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Qwen2DecoderLayer`]
+
+    Args:
+        config: Qwen2Config
+    """
+
+    def __init__(self, config: Qwen2Config):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList(
+            [Qwen2DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self._attn_implementation = config._attn_implementation
+        self.norm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        key_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,
+        cache_position: Optional[torch.LongTensor] = None,
+        blocks_positions=None
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        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
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError(
+                "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
+            )
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        use_legacy_cache = False
+        if use_cache and not isinstance(past_key_values, Cache):
+            use_legacy_cache = True
+            past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+            logger.warning_once(
+                "We detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. "
+                "Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/v4.41.3/en/internal/generation_utils#transformers.Cache)"
+            )
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position = torch.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        causal_mask = self._update_causal_mask(
+            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
+        )
+
+        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 = None
+
+        for decoder_layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states,
+                    causal_mask,
+                    position_ids,
+                    past_key_values,
+                    output_attentions,
+                    use_cache,
+                    cache_position,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=causal_mask,
+                    position_ids=position_ids,
+                    key_position_ids=key_position_ids,
+                    past_key_value=past_key_values,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                    cache_position=cache_position,
+                    blocks_positions=blocks_positions
+                )
+
+            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 = None
+        if use_cache:
+            next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache
+
+        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,
+        )
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaModel._update_causal_mask
+    def _update_causal_mask(
+        self,
+        attention_mask: torch.Tensor,
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_key_values: Cache,
+        output_attentions: bool,
+    ):
+        # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static
+        # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes.
+        # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using
+        # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114
+
+        if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and 0.0 in attention_mask:
+                return attention_mask
+            return None
+
+        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
+        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
+        # to infer the attention mask.
+        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+        using_static_cache = isinstance(past_key_values, StaticCache)
+
+        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+        if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
+            if AttentionMaskConverter._ignore_causal_mask_sdpa(
+                attention_mask,
+                inputs_embeds=input_tensor,
+                past_key_values_length=past_seen_tokens,
+                is_training=self.training,
+            ):
+                return None
+
+        dtype, device = input_tensor.dtype, input_tensor.device
+        min_dtype = torch.finfo(dtype).min
+        sequence_length = input_tensor.shape[1]
+        if using_static_cache:
+            target_length = past_key_values.get_max_length()
+        else:
+            target_length = (
+                attention_mask.shape[-1]
+                if isinstance(attention_mask, torch.Tensor)
+                else past_seen_tokens + sequence_length + 1
+            )
+
+        if attention_mask is not None and attention_mask.dim() == 4:
+            # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
+            if attention_mask.max() != 0:
+                raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")
+            causal_mask = attention_mask
+        else:
+            causal_mask = torch.full(
+                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
+            )
+            if sequence_length != 1:
+                causal_mask = torch.triu(causal_mask, diagonal=1)
+            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+            causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
+            if attention_mask is not None:
+                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+                mask_length = attention_mask.shape[-1]
+                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+                padding_mask = padding_mask == 0
+                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                    padding_mask, min_dtype
+                )
+        if (
+            self.config._attn_implementation == "sdpa"
+            and attention_mask is not None
+            and attention_mask.device.type == "cuda"
+            and not output_attentions
+        ):
+            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+            # Details: https://github.com/pytorch/pytorch/issues/110213
+            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
+
+        return causal_mask
+
+
+class Qwen2ForCausalLM(Qwen2PreTrainedModel):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = Qwen2Model(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    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
+
+    @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        key_position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        blocks_positions=None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, Qwen2ForCausalLM
+
+        >>> model = Qwen2ForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
+        >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```"""
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            key_position_ids=key_position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            cache_position=cache_position,
+            blocks_positions=blocks_positions,
+        )
+
+        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,
+        )
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.prepare_inputs_for_generation
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        cache_position=None,
+        position_ids=None,
+        use_cache=True,
+        **kwargs,
+    ):
+        # If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens
+        # Exception 1: when passing input_embeds, input_ids may be missing entries
+        # Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here
+        if past_key_values is not None:
+            if inputs_embeds is not None:  # Exception 1
+                input_ids = input_ids[:, -cache_position.shape[0] :]
+            elif input_ids.shape[1] != cache_position.shape[0]:  # Default case (the "else", a no op, is Exception 2)
+                input_ids = input_ids[:, cache_position]
+
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -input_ids.shape[1] :]
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and cache_position[0] == 0:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids.contiguous()}  # `contiguous()` needed for compilation use cases
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "cache_position": cache_position,
+                "past_key_values": past_key_values,
+                "use_cache": use_cache,
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+
+
+@add_start_docstrings(
+    """
+    The Qwen2 Model transformer with a sequence classification head on top (linear layer).
+
+    [`Qwen2ForSequenceClassification`] uses the last token in order to do the classification, as other causal models
+    (e.g. GPT-2) do.
+
+    Since it does classification on the last token, it requires to know the position of the last token. If a
+    `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
+    no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
+    padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
+    each row of the batch).
+    """,
+    QWEN2_START_DOCSTRING,
+)
+class Qwen2ForSequenceClassification(Qwen2PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = Qwen2Model(config)
+        self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
+
+        # 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
+
+    @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_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,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = transformer_outputs[0]
+        logits = self.score(hidden_states)
+
+        if input_ids is not None:
+            batch_size = input_ids.shape[0]
+        else:
+            batch_size = inputs_embeds.shape[0]
+
+        if self.config.pad_token_id is None and batch_size != 1:
+            raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
+        if self.config.pad_token_id is None:
+            sequence_lengths = -1
+        else:
+            if input_ids is not None:
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
+            else:
+                sequence_lengths = -1
+
+        pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
+
+        loss = None
+        if labels is not None:
+            labels = labels.to(logits.device)
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(pooled_logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(pooled_logits, labels)
+        if not return_dict:
+            output = (pooled_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutputWithPast(
+            loss=loss,
+            logits=pooled_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    The Qwen2 Model transformer with a token classification head on top (a linear layer on top of the hidden-states
+    output) e.g. for Named-Entity-Recognition (NER) tasks.
+    """,
+    QWEN2_START_DOCSTRING,
+)
+# Copied from transformers.models.llama.modeling_llama.LlamaForTokenClassification with Llama->Qwen2, LLAMA->QWEN2
+class Qwen2ForTokenClassification(Qwen2PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = Qwen2Model(config)
+        if getattr(config, "classifier_dropout", None) is not None:
+            classifier_dropout = config.classifier_dropout
+        elif getattr(config, "hidden_dropout", None) is not None:
+            classifier_dropout = config.hidden_dropout
+        else:
+            classifier_dropout = 0.1
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.score = nn.Linear(config.hidden_size, config.num_labels)
+
+        # 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
+
+    @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_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,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, TokenClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = outputs[0]
+        sequence_output = self.dropout(sequence_output)
+        logits = self.score(sequence_output)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )