Upload 10 files

.gitattributes

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

Param_Calculation.md

@@ -0,0 +1,24 @@
+# Documentation for Parameter Calculation
+
+Thank you for your interest in our Yuan3.0 Model!
+Our community is open to everyone and welcomes all kinds of comments.
+This document is an explanation of Parameter Calculation script.
+
+## Setup for Parameter Calculation
+
+### 1.Download Yuan3.0 Model Locally
+
+### 2.Modify the "MODEL PATH" to Your Local Download Path
+
+```bash
+vim Param_Calculation.py
+# modify the above line
+MODEL_PATH = "/path/to/Yuan3.0-Model"
+```
+
+### 3.Run Parameter Calculation script
+
+```bash
+python Param_Calculation.py
+```
+

Param_Calculation.py

@@ -0,0 +1,51 @@
+import os
+import sys
+import torch
+
+# ================= 配置区域 =================
+# 添加Yuan3.0模型完整路径,将如下路径替换成你的路径
+MODEL_PATH = "/path/to/Yuan3.0-Model"
+
+# 将模型目录设为 Python 搜索路径的第一优先级
+if MODEL_PATH not in sys.path:
+    sys.path.insert(0, MODEL_PATH)
+
+# 设置环境变量，强制离线模式，禁止 HF 联网或访问远程缓存校验
+os.environ["TRANSFORMERS_OFFLINE"] = "1"
+os.environ["HF_DATASETS_OFFLINE"] = "1"
+os.environ["HF_EVALUATE_OFFLINE"] = "1"
+
+from transformers import AutoModel, AutoTokenizer, AutoConfig
+
+print(f"🚀 开始从本地加载模型：{MODEL_PATH}")
+
+# 加载模型
+model = AutoModel.from_pretrained(
+    MODEL_PATH,
+    torch_dtype=torch.bfloat16,
+    low_cpu_mem_usage=True,
+    use_flash_attn=False,
+    device_map="cpu",
+    local_files_only=True,
+    trust_remote_code=True,
+)
+
+print("\n" + "="*30)
+print("--Yuan3.0 Model Parameter--")
+print("="*30)
+
+# 统计参数
+vit_params = 0
+yuan_params = 0
+total_params = model.num_parameters()
+for n, p in model.named_parameters():
+    if 'vision_model' in n:
+        vit_params += p.numel()
+    else:
+        yuan_params += p.numel()
+
+print(f"Vit Model Parameters:     {vit_params:,}")
+print(f"Yuan Model Parameters:     {yuan_params:,}")
+print(f"Total Parameters:     {total_params:,}")
+print("="*30)
+

modeling_intern_vit.py

@@ -0,0 +1,364 @@
+# --------------------------------------------------------
+# InternVL
+# Copyright (c) 2023 OpenGVLab
+# Licensed under The MIT License [see LICENSE for details]
+# --------------------------------------------------------
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from einops import rearrange
+from timm.models.layers import DropPath
+from torch import nn
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import (BaseModelOutput,
+                                           BaseModelOutputWithPooling)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import logging
+
+from .configuration_intern_vit import InternVisionConfig
+
+try:
+    from flash_attention import FlashAttention
+    has_flash_attn = True
+except:
+    print('internvit FlashAttention is not installed.')
+    has_flash_attn = False
+
+
+logger = logging.get_logger(__name__)
+
+
+class InternRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        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)
+
+
+try:
+    from apex.normalization import FusedRMSNorm
+
+    InternRMSNorm = FusedRMSNorm  # noqa
+
+    logger.info('Discovered apex.normalization.FusedRMSNorm - will use it instead of InternRMSNorm')
+except ImportError:
+    # using the normal InternRMSNorm
+    pass
+except Exception:
+    logger.warning('discovered apex but it failed to load, falling back to InternRMSNorm')
+    pass
+
+
+NORM2FN = {
+    'rms_norm': InternRMSNorm,
+    'layer_norm': nn.LayerNorm,
+}
+
+
+class InternVisionEmbeddings(nn.Module):
+    def __init__(self, config: InternVisionConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+
+        self.class_embedding = nn.Parameter(
+            torch.randn(1, 1, self.embed_dim),
+        )
+
+        self.patch_embedding = nn.Conv2d(
+            in_channels=3, out_channels=self.embed_dim, kernel_size=self.patch_size, stride=self.patch_size
+        )
+
+        self.num_patches = (self.image_size // self.patch_size) ** 2
+        self.num_positions = self.num_patches + 1
+
+        self.position_embedding = nn.Parameter(torch.randn(1, self.num_positions, self.embed_dim))
+
+    def _get_pos_embed(self, pos_embed, H, W):
+        target_dtype = pos_embed.dtype
+        pos_embed = pos_embed.float().reshape(
+            1, self.image_size // self.patch_size, self.image_size // self.patch_size, -1).permute(0, 3, 1, 2)
+        pos_embed = F.interpolate(pos_embed, size=(H, W), mode='bicubic', align_corners=False).\
+            reshape(1, -1, H * W).permute(0, 2, 1).to(target_dtype)
+        return pos_embed
+
+    def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
+        target_dtype = self.patch_embedding.weight.dtype
+        patch_embeds = self.patch_embedding(pixel_values)  # shape = [*, channel, width, height]
+        batch_size, _, height, width = patch_embeds.shape
+        patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
+        class_embeds = self.class_embedding.expand(batch_size, 1, -1).to(target_dtype)
+        embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
+        position_embedding = torch.cat([
+            self.position_embedding[:, :1, :],
+            self._get_pos_embed(self.position_embedding[:, 1:, :], height, width)
+        ], dim=1)
+        embeddings = embeddings + position_embedding.to(target_dtype)
+        return embeddings
+
+
+class InternAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: InternVisionConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.use_flash_attn = config.use_flash_attn and has_flash_attn
+        if config.use_flash_attn and not has_flash_attn:
+            print('Warning: Flash Attention is not available, use_flash_attn is set to False.')
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f'embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:'
+                f' {self.num_heads}).'
+            )
+
+        self.scale = self.head_dim ** -0.5
+        self.qkv = nn.Linear(self.embed_dim, 3 * self.embed_dim, bias=config.qkv_bias)
+        self.attn_drop = nn.Dropout(config.attention_dropout)
+        self.proj_drop = nn.Dropout(config.dropout)
+
+        self.qk_normalization = config.qk_normalization
+
+        if self.qk_normalization:
+            self.q_norm = InternRMSNorm(self.embed_dim, eps=config.layer_norm_eps)
+            self.k_norm = InternRMSNorm(self.embed_dim, eps=config.layer_norm_eps)
+
+        if self.use_flash_attn:
+            self.inner_attn = FlashAttention(attention_dropout=config.attention_dropout)
+        self.proj = nn.Linear(self.embed_dim, self.embed_dim)
+
+    def _naive_attn(self, x):
+        B, N, C = x.shape
+        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv.unbind(0)  # make torchscript happy (cannot use tensor as tuple)
+
+        if self.qk_normalization:
+            B_, H_, N_, D_ = q.shape
+            q = self.q_norm(q.transpose(1, 2).flatten(-2, -1)).view(B_, N_, H_, D_).transpose(1, 2)
+            k = self.k_norm(k.transpose(1, 2).flatten(-2, -1)).view(B_, N_, H_, D_).transpose(1, 2)
+
+        attn = ((q * self.scale) @ k.transpose(-2, -1))
+        attn = attn.softmax(dim=-1)
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+    def _flash_attn(self, x, key_padding_mask=None, need_weights=False):
+        qkv = self.qkv(x)
+        qkv = rearrange(qkv, 'b s (three h d) -> b s three h d', three=3, h=self.num_heads)
+
+        if self.qk_normalization:
+            q, k, v = qkv.unbind(2)
+            q = self.q_norm(q.flatten(-2, -1)).view(q.shape)
+            k = self.k_norm(k.flatten(-2, -1)).view(k.shape)
+            qkv = torch.stack([q, k, v], dim=2)
+
+        context, _ = self.inner_attn(
+            qkv, key_padding_mask=key_padding_mask, need_weights=need_weights, causal=False
+        )
+        outs = self.proj(rearrange(context, 'b s h d -> b s (h d)'))
+        outs = self.proj_drop(outs)
+        return outs
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        x = self._naive_attn(hidden_states) if not self.use_flash_attn else self._flash_attn(hidden_states)
+        return x
+
+
+class InternMLP(nn.Module):
+    def __init__(self, config: InternVisionConfig):
+        super().__init__()
+        self.config = config
+        self.act = ACT2FN[config.hidden_act]
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+        return hidden_states
+
+
+class InternVisionEncoderLayer(nn.Module):
+    def __init__(self, config: InternVisionConfig, drop_path_rate: float):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.norm_type = config.norm_type
+
+        self.attn = InternAttention(config)
+        self.mlp = InternMLP(config)
+        self.norm1 = NORM2FN[self.norm_type](self.embed_dim, eps=config.layer_norm_eps)
+        self.norm2 = NORM2FN[self.norm_type](self.embed_dim, eps=config.layer_norm_eps)
+
+        self.ls1 = nn.Parameter(config.initializer_factor * torch.ones(self.embed_dim))
+        self.ls2 = nn.Parameter(config.initializer_factor * torch.ones(self.embed_dim))
+        self.drop_path1 = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity()
+        self.drop_path2 = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity()
+
+    def forward(
+            self,
+            hidden_states: torch.Tensor,
+    ) -> Tuple[torch.FloatTensor, Optional[torch.FloatTensor], Optional[Tuple[torch.FloatTensor]]]:
+        """
+        Args:
+            hidden_states (`Tuple[torch.FloatTensor, Optional[torch.FloatTensor]]`): input to the layer of shape `(batch, seq_len, embed_dim)`
+        """
+
+        hidden_states = hidden_states + self.drop_path1(self.attn(self.norm1(hidden_states)) * self.ls1)
+
+        hidden_states = hidden_states + self.drop_path2(self.mlp(self.norm2(hidden_states)) * self.ls2)
+
+        return hidden_states
+
+
+class InternVisionEncoder(nn.Module):
+    """
+    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
+    [`InternEncoderLayer`].
+
+    Args:
+        config (`InternConfig`):
+            The corresponding vision configuration for the `InternEncoder`.
+    """
+
+    def __init__(self, config: InternVisionConfig):
+        super().__init__()
+        self.config = config
+        # stochastic depth decay rule
+        dpr = [x.item() for x in torch.linspace(0, config.drop_path_rate, config.num_hidden_layers)]
+        self.layers = nn.ModuleList([
+            InternVisionEncoderLayer(config, dpr[idx]) for idx in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = True
+
+    def forward(
+            self,
+            inputs_embeds,
+            output_hidden_states: Optional[bool] = None,
+            return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutput]:
+        r"""
+        Args:
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+                Embedded representation of the inputs. Should be float, not int tokens.
+            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.
+        """
+        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
+
+        encoder_states = () if output_hidden_states else None
+        hidden_states = inputs_embeds
+
+        for idx, encoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    encoder_layer,
+                    hidden_states)
+            else:
+                layer_outputs = encoder_layer(
+                    hidden_states,
+                )
+            hidden_states = layer_outputs
+
+        if output_hidden_states:
+            encoder_states = encoder_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, encoder_states] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=encoder_states
+        )
+
+
+class InternVisionModel(PreTrainedModel):
+    main_input_name = 'pixel_values'
+    config_class = InternVisionConfig
+    _no_split_modules = ['InternVisionEncoderLayer']
+
+    def __init__(self, config: InternVisionConfig):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = InternVisionEmbeddings(config)
+        self.encoder = InternVisionEncoder(config)
+
+    def resize_pos_embeddings(self, old_size, new_size, patch_size):
+        pos_emb = self.embeddings.position_embedding
+        _, num_positions, embed_dim = pos_emb.shape
+        cls_emb = pos_emb[:, :1, :]
+        pos_emb = pos_emb[:, 1:, :].reshape(1, old_size // patch_size, old_size // patch_size, -1).permute(0, 3, 1, 2)
+        pos_emb = F.interpolate(pos_emb.float(), size=new_size // patch_size, mode='bicubic', align_corners=False)
+        pos_emb = pos_emb.to(cls_emb.dtype).reshape(1, embed_dim, -1).permute(0, 2, 1)
+        pos_emb = torch.cat([cls_emb, pos_emb], dim=1)
+        self.embeddings.position_embedding = nn.Parameter(pos_emb)
+        self.embeddings.image_size = new_size
+        logger.info('Resized position embeddings from {} to {}'.format(old_size, new_size))
+
+    def get_input_embeddings(self):
+        return self.embeddings
+
+    def forward(
+            self,
+            pixel_values: Optional[torch.FloatTensor] = None,
+            output_hidden_states: Optional[bool] = None,
+            return_dict: Optional[bool] = None,
+            pixel_embeds: Optional[torch.FloatTensor] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if pixel_values is None and pixel_embeds is None:
+            raise ValueError('You have to specify pixel_values or pixel_embeds')
+
+        if pixel_embeds is not None:
+            hidden_states = pixel_embeds
+        else:
+            if len(pixel_values.shape) == 4:
+                hidden_states = self.embeddings(pixel_values)
+            else:
+                raise ValueError(f'wrong pixel_values size: {pixel_values.shape}')
+        encoder_outputs = self.encoder(
+            inputs_embeds=hidden_states,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        last_hidden_state = encoder_outputs.last_hidden_state
+        pooled_output = last_hidden_state[:, 0, :]
+
+        if not return_dict:
+            return (last_hidden_state, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=last_hidden_state,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )

modeling_yuanlm2.py

@@ -0,0 +1,1502 @@
+# coding=utf-8
+# Copyright 2022 EleutherAI 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 Yuan model."""
+import math
+from typing import List, Optional, Tuple, Union
+import torch.nn.functional as F
+import torch
+import torch.utils.checkpoint
+from torch import einsum, nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, SequenceClassifierOutputWithPast
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings
+from .configuration_yuan import YuanConfig
+from einops import rearrange
+from transformer_engine.pytorch import RMSNorm
+import pdb
+import copy
+try:
+    import grouped_gemm as gg
+except ImportError:
+    gg = None
+try:
+    from flash_attn import flash_attn_varlen_func as flash_attn_unpadded_func
+    from flash_attn import flash_attn_func
+except ImportError:
+    flash_attn_unpadded_func = None
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "YuanConfig"
+
+class YuanRotaryEmbedding(nn.Module):
+    def __init__(self, dim, base=10000, dtype=torch.float32, rotary_interleaved=False, seq_len_interpolation_factor=None):
+        super().__init__()
+        self.base = base
+        self.dim = dim
+        self.rotary_interleaved = rotary_interleaved
+        self.seq_len_interpolation_factor = seq_len_interpolation_factor
+
+    def get_rotary_seq_len(
+            self,
+            inference_param=None,
+            transformer_input: torch.Tensor=None,
+            transformer_config=None,
+    ):
+        if inference_param is not None:
+            rotary_seq_len = inference_param.max_sequence_length
+        else:
+            rotary_seq_len = transformer_input.size[0]
+        if transformer_config.sequence_parallel:
+            rotary_seq_len *= transformer_config.tensor_model_parallel_size
+        
+        return rotary_seq_len
+
+    def forward(self, max_seq_len, offset=0):
+
+        """Forward pass of RoPE embedding.
+
+        Args:
+            max_seq_len (int): Maximum size of sequence
+            offset (int, optional): _description_. Defaults to 0.
+
+        Returns:
+            Tensor: Embeddings after applying RoPE.
+        """
+        inv_freq = (1.0 / ( self.base**(torch.arange(0, self.dim, 2, dtype=torch.float32, device=torch.cuda.current_device()) / self.dim))).to(torch.float32)
+        
+        #max_seq_len_int = max_seq_len.item() if max_seq_len.numel() == 1 else max_seq_len.max().item()
+        seq = (
+            torch.arange(max_seq_len, device=inv_freq.device, dtype=inv_freq.dtype)
+            + offset
+        )
+
+        if self.seq_len_interpolation_factor is not None:
+            seq *= 1 / self.seq_len_interpolation_factor
+
+        freqs = torch.outer(seq, inv_freq)
+        # first part even vector components, second part odd vector components,
+        #  2 * dim in dimension size
+        if not self.rotary_interleaved:
+            emb = torch.cat((freqs, freqs), dim=-1)
+        else:
+            emb = torch.stack((freqs.view(-1, 1), freqs.view(-1, 1)), dim=-1).view(
+                freqs.shape[0], -1
+            )
+        # emb [seq_length, .., dim]
+        emb = emb[:, None, None, :]
+        #emb = emb[:, None, :] 
+        return emb
+
+
+def _rotate_half(x, rotary_interleaved):
+    """huggingface version
+    change sign so the last dimension becomes [-odd, +even]
+    
+    x1, x2 = torch.chunk(x, 2, dim=-1)
+    return torch.cat((-x2, x1), dim=-1)
+    """
+    if not rotary_interleaved:
+        x1, x2 = torch.chunk(x, 2, dim=-1)
+        return torch.cat((-x2, x1), dim=-1)
+    else:
+        x1 = x[:, :, :, ::2]
+        x2 = x[:, :, :, 1::2]
+        x_new = torch.stack((-x2, x1), dim=-1)
+        return x_new.view(x_new.shape[0], x_new.shape[1], x_new.shape[2], -1)
+
+def apply_rotary_pos_emb(t, freqs, position_ids, rotary_interleaved=False):
+
+    rot_dim = freqs.shape[-1]
+    #if position_ids.shape[1] > 1:
+    freqs = freqs[position_ids]
+    freqs = freqs.view(t.shape[1],freqs.shape[1],freqs.shape[2],freqs.shape[4]).transpose(0,1)
+    # ideally t_pass is empty so rotary pos embedding is applied to all tensor t
+    t, t_pass = t[..., :rot_dim], t[..., rot_dim:]
+
+    # first part is cosine component
+    # second part is sine component, need to change signs with _rotate_half method
+    t_type = t.dtype
+    cos_ = torch.cos(freqs).to(t.dtype)
+    sin_ = torch.sin(freqs).to(t.dtype)
+
+    t = (t * cos_) + (_rotate_half(t, rotary_interleaved) * sin_)
+    return torch.cat((t, t_pass), dim=-1)
+
+class LocalizedFiltering(torch.nn.Module):
+    """
+    Mega's Exponential Moving Average layer, largely left unmodified from the original repo with the exception of
+    variable names and moving away from the stateful representation of incremental decoding state. See
+    "https://arxiv.org/abs/2209.10655" for more details.
+    """
+
+    def __init__(self, hidden_size, lf_conv2d_group, lf_conv2d_num_pad, use_lfa_bias):
+        super().__init__()
+
+        self.embed_dim = hidden_size
+        self.lf_conv2d_group = lf_conv2d_group
+        self.lf_conv2d_num_pad = lf_conv2d_num_pad
+        self.use_lfa_bias = use_lfa_bias
+        if self.lf_conv2d_num_pad == 1:
+            self.training = True
+        self.conv1 = torch.nn.Conv2d(self.embed_dim, self.embed_dim // 2, (2, 1), stride=(1, 1), padding=(self.lf_conv2d_num_pad, 0), groups=self.lf_conv2d_group, bias=use_lfa_bias)
+        self.conv2 = torch.nn.Conv2d(self.embed_dim // 2, self.embed_dim, (2, 1), stride=(1, 1), padding=(self.lf_conv2d_num_pad, 0), groups=self.lf_conv2d_group, bias=use_lfa_bias)
+        self.output_layernorm = RMSNorm(self.embed_dim, eps=1e-6)
+
+    def _train_forward(self, inputs):
+        inputs = inputs.transpose(0,1)
+        seq_len, bsz, embed_dim = inputs.size()
+        if embed_dim != self.embed_dim:
+            raise ValueError(
+                f"Unexpected embedding dimension received: input is {embed_dim}, model expects {self.embed_dim}"
+            )
+        residual = inputs
+
+        inputs = inputs.view(seq_len, 1, bsz, embed_dim).permute(2, 3, 0, 1)
+        output1 = self.conv1(inputs)
+        output1 = output1[:, :, :seq_len, :]
+
+        output2 = self.conv2(output1)
+        output2 = output2[:, :, :seq_len, :].permute(2, 3, 0, 1).contiguous()
+        output2 = output2.view(seq_len, bsz, embed_dim)
+        assert output2.shape == residual.shape
+
+        torch.cuda.set_device(output2.device)
+        lf_output = self.output_layernorm(output2 + residual)
+        lf_output = lf_output.transpose(0,1)
+        return lf_output
+
+    def _inference_forward(self, inputs, before_hidden_states):
+
+        if before_hidden_states is None:
+            residual = inputs
+            seq_len, bsz, embed_dim = inputs.size()
+
+            inputs = inputs.view(seq_len, 1, bsz, embed_dim).permute(2, 3, 0, 1)
+
+            pad_zero1 = torch.zeros(bsz, embed_dim, 1, 1).to(inputs)
+            inputs = torch.cat((pad_zero1, inputs), dim=2).contiguous()
+            output1 = self.conv1(inputs)
+
+            pad_zero2 = torch.zeros(bsz, embed_dim // 2, 1, 1).to(output1)
+            output1 = torch.cat((pad_zero2, output1), dim=2).contiguous()
+            output2 = self.conv2(output1)
+
+            output2 = output2.permute(2, 3, 0, 1).contiguous()
+
+            output2 = output2.view(seq_len, bsz, embed_dim)
+
+            assert output2.shape == residual.shape
+
+            lf_output = self.output_layernorm(output2 + residual)
+
+        else:
+            residual = inputs
+
+            seq_len, bsz, embed_dim = inputs.size()
+            seq_len_before, _, _ = before_hidden_states.size()
+
+            assert seq_len == 1 and seq_len_before == 2
+
+            inputs = torch.cat((before_hidden_states, inputs), dim=0)
+            inputs = inputs.view(3, 1, bsz, embed_dim).permute(2, 3, 0, 1)
+
+            output1 = self.conv1(inputs)
+            output2 = self.conv2(output1)
+            output2 = output2.view(1, bsz, embed_dim)
+
+            assert output2.shape == residual.shape
+
+            lf_output = self.output_layernorm(output2 + residual)
+
+        return lf_output
+
+    def forward(
+        self,
+        inputs,
+        before_hidden_states = None,
+    ) -> torch.Tensor:
+        # assert self.lf_conv2d_num_pad == 1
+        if self.training:
+            lf_output = self._train_forward(inputs)
+        else:
+            lf_output = self._inference_forward(inputs, before_hidden_states)
+
+        return lf_output
+
+
+# Copied from transformers.models.bart.modeling_bart._make_causal_mask
+def _make_causal_mask(
+    input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
+):
+    """
+    Make causal mask used for bi-directional self-attention.
+    """
+    bsz, tgt_len = input_ids_shape
+    mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min, device=device), device=device)
+    mask_cond = torch.arange(mask.size(-1), device=device)
+    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
+    mask = mask.to(dtype)
+
+    if past_key_values_length > 0:
+        mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
+    return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
+
+
+# Copied from transformers.models.bart.modeling_bart._expand_mask
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    """
+    bsz, src_len = mask.size()
+    tgt_len = tgt_len if tgt_len is not None else src_len
+
+    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
+
+    inverted_mask = 1.0 - expanded_mask
+
+    return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
+
+
+class YuanRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        YuanRMSNorm is equivalent to LlamaRMSNorm
+        """
+        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)
+
+# flash attn
+class FlashSelfAttention(torch.nn.Module):
+    """Implement the scaled dot product attention with softmax.
+    Arguments
+    ---------
+        softmax_scale: The temperature to use for the softmax attention.
+                      (default: 1/sqrt(d_keys) where d_keys is computed at
+                      runtime)
+        attention_dropout: The dropout rate to apply to the attention
+                           (default: 0.0)
+    """
+    def __init__(self, causal=False, softmax_scale=None, attention_dropout=0.0,
+                 device=None, dtype=None):
+        super().__init__()
+        assert flash_attn_unpadded_func is not None, ('Please install FlashAttention first, '
+                                                      'e.g., with pip install flash-attn')
+        assert rearrange is not None, 'Please install einops first, e.g., with pip install einops'
+        self.causal = causal
+        self.softmax_scale = softmax_scale
+        self.dropout_p = attention_dropout
+
+    def forward(self, q, k, v):
+        """Implements the multihead softmax attention.
+        Arguments
+        ---------
+            q, k, v: The tensor containing the query, key, and value. (B, S, H, D)
+        """
+
+        assert all((i.dtype in [torch.float16, torch.bfloat16] for i in (q,k,v)))
+        assert all((i.is_cuda for i in (q,k,v)))
+
+        batch_size, seqlen_q = q.shape[1], q.shape[0]
+        seqlen_k = k.shape[0]
+        q, k, v = [rearrange(x, 'b s ... -> (b s) ...') for x in [q, k, v]]
+        cu_seqlens_q = torch.arange(0, (batch_size + 1) * seqlen_q, step=seqlen_q, dtype=torch.int32, device=q.device)
+        if self.training:
+            # during training q,k,v always have same seqlen
+            assert seqlen_k == seqlen_q
+            is_causal = self.causal
+            cu_seqlens_k = cu_seqlens_q
+            dropout_p = self.dropout_p
+        else:
+            # turn off FA causal mask after first inference autoregressive iteration
+            # only on first autoregressive step q,k,v have same seqlen
+            is_causal = seqlen_q == seqlen_k
+            cu_seqlens_k = torch.arange(0, (batch_size + 1) * seqlen_k, step=seqlen_k, dtype=torch.int32, device=q.device) 
+            #cu_seqlens_q = [cu_seqlens_q[0], cu_seqlens_q[-1]]
+            #cu_seqlens_k = [cu_seqlens_k[0], cu_seqlens_k[-1]]
+            dropout_p = 0
+
+        output = flash_attn_unpadded_func(q, k, v, cu_seqlens_q, cu_seqlens_k, seqlen_q, seqlen_k, dropout_p, softmax_scale=self.softmax_scale, causal=is_causal)
+
+        output = rearrange(output, '(b s) ... -> b s ...', b=batch_size)
+        return output
+
+class ParallelAttention_router(nn.Module):
+    def __init__(self, config, num_experts):
+        super(ParallelAttention_router, self).__init__()
+        layer_number=0
+        self.layer_number = max(1, layer_number)
+        
+        self.hidden_size = config.hidden_size
+        self.projection_size = num_experts
+        
+        self.num_attention_router_heads = config.moe_config['num_attention_router_heads']
+        self.hidden_size_per_attention_head = config.max_position_embeddings // self.num_attention_router_heads
+        self.query_key_value = nn.Linear(self.hidden_size, self.projection_size*3, bias=False)
+
+    def forward(self, hidden_states, attention_mask=None, enc_position_ids=None,
+                encoder_output=None, inference_params=None,
+                rotary_pos_emb=None):
+        is_first_step = False
+        before_hidden_states = None
+        
+        #mixed_x_layer = torch.matmul(hidden_states, self.query_key_value)
+        mixed_x_layer = self.query_key_value(hidden_states)
+        (query_layer, key_layer, value_layer) = torch.split(mixed_x_layer, self.projection_size, -1)
+        b, s, z = query_layer.shape
+
+        # use fp32 router
+        query_layer = query_layer.float().view(b,s,z,1)
+        key_layer = key_layer.float().view(b,s,z,1)
+        value_layer = value_layer.float().view(b,s,z,1)
+
+        attn_weights = torch.matmul(query_layer, key_layer.transpose(2, 3))
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+        attn_output = torch.matmul(attn_weights, value_layer)
+        router_output = attn_output.view(-1, z)
+        return router_output 
+
+class YuanExpertMLP(nn.Module):
+    def __init__(self, config):
+        super(YuanExpertMLP, self).__init__()
+        self.gated_linear_unit = config.moe_config['gated_linear_unit']
+        #self.ffn_hidden_size = config.moe_config['ffn_hidden_size']
+        self.ffn_hidden_size = config.ffn_hidden_size
+
+
+        if self.gated_linear_unit:
+            self.w1 = nn.Linear(config.hidden_size, self.ffn_hidden_size*2, bias=False)
+              
+        else:
+            self.w1 = nn.Linear(config.hidden_size, self.ffn_hidden_size, bias=False)
+            
+        self.act_fn = ACT2FN[config.hidden_act]
+        self.w2 = nn.Linear(self.ffn_hidden_size, config.hidden_size, bias=False)
+        
+
+    def forward(self, x):
+        x = self.w1(x)
+        if self.gated_linear_unit:
+            x = torch.chunk(x, 2, dim=-1)
+            x = self.act_fn(x[0]) * x[1]
+        else:
+            x = self.act_fn(x)  
+        x = self.w2(x)  
+        return x
+    
+
+
+class YuanMLP(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str
+    ):
+        super().__init__()
+        self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
+        self.act_fn = ACT2FN[hidden_act]
+
+    def forward(self, x):
+        return self.down_proj(self.gate_proj(x) * self.act_fn(self.up_proj(x)))
+        
+
+class YuanAttention(nn.Module):
+    """Localized Filtering-based Attention 'YUAN 2.0: A Large Language Model with Localized Filtering-based Attention' paper"""
+
+    def __init__(self, config: YuanConfig):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.lf_conv2d_group = config.lf_conv2d_group
+        self.lf_conv2d_num_pad = config.lf_conv2d_num_pad
+        self.use_lfa_bias = config.use_lfa_bias
+        try:
+            self.attention_projection_size = config.attention_projection_size
+        except:
+            self.attention_projection_size = None
+        
+        if self.attention_projection_size is None:
+            self.head_dim = self.hidden_size // self.num_heads
+        else:
+            self.head_dim = self.attention_projection_size // self.num_heads
+
+        self.max_position_embeddings = config.max_position_embeddings
+        self.causal_mask = config.causal_mask
+        self.attn_mask_type = config.attn_mask_type
+        self.softmax_scale = 1.0 / math.sqrt(self.head_dim)
+        self.use_flash_attention = config.use_flash_attention
+        try:
+            self.use_shareqk = config.use_shareqk
+        except Exception as e:
+            self.use_shareqk=False
+        self.dropout = 0.0
+        self.attention_projection_size = config.attention_projection_size
+        self.v_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+        
+        if self.use_shareqk:
+            self.qk_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+            self.qk_weight = nn.Parameter(torch.Tensor(2, self.hidden_size))
+            self.qk_bias = nn.Parameter(torch.Tensor(2, self.hidden_size))
+        else:
+            self.lf_gate = LocalizedFiltering(self.hidden_size, self.lf_conv2d_group, self.lf_conv2d_num_pad, self.use_lfa_bias)
+            self.get_query_key = nn.Linear(self.hidden_size, 2 * self.attention_projection_size, bias=False)
+        self.core_attention = FlashSelfAttention(causal=True, attention_dropout=config.attn_dropout, softmax_scale=self.softmax_scale)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        position_ids_k: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        rotary_pos_emb: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        
+        q_len, bsz, _ = hidden_states.size()
+        hidden_states = hidden_states#.to('cuda:1')
+        is_first_step = False
+        if use_cache:
+            if past_key_value is None:
+                before_hidden_states = None
+                is_first_step = True
+                if q_len > 1:
+                    inference_hidden_states_memory = hidden_states[-2:, :, :]
+                else:
+                    inference_hidden_states_memory = torch.cat((torch.zeros_like(hidden_states), hidden_states), dim=0)
+            else:
+                before_hidden_states = past_key_value[2]
+                inference_hidden_states_memory = torch.cat((before_hidden_states[-1:, :, :], hidden_states), dim=0)
+        value_states = self.v_proj(hidden_states).view(q_len, bsz, self.num_heads, self.head_dim)
+        if self.use_shareqk:
+            qk_states = self.qk_proj(hidden_states).view(q_len, bsz, self.num_heads*self.head_dim)
+            query_key = qk_states.unsqueeze(2) * self.qk_weight + self.qk_bias
+            query_states, key_states = torch.unbind(query_key, dim=2)
+
+            query_states = query_states.view(q_len, bsz, self.num_heads, self.head_dim).transpose(1, 2)
+            key_states = key_states.view(q_len, bsz, self.num_heads, self.head_dim).transpose(1, 2)
+        else:
+            hidden_states = self.lf_gate(hidden_states, before_hidden_states)
+            mixed_qk_layer = self.get_query_key(hidden_states)
+            #mixed_qk_layer = torch.matmul(hidden_states, qk_tensor)
+            new_tensor_shape = mixed_qk_layer.size()[:-1] + (self.num_heads, 2 * self.head_dim)
+            mixed_qk_layer = mixed_qk_layer.view(*new_tensor_shape)
+            (query_states, key_states) = torch.split(mixed_qk_layer, self.head_dim, dim=-1)
+
+        
+        kv_seq_len = key_states.shape[1]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value[0].shape[1]
+        
+        # duplicate the pos_emb for self attention
+        if rotary_pos_emb is not None:
+            if position_ids.shape[1] == 1:
+                q_seq_start = position_ids[0,-1]
+                #seq_start = past_key_value[0].shape[0]
+                q_seq_end = q_seq_start + 1
+                k_seq_end = q_seq_end
+            else:
+                q_seq_start = 0
+                q_seq_end = q_seq_start+key_states.shape[0]
+                k_seq_end = q_seq_end
+            
+            rotary_pos_shape = rotary_pos_emb.shape
+            if isinstance(rotary_pos_emb, tuple):
+                rotary_pos_emb = rotary_pos_emb
+            else:
+                rotary_pos_emb = ((rotary_pos_emb,) * 2)
+            q_pos_emb, k_pos_emb = rotary_pos_emb
+        if past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = torch.cat([past_key_value[0], key_states], dim=0)
+            value_states = torch.cat([past_key_value[1], value_states], dim=0)
+        past_key_value = (key_states, value_states, inference_hidden_states_memory) if use_cache else None
+        query_states = apply_rotary_pos_emb(query_states, q_pos_emb, position_ids)
+        key_states = apply_rotary_pos_emb(key_states, k_pos_emb, position_ids_k)
+
+        attn_weights = None
+        attn_output = self.core_attention(query_states, key_states, value_states)
+        q_len, bsz, _, _ = attn_output.shape
+        attn_output = attn_output.reshape(q_len, bsz, -1)
+        
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output, attn_weights, past_key_value
+
+class MoEDroplessTokenDispatcher:
+    def __init__(self, num_experts: int, config: YuanConfig) -> None:
+        self.num_experts = num_experts
+        assert self.num_experts > 0, "Expected at least one expert"
+        self.router_topk = config.moe_config['moe_top_k']
+
+    def token_permutation(
+        self, hidden_states: torch.Tensor, max_prob: torch.Tensor, max_ind: torch.Tensor
+    ):
+        self.hidden_shape = hidden_states.shape
+        hidden_states = hidden_states.view(-1, self.hidden_shape[-1])
+
+        if self.router_topk > 1:
+            global_local_map = torch.ones_like(max_ind).bool()
+            local_indices = max_ind.masked_select(global_local_map)
+            local_probs = max_prob.masked_select(global_local_map)
+            global_local_map = global_local_map.nonzero()[:, 0]
+            global_local_map = global_local_map.view(-1, 1).expand(-1, hidden_states.shape[-1])
+            local_hidden_states = torch.gather(hidden_states, 0, global_local_map)
+
+        indices = torch.argsort(local_indices, dim=0)
+        tokens_per_expert = torch.histc(
+            local_indices,
+            bins=self.num_experts,
+            min=0,
+            max=self.num_experts - 1,
+        )
+        tokens_per_expert = tokens_per_expert.cpu().to(torch.long)
+
+        indices = indices.view(-1, 1).expand(-1, hidden_states.shape[-1])
+        permuted_local_hidden_states = torch.gather(local_hidden_states, 0, indices)
+        return (permuted_local_hidden_states, tokens_per_expert, local_probs, indices, global_local_map)
+
+    def token_unpermutation(
+        self,
+        hidden_states: torch.Tensor,
+        scores: torch.Tensor,
+        indices: torch.Tensor,
+        global_local_map: torch.Tensor = None,
+    ):
+        scores = scores.to(dtype=hidden_states.dtype)
+        unpermuted_local_hidden = torch.zeros_like(hidden_states)
+        assert indices.shape == hidden_states.shape, f'{indices.shape}, {hidden_states.shape}'
+        unpermuted_local_hidden = unpermuted_local_hidden.scatter(0, indices, hidden_states)
+
+        if self.router_topk > 1:
+            unpermuted_local_hidden = unpermuted_local_hidden * scores.view(-1, 1)
+        unpermuted_local_bias = None
+        output_total = unpermuted_local_hidden
+        output_bias_total = unpermuted_local_bias
+
+        if self.router_topk > 1:
+            global_num_tokens = self.hidden_shape[0] * self.hidden_shape[1]
+            global_hidden_shape = [global_num_tokens, hidden_states.shape[-1]]
+            unpermuted_global_hidden = torch.zeros(
+                global_hidden_shape,
+                dtype=hidden_states.dtype,
+                device=hidden_states.device,
+            )
+            output_total = unpermuted_global_hidden.scatter_add(
+                0, global_local_map, unpermuted_local_hidden
+            )
+
+        output_total = output_total.view(self.hidden_shape)
+
+        return output_total
+
+class GroupedMLP(nn.Module):
+    """An efficient implementation of the Experts layer using CUTLASS GroupedGEMM.
+
+    This class is designed to execute multiple experts in parallel, thereby maximizing computational efficiency.
+    """
+
+    def __init__(self, num_experts: int, config: YuanConfig):
+        super().__init__()
+        self.num_experts = num_experts
+        self.config = config
+
+        def glu(x):
+            x = torch.chunk(x, 2, dim=-1)
+            return torch.nn.functional.silu(x[0]) * x[1]
+
+        self.activation_func = glu
+        self.ffn_hidden_size = config.ffn_hidden_size
+        fc1_output_size_per_partition = self.ffn_hidden_size * 2
+        fc2_input_size = self.ffn_hidden_size
+        
+        self.w1 = nn.ModuleList([nn.Linear(self.config.hidden_size, self.ffn_hidden_size * 2, bias=False) for _ in range(num_experts)])
+        self.w2 = nn.ModuleList([nn.Linear(self.ffn_hidden_size, self.config.hidden_size, bias=False) for _ in range(num_experts)])
+    def forward(self, permuted_hidden_states, tokens_per_expert):
+        torch.cuda.set_device(permuted_hidden_states.device)
+        permuted_hidden_states = permuted_hidden_states#.to('cuda:0')
+
+        fc2_outputs = []
+        start_idx = 0
+        for i in range(self.num_experts):
+            if tokens_per_expert[i] == 0:
+                continue
+            end_idx = start_idx + tokens_per_expert[i]
+            # Use custom attributes for each expert's Linear layers
+            
+            fc1_output = self.w1[i](permuted_hidden_states[start_idx:end_idx])
+            intermediate_parallel = self.activation_func(fc1_output)
+            fc2_output = self.w2[i](intermediate_parallel)
+            fc2_outputs.append(fc2_output)
+            start_idx = end_idx
+        fc2_output = torch.cat(fc2_outputs, dim=0)
+        return fc2_output#.to('cuda:1')
+
+class YuanMoeLayer(nn.Module):
+    def __init__(self, config:YuanConfig, num_layer):
+        super().__init__()
+        self.config = config
+        if 'per_layer_experts_blocks' in config.moe_config:
+            assert config.moe_config['per_layer_experts_blocks'] != None
+            self.num_experts = config.moe_config['per_layer_experts_blocks'][num_layer]
+        elif 'moe_num_experts' in config.moe_config:
+            assert config.moe_config['moe_num_experts'] != None
+            self.num_experts = config.moe_config['moe_num_experts']
+        # self.num_experts = config.moe_config['moe_num_experts']
+        self.top_k = config.moe_config['moe_top_k']
+        self.norm_topk_prob = config.moe_config['norm_topk_prob']
+        self.hidden_size = config.hidden_size
+        
+        expert_indices_offset = (0)
+
+        if config.moe_config['router_type'] == 'attn_router':
+            self.router = ParallelAttention_router(config, self.num_experts)
+        else:
+            self.router = nn.Linear(config.hidden_size, self.num_experts, bias=False)
+
+        self.token_dispatcher = MoEDroplessTokenDispatcher(self.num_experts, config=self.config)
+        self.experts = GroupedMLP(self.num_experts, self.config)
+
+    def routing(self, logits: torch.Tensor) -> torch.Tensor:
+        top_logits, indices = torch.topk(logits, k=self.top_k, dim=1)
+        scores = torch.softmax(top_logits, dim=-1, dtype=torch.float32).type_as(logits)
+        return scores, indices
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        batch_size, sequence_length, hidden_dim = hidden_states.shape
+        logits = self.router(hidden_states)
+        scores, indices = self.routing(logits)
+        scores = scores.to(hidden_states.dtype)
+        (dispatched_input, tokens_per_expert, scores, indices, global_local_map, ) = self.token_dispatcher.token_permutation(hidden_states, scores, indices)
+        expert_output = self.experts(dispatched_input, tokens_per_expert)
+        output = self.token_dispatcher.token_unpermutation(expert_output, scores, indices, global_local_map)
+        return output
+
+class YuanDecoderLayer(nn.Module):
+    def __init__(self, config: YuanConfig, num_layer):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = YuanAttention(config=config)
+        self.num_layer = num_layer
+        
+        if hasattr(config, 'moe_config'): #moe_config['num_moe_experts'] > 0 or hasattr(config.moe_config, 'per_layer_experts_blocks'):
+            self.mlp = YuanMoeLayer(config, num_layer)
+        else:
+            self.mlp = YuanMLP(
+                hidden_size=self.hidden_size,
+                intermediate_size=config.intermediate_size,
+                hidden_act=config.hidden_act,
+            )
+        
+        
+        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        position_ids_k: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        rotary_pos_emb: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+    ) -> 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, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+        """
+        residual = hidden_states#.to('cuda:1')
+        torch.cuda.set_device(hidden_states.device)
+        hidden_states = self.input_layernorm(hidden_states) #.to('cuda:0')).to('cuda:1')
+        
+        # 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,
+            position_ids_k=position_ids_k,
+            past_key_value=past_key_value,
+            rotary_pos_emb=rotary_pos_emb,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+        )
+
+        hidden_states = residual + hidden_states.permute(1, 0, 2)
+
+        # Fully Connected
+        residual = hidden_states#.to('cuda:1')
+        torch.cuda.set_device(hidden_states.device)
+        hidden_states = self.post_attention_layernorm(hidden_states) #.to('cuda:0')).to('cuda:1')
+        hidden_states = self.mlp(hidden_states)# .to('cuda:1')
+        hidden_states = residual + hidden_states
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+    
+        return outputs
+
+
+YUAN_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 ([`YuanConfig`]):
+            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 Yuan Model outputting raw hidden-states without any specific head on top.",
+   YUAN_START_DOCSTRING,
+)
+class YuanPreTrainedModel(PreTrainedModel):
+    config_class = YuanConfig 
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["YuanDecoderLayer"]
+    _skip_keys_device_placement = "past_key_values"
+    _keys_to_ignore_on_load_unexpected = [r"decoder\.version"]
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, YuanModel):
+            module.gradient_checkpointing = value
+
+
+YUAN_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 (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            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)`) and 2 additional tensors of shape
+            `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_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.
+"""
+
+
+@add_start_docstrings(
+    "The bare Yuan Model outputting raw hidden-states without any specific head on top.",
+    YUAN_START_DOCSTRING,
+)
+class YuanModel(YuanPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`YuanDecoderLayer`]
+
+    Args:
+        config: YuanConfig
+    """
+
+    def __init__(self, config: YuanConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+        
+        #TODO: control it by config
+        self.eod_token = config.eod_token
+        self.reset_attention_mask = config.reset_attention_mask
+        self.reset_position_ids = config.reset_position_ids
+        self.max_position_embeddings = config.max_position_embeddings
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList([YuanDecoderLayer(config, i) for i in range(config.num_hidden_layers)])
+        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+        self.seq_length = config.max_position_embeddings
+        rotary_dim = config.hidden_size // config.num_attention_heads
+        if config.rotary_percent < 1.0:
+            rotary_dim = int(rotary_dim * config.rotary_percent)
+        self.rotary_pos_emb = YuanRotaryEmbedding(rotary_dim, base=config.rotary_base, dtype=config.torch_dtype)
+
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.embed_tokens(input_ids)
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
+    def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
+        # create causal mask
+        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+        combined_attention_mask = None
+        if input_shape[-1] > 1:
+            combined_attention_mask = _make_causal_mask(
+                input_shape,
+                inputs_embeds.dtype,
+                device=inputs_embeds.device,
+                past_key_values_length=past_key_values_length,
+            )
+
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
+                inputs_embeds.device
+            )
+            combined_attention_mask = (
+                expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
+            )
+
+        return combined_attention_mask
+
+    def _prepare_decoder_attention_mask_training(self, input_id, inputs_embeds, eod_token, reset_mask_flag ,reset_attention_mask=True, reset_position_ids=True):
+    
+        micro_batch_size, seq_length = input_id.size()
+        
+        attention_mask = torch.tril(torch.ones(
+            (micro_batch_size, seq_length, seq_length), device=inputs_embeds.device)).view(
+                micro_batch_size, 1, seq_length, seq_length)    
+                
+        position_ids = torch.arange(seq_length, dtype=torch.long,
+                                    device=inputs_embeds.device)
+        position_ids = position_ids.unsqueeze(0).expand_as(input_id)
+               
+        if reset_position_ids:
+            position_ids = position_ids.clone()
+        
+        if reset_position_ids or reset_attention_mask:
+            # Loop through the batches:
+            for b in range(micro_batch_size):
+
+                # Find indecies where EOD token is.
+                eod_index = position_ids[b, input_id[b] == eod_token]
+
+                # Detach indecies from positions if going to modify positions.
+                if reset_position_ids:
+                    eod_index = eod_index.clone()
+                # Loop through EOD indecies:
+                prev_index = 0
+                for j in range(eod_index.size()[0]):
+                    i = eod_index[j]
+                    # Mask attention loss.
+                    if reset_attention_mask:
+                        attention_mask[b, 0, (i + 1):, :(i + 1)] = 0
+                    # Reset positions.
+                    if reset_position_ids:
+                        position_ids[b, (i + 1):] -= (i + 1 - prev_index)
+                        prev_index = i + 1
+                        
+        inverted_mask = 1 - attention_mask   
+        output_attn_mask = inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(inputs_embeds.dtype).min)
+        if reset_mask_flag:
+            output_attn_mask = output_attn_mask[:,:,-1:,:]
+        return output_attn_mask, position_ids
+
+    @add_start_docstrings_to_model_forward(YUAN_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,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_router_logits: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast, torch.Tensor]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_router_logits = (
+            output_router_logits if output_router_logits is not None else self.config.output_router_logits
+        )
+        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
+        input_ids1 = copy.deepcopy(input_ids)
+        reset_mask_flag = False
+        if past_key_values:
+            input_ids = input_ids
+            input_ids = input_ids[:,-1:]
+            if use_cache:
+                reset_mask_flag = True
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_ids = input_ids
+            batch_size, seq_length = input_ids.shape
+        elif inputs_embeds is not None:
+            inputs_embeds = inputs_embeds.transpose(0,1)
+            batch_size, seq_length, _ = inputs_embeds.shape
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        seq_length_with_past = seq_length
+        past_key_values_length = 0
+
+        if past_key_values is not None:
+            past_key_values_length = past_key_values[0][0].shape[0]
+            seq_length_with_past = seq_length_with_past + past_key_values_length
+
+        # modify to reset position ids
+        if past_key_values is not None:
+            pos_start = position_ids[:,-1]+1
+            pos_end = pos_start+past_key_values[0][0].shape[0]-position_ids.shape[1]+1
+            position_ids_k = torch.arange(pos_start.item(), pos_end.item()).to(position_ids.device)
+            position_ids_k = position_ids_k.unsqueeze(0)
+            position_ids_k = torch.cat((position_ids, position_ids_k), dim=1)
+            position_ids = position_ids[:,-1]+past_key_values[0][0].shape[0]-position_ids.shape[1]+1
+            position_ids = position_ids.unsqueeze(0)
+        else:
+            position_ids_k = position_ids
+
+        if position_ids is None:
+            device = input_ids.device if input_ids is not None else inputs_embeds.device
+            position_ids = torch.arange(
+                past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
+            )
+            position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
+        else:
+            pass
+        
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids).transpose(0,1)
+        
+        if self.training or self.reset_position_ids:
+            attention_mask, _ = self._prepare_decoder_attention_mask_training(input_ids1, inputs_embeds, self.eod_token, reset_mask_flag, self.reset_attention_mask, self.reset_position_ids)
+        else: 
+            if attention_mask is None:
+                attention_mask = torch.ones(
+                    (batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device
+                )
+                attention_mask = self._prepare_decoder_attention_mask(
+                    attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
+                )
+
+        rotary_pos_emb = None
+        rotary_pos_emb = self.rotary_pos_emb(self.max_position_embeddings)
+
+        hidden_states = inputs_embeds
+        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
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = () if use_cache else None
+        position_ids = position_ids.cpu()
+        position_ids_k = position_ids_k.cpu()
+        for idx, decoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, output_attentions, None)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(decoder_layer),
+                    hidden_states,
+                    attention_mask,
+                    position_ids,
+                    None,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    position_ids_k=position_ids_k,
+                    past_key_value=past_key_value,
+                    rotary_pos_emb=rotary_pos_emb,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                )
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+        hidden_states = hidden_states#.to('cuda:0')
+        torch.cuda.set_device(hidden_states.device)
+        hidden_states = self.norm(hidden_states)
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+        next_cache = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+
+class YuanForCausalLM(YuanPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.model = YuanModel(config)
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+    
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    def get_loss_mask(self, input_ids, labels, eod_token, sep_token):
+        micro_batch_size, seq_length = input_ids.size()
+        loss_mask = torch.ones(input_ids.size(), dtype=torch.float, device=input_ids.device)
+        position_ids = torch.arange(seq_length, dtype=torch.long, device=input_ids.device)
+        position_ids = position_ids.unsqueeze(0).expand_as(input_ids) 
+
+
+        """modify loss_mask to only calculate the loss of the answer (separated with [SEP])"""
+
+        for b in range(micro_batch_size):
+            eod_indexs = position_ids[b, input_ids[b] == eod_token]
+            sep_indexs = position_ids[b, input_ids[b] == sep_token]
+
+            if len(eod_indexs) == 0 or len(sep_indexs) == 0:
+                loss_mask[b] = 1.0
+            else:
+                if eod_indexs[0] > sep_indexs[0]:
+                    loss_mask[b, 0:sep_indexs[0]] = 0
+
+                    if len(eod_indexs) == len(sep_indexs):
+                        for ii, eod_index in enumerate(eod_indexs):
+                            start_index = eod_index
+                            if ii == (len(sep_indexs) - 1):
+                                stop_index = seq_length
+                            else:
+                                stop_index = sep_indexs[ii + 1]
+                            loss_mask[b, start_index:stop_index] = 0.0
+                    else:
+                        if len(eod_indexs) > len(sep_indexs):
+                            loss_mask[b,:] = 1.0
+                        else:
+                            for ii, eod_index in enumerate(eod_indexs):
+                                start_index = eod_index
+                                stop_index = sep_indexs[ii + 1]
+
+                                loss_mask[b, start_index:stop_index] = 0.0
+
+                elif eod_indexs[0] < sep_indexs[0]:
+
+                    if len(eod_indexs) == len(sep_indexs):
+                        for ii, eod_index in enumerate(eod_indexs):
+                            start_index = eod_index
+                            stop_index = sep_indexs[ii]
+                            loss_mask[b, start_index:stop_index] = 0.0
+
+                    else:
+                        if len(eod_indexs) < len(sep_indexs):
+                            loss_mask[b,:] = 1.0
+                        else:
+                            for ii, eod_index in enumerate(eod_indexs):
+                                start_index = eod_index
+                                if ii >= len(sep_indexs):
+                                    stop_index = seq_length
+                                else:
+                                    stop_index = sep_indexs[ii]
+                                loss_mask[b, start_index:stop_index] = 0.0
+
+        loss_mask[input_ids == eod_token] = 1.0
+        return loss_mask
+    @add_start_docstrings_to_model_forward(YUAN_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,
+        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, CausalLMOutputWithPast]:
+        """
+        ## modify delete routers
+        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, YuanForCausalLM
+
+        >>> model = YuanForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
+        >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
+
+        >>> prompt = "Hey, are you consciours? 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 consciours? Can you talk to me?\nI'm not consciours, but I can talk to you."
+        ```"""
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0].transpose(0,1)
+        
+        logits = self.lm_head(hidden_states)
+        
+        loss = None
+        if labels is not None:
+            if self.use_loss_mask:
+                loss_mask = self.get_loss_mask(input_ids, labels, self.eod_token, self.sep_token)
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            if self.use_loss_mask:
+                loss_fct = CrossEntropyLoss(reduction='none')
+                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)
+                loss = torch.sum(loss * loss_mask) / loss_mask.sum()
+            else:
+                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=hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def prepare_inputs_for_generation(
+        self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
+    ):
+
+        position_ids = kwargs.get("position_ids", None)
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -1].unsqueeze(-1)
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+
+    @staticmethod
+    def _reorder_cache(past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
+        return reordered_past
+
+
+@add_start_docstrings(
+    """
+    The Yuan Model transformer with a sequence classification head on top (linear layer).
+
+    [`YuanForSequenceClassification`] 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).
+    """,
+    YUAN_START_DOCSTRING,
+)
+class YuanForSequenceClassification(YuanPreTrainedModel):
+    #_keys_to_ignore_on_load_missing = [r"lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = YuanModel(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(YUAN_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]:
+        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:
+                sequence_lengths = (torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1).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,
+        )
+
+
+

modeling_yuanvl_chat.py

@@ -0,0 +1,363 @@
+# --------------------------------------------------------
+# YuanVL
+# Copyright (c) 2024 OpenGVLab
+# Licensed under The MIT License [see LICENSE for details]
+# --------------------------------------------------------
+
+import warnings
+from typing import (Any, Callable, Iterable, List, Literal, Mapping, Optional,
+                    Set, Tuple, Type, TypedDict, Union)
+
+import torch.utils.checkpoint
+import transformers
+import torch
+from torch import nn
+from torch.nn import CrossEntropyLoss
+from transformers import (AutoModel, GenerationConfig, LlamaForCausalLM,
+                          LlamaTokenizer)
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import ModelOutput, logging
+
+from transformer_engine.pytorch import RMSNorm
+from transformers.activations import ACT2FN
+
+from .configuration_yuanvl import YuanVLChatConfig
+from .conversation import get_conv_template
+from .modeling_intern_vit import InternVisionModel, has_flash_attn
+from .modeling_yuanlm2 import YuanForCausalLM
+from .utils import flatten_bn, merge_multimodal_embeddings
+
+logger = logging.get_logger(__name__)
+
+
+class InternVLImagePixelInputs(TypedDict):
+    type: Literal["pixel_values"]
+    data: Union[torch.Tensor, List[torch.Tensor]]
+    """
+    Shape: `(batch_size, 1 + num_patches, num_channels, height, width)`
+
+    Note that `num_patches` may be different for each batch, in which case
+    the data is passed as a list instead of a batched tensor.
+    """
+    patches_per_image: List[int]
+    """
+    List of number of total patches for each image in the batch.
+    """
+
+
+class InternVLImageEmbeddingInputs(TypedDict):
+    type: Literal["image_embeds"]
+    data: Any # in vllm vision this is a NestedTensors
+    """
+    A tensor of shape `(num_images, total_image_feature_size, hidden_size)`
+    or a list of tensors of shape `(total_image_feature_size, hidden_size)`
+
+    `hidden_size` must match the hidden size of language model backbone.
+    """
+
+
+InternVLImageInputs = Union[InternVLImagePixelInputs,
+                            InternVLImageEmbeddingInputs]
+
+
+def version_cmp(v1, v2, op='eq'):
+    import operator
+
+    from packaging import version
+    op_func = getattr(operator, op)
+    return op_func(version.parse(v1), version.parse(v2))
+
+class YuanImageMLP(nn.Module):
+    
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        output_size: int,
+        hidden_act: str,
+    ) -> None:
+        super().__init__()
+        self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.down_proj = nn.Linear(intermediate_size, output_size, bias=False)
+
+        if hidden_act != "silu":
+            raise ValueError(f"Unsupported activation: {hidden_act}. Only silu is supported for now.")
+        
+        self.act_fn = ACT2FN[hidden_act]
+    
+    @torch.compile
+    def swiglu(self, y_1, y_2):
+        return self.act_fn(y_1) * y_2
+    
+    def forward(self, x):
+        x1 = self.up_proj(x)
+        x2 = self.gate_proj(x)
+        x3 = self.swiglu(x1, x2)
+        x = self.down_proj(x3)
+        return x
+
+class YuanVLChatModel(PreTrainedModel):
+    config_class = YuanVLChatConfig
+    main_input_name = 'pixel_values'
+    base_model_prefix = 'language_model'
+    _supports_flash_attn_2 = True
+    _no_split_modules = ['InternVisionModel', 'YuanDeocderLayer']
+
+    def __init__(self, config: YuanVLChatConfig, vision_model=None, language_model=None, use_flash_attn=True):
+        super().__init__(config)
+
+        assert version_cmp(transformers.__version__, '4.37.0', 'ge')
+        image_size = config.force_image_size or config.vision_config.image_size
+        patch_size = config.vision_config.patch_size
+        self.patch_size = patch_size
+        self.select_layer = config.select_layer
+        self.template = config.template
+        self.num_image_token = int((image_size // patch_size) ** 2 * (config.downsample_ratio ** 2))
+        self.downsample_ratio = config.downsample_ratio
+        self.ps_version = config.ps_version
+        use_flash_attn = use_flash_attn if has_flash_attn else False
+        config.vision_config.use_flash_attn = True if use_flash_attn else False
+        config.llm_config._attn_implementation = 'flash_attention_2' if use_flash_attn else 'eager'
+
+        logger.info(f'num_image_token: {self.num_image_token}')
+        logger.info(f'ps_version: {self.ps_version}')
+        if vision_model is not None:
+            self.vision_model = vision_model
+        else:
+            self.vision_model = InternVisionModel(config.vision_config)
+        if language_model is not None:
+            self.language_model = language_model
+        else:
+            if config.llm_config.architectures[0] == 'YuanForCausalLM':
+                self.language_model = YuanForCausalLM(config.llm_config)
+            else:
+                raise NotImplementedError(f'{config.llm_config.architectures[0]} is not implemented.')
+
+        self.pixel_unshuffle = torch.nn.PixelUnshuffle(downscale_factor=2)
+        layernorm_epsilon = config.llm_config.rms_norm_eps
+
+        self.imagemlp_input_hiddensize = int(config.vision_config.hidden_size / self.downsample_ratio ** 2)
+        self.imagemlp_ffn_hidden_size = config.llm_config.ffn_hidden_size
+
+        self.imagemlp = YuanImageMLP(self.imagemlp_input_hiddensize, self.imagemlp_ffn_hidden_size,
+                     output_size=config.llm_config.hidden_size, hidden_act="silu")
+        self.imagemlp_layernorm = RMSNorm(config.llm_config.hidden_size, eps=layernorm_epsilon)
+
+        self.img_context_token_id = config.img_context_token_id
+        self.conv_template = get_conv_template(self.template)
+        self.system_message = self.conv_template.system_message
+
+    def _validate_pixel_values(self,
+                                data: Union[torch.Tensor, List[torch.Tensor]]
+                                ) -> Union[torch.Tensor, List[torch.Tensor]]:
+        
+        h = w = self.config.vision_config.image_size
+        expected_dims = (3, h, w)
+
+        def _validate_shape(d: torch.Tensor):
+            actual_dims = tuple(d.shape)
+            if actual_dims != expected_dims:
+                # expected_expr = ("num_patches", *map(str, expected_dims))
+                expected_expr = (expected_dims)
+                raise ValueError("The expected shape of pixel values in each batch element "
+                                 f" is {expected_expr}. You supplied {tuple(d.shape)}.")
+        for d in data:
+            _validate_shape(d)
+        return data
+
+
+
+    def _parse_and_validate_image_input(self, 
+                                        pixel_values: List[torch.Tensor] = None, 
+                                        image_token_id: torch.Tensor = None, 
+                                        image_embeds: torch.Tensor = None,
+                                        ) -> Optional[InternVLImagePixelInputs]:
+        if pixel_values is None and image_embeds is None:
+            return None
+        
+        if image_embeds is not None:
+            if not isinstance(image_embeds, torch.Tensor):
+                raise ValueError("Incorrect type of image embeddings. "
+                                 f"Got type: {type(image_embeds)}")
+            return InternVLImageEmbeddingInputs(
+                type="image_embeds",
+                data=flatten_bn(image_embeds),
+            )
+        
+        if pixel_values is not None:
+            if not isinstance(pixel_values, (torch.Tensor, list)):
+                raise ValueError("Incorrect type of pixel values. "
+                                 f"Got type: {type(pixel_values)}")
+            patches_per_image = []
+            for request_pixel_values in pixel_values:
+                patches_per_image.append(request_pixel_values.shape[0])
+
+            # We need to flatten (B, N, P) to (B*N*P)
+            # so we call flatten_bn twice.
+            # (total_patches, 3, h, w)
+            return InternVLImagePixelInputs(
+                type="pixel_values",
+                data=self._validate_pixel_values(flatten_bn(pixel_values)),
+                patches_per_image=patches_per_image)
+        raise AssertionError("This line should be unreachable")
+    
+    def _process_image_input(
+            self,
+            image_input: InternVLImageInputs,
+    ) -> Tuple[torch.Tensor] :
+        if image_input["type"] == "image_embeds":
+            return image_input["data"]
+        assert self.vision_model is not None
+        # (total_patches, tokens_per_image, llm_config.hidden_size)
+        image_embeds = self.extract_feature(image_input["data"])
+
+        patches_per_image = image_input["patches_per_image"]
+
+        # Only one image in the current batch
+        if len(patches_per_image) == 1:
+            image_embeds = image_embeds.view(-1, self.config.llm_config.hidden_size).unsqueeze(1)
+            return image_embeds
+        # NOTE: Image embeddings are split into separate tensors for each image
+        # by the size of each embedding.
+        # feature_size 每个patch 256个token位置
+        feature_size = image_embeds.shape[1]
+        # (total_image_tokens, llm_config.hidden_size)
+        image_embeds = image_embeds.view(-1, self.config.llm_config.hidden_size)
+        image_feature_sizes = [num_patches * feature_size for num_patches in patches_per_image]
+        image_embeds = image_embeds.split(image_feature_sizes)
+
+        return image_embeds
+        
+
+    
+    def get_multimodal_embeddings(self,
+                                  pixel_values: Optional[List[torch.Tensor]] = None,
+                                  image_token_id: Optional[List[torch.Tensor]] = None,
+                                  image_embeds: Optional[List[torch.Tensor]] = None,
+                                  image_input: InternVLImageInputs = None,
+                                  ):
+        image_input = self._parse_and_validate_image_input(pixel_values, image_token_id, image_embeds)
+        if image_input is None:
+            return None
+        
+        # image_input: (total_patches, 3, h, w)
+        vision_embeddings = self._process_image_input(image_input)
+        return vision_embeddings
+    
+    def get_input_embeddings(
+            self,
+            input_ids: torch.Tensor,
+            multimodal_embeddings: Optional[torch.Tensor]
+            ) -> torch.Tensor:
+        inputs_embeds = self.language_model.model.get_input_embeddings(input_ids)
+        if multimodal_embeddings is not None:
+            assert self.img_context_token_id is not None
+            inputs_embeds = merge_multimodal_embeddings(
+                input_ids, inputs_embeds, multimodal_embeddings,
+                self.img_context_token_id)
+        return inputs_embeds
+    
+    def forward(
+            self,
+            input_ids: torch.LongTensor = None,
+            attention_mask: torch.Tensor = None,
+            position_ids: torch.LongTensor = None,
+            past_key_values: 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,
+            pixel_values: Optional[List[torch.Tensor]] = None,
+            image_token_id: Optional[List[torch.Tensor]] = None,
+            image_embeds: Optional[List[torch.Tensor]] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+
+        if inputs_embeds is None:
+            # (images, patches * token_per_image)
+            vision_embeddings = self.get_multimodal_embeddings(pixel_values, image_token_id, image_embeds)
+            # (tokens, hidden_size)
+            inputs_embeds = self.get_input_embeddings(input_ids, vision_embeddings).permute(1, 0, 2)
+            input_ids = None
+        
+        hidden_states = self.language_model.model(input_ids, attention_mask, position_ids, past_key_values, 
+                                                  inputs_embeds, labels, use_cache, output_attentions, 
+                                                  output_hidden_states, return_dict)
+
+        return hidden_states
+
+    def pixel_shuffle(self, x, scale_factor=0.5):
+        n, w, h, c = x.size()
+        # N, W, H, C --> N, W, H * scale, C // scale
+        x = x.view(n, w, int(h * scale_factor), int(c / scale_factor))
+        # N, W, H * scale, C // scale --> N, H * scale, W, C // scale
+        x = x.permute(0, 2, 1, 3).contiguous()
+        # N, H * scale, W, C // scale --> N, H * scale, W * scale, C // (scale ** 2)
+        x = x.view(n, int(h * scale_factor), int(w * scale_factor),
+                   int(c / (scale_factor * scale_factor)))
+        if self.ps_version == 'v1':
+            warnings.warn("In ps_version 'v1', the height and width have not been swapped back, "
+                          'which results in a transposed image.')
+        else:
+            x = x.permute(0, 2, 1, 3).contiguous()
+        return x
+
+    # Internvl vision
+    def extract_feature(self, pixel_values):
+        # pixel_values: (imbs * num_image, ic, ih, iw)
+        pixel_values = pixel_values.to(torch.bfloat16)
+        output = self.vision_model(pixel_values=pixel_values)
+        vit_embeds=output[0]
+        # vit_embeds: (imbs * num_images, h*w, vit_dim)
+        vit_embeds = vit_embeds[:, 1:, :]
+
+        pn, phw, pc = vit_embeds.shape
+        ph = pw = int(phw**0.5)
+        vit_embeds = vit_embeds.view(pn, ph, pw, pc).permute(0, 3, 1, 2)
+        vit_embeds = self.pixel_unshuffle(vit_embeds)
+        pn, pc, ph, pw = vit_embeds.shape
+        vit_embeds = vit_embeds.view(pn, pc, ph * pw).permute(0, 2, 1)
+        num_images, cvs, chs = vit_embeds.shape
+        vit_embeds = vit_embeds.reshape(1, -1, vit_embeds.shape[-1]).permute(1, 0, 2)
+        vit_embeds = self.imagemlp(vit_embeds)
+        vit_embeds = self.imagemlp_layernorm(vit_embeds)
+        vit_embeds = vit_embeds.view(num_images, cvs, -1)
+        return vit_embeds
+    
+    @torch.no_grad()
+    def generate(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        input_ids: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+        visual_features: Optional[torch.FloatTensor] = None,
+        generation_config: Optional[GenerationConfig] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        output_hidden_states: Optional[bool] = None,
+    ) -> torch.LongTensor:
+
+        
+        if pixel_values is not None:
+            if visual_features is not None:
+                vit_embeds = visual_features
+            else:
+                vit_embeds = self.get_multimodal_embeddings(pixel_values)
+                inputs_embeds = self.get_input_embeddings(input_ids, vit_embeds)
+                input_ids = None
+        
+        
+        outputs = self.language_model.generate(
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            generation_config=generation_config,
+            output_hidden_states=output_hidden_states,
+            position_ids=position_ids,
+            max_length=8192,
+            use_cache=True,
+        )
+
+
+        return outputs

special_tokens_map.json

@@ -0,0 +1,1092 @@
+{
+  "additional_special_tokens": [
+    "<s>",
+    "<eod>",
+    "<unk>",
+    "<sep>",
+    "<pad>",
+    "<mask>",
+    "<predict>",
+    "<FIM_SUFFIX>",
+    "<FIM_PREFIX>",
+    "<FIM_MIDDLE>",
+    "<commit_before>",
+    "<commit_msg>",
+    "<commit_after>",
+    "<jupyter_start>",
+    "<jupyter_text>",
+    "<jupyter_code>",
+    "<jupyter_output>",
+    "<empty_output>",
+    "<repo_name>",
+    "<file_sep>",
+    "<BOS>",
+    "<IMAGE>",
+    "</IMAGE>",
+    "<grounding>",
+    "<obj>",
+    "</obj>",
+    "<box>",
+    "</box>",
+    "<point>",
+    "</point>",
+    "<3dbox>",
+    "</3dbox>",
+    "<depth>",
+    "</depth>",
+    "s000",
+    "s001",
+    "s002",
+    "s003",
+    "s004",
+    "s005",
+    "s006",
+    "s007",
+    "s008",
+    "s009",
+    "s010",
+    "s011",
+    "s012",
+    "s013",
+    "s014",
+    "s015",
+    "s016",
+    "s017",
+    "s018",
+    "s019",
+    "s020",
+    "s021",
+    "s022",
+    "s023",
+    "s024",
+    "s025",
+    "s026",
+    "s027",
+    "s028",
+    "s029",
+    "s030",
+    "s031",
+    "s032",
+    "s033",
+    "s034",
+    "s035",
+    "s036",
+    "s037",
+    "s038",
+    "s039",
+    "s040",
+    "s041",
+    "s042",
+    "s043",
+    "s044",
+    "s045",
+    "s046",
+    "s047",
+    "s048",
+    "s049",
+    "s050",
+    "s051",
+    "s052",
+    "s053",
+    "s054",
+    "s055",
+    "s056",
+    "s057",
+    "s058",
+    "s059",
+    "s060",
+    "s061",
+    "s062",
+    "s063",
+    "s064",
+    "s065",
+    "s066",
+    "s067",
+    "s068",
+    "s069",
+    "s070",
+    "s071",
+    "s072",
+    "s073",
+    "s074",
+    "s075",
+    "s076",
+    "s077",
+    "s078",
+    "s079",
+    "s080",
+    "s081",
+    "s082",
+    "s083",
+    "s084",
+    "s085",
+    "s086",
+    "s087",
+    "s088",
+    "s089",
+    "s090",
+    "s091",
+    "s092",
+    "s093",
+    "s094",
+    "s095",
+    "s096",
+    "s097",
+    "s098",
+    "s099",
+    "s100",
+    "s101",
+    "s102",
+    "s103",
+    "s104",
+    "s105",
+    "s106",
+    "s107",
+    "s108",
+    "s109",
+    "s110",
+    "s111",
+    "s112",
+    "s113",
+    "s114",
+    "s115",
+    "s116",
+    "s117",
+    "s118",
+    "s119",
+    "s120",
+    "s121",
+    "s122",
+    "s123",
+    "s124",
+    "s125",
+    "s126",
+    "s127",
+    "s128",
+    "s129",
+    "s130",
+    "s131",
+    "s132",
+    "s133",
+    "s134",
+    "s135",
+    "s136",
+    "s137",
+    "s138",
+    "s139",
+    "s140",
+    "s141",
+    "s142",
+    "s143",
+    "s144",
+    "s145",
+    "s146",
+    "s147",
+    "s148",
+    "s149",
+    "s150",
+    "s151",
+    "s152",
+    "s153",
+    "s154",
+    "s155",
+    "s156",
+    "s157",
+    "s158",
+    "s159",
+    "s160",
+    "s161",
+    "s162",
+    "s163",
+    "s164",
+    "s165",
+    "s166",
+    "s167",
+    "s168",
+    "s169",
+    "s170",
+    "s171",
+    "s172",
+    "s173",
+    "s174",
+    "s175",
+    "s176",
+    "s177",
+    "s178",
+    "s179",
+    "s180",
+    "s181",
+    "s182",
+    "s183",
+    "s184",
+    "s185",
+    "s186",
+    "s187",
+    "s188",
+    "s189",
+    "s190",
+    "s191",
+    "s192",
+    "s193",
+    "s194",
+    "s195",
+    "s196",
+    "s197",
+    "s198",
+    "s199",
+    "s200",
+    "s201",
+    "s202",
+    "s203",
+    "s204",
+    "s205",
+    "s206",
+    "s207",
+    "s208",
+    "s209",
+    "s210",
+    "s211",
+    "s212",
+    "s213",
+    "s214",
+    "s215",
+    "s216",
+    "s217",
+    "s218",
+    "s219",
+    "s220",
+    "s221",
+    "s222",
+    "s223",
+    "s224",
+    "s225",
+    "s226",
+    "s227",
+    "s228",
+    "s229",
+    "s230",
+    "s231",
+    "s232",
+    "s233",
+    "s234",
+    "s235",
+    "s236",
+    "s237",
+    "s238",
+    "s239",
+    "s240",
+    "s241",
+    "s242",
+    "s243",
+    "s244",
+    "s245",
+    "s246",
+    "s247",
+    "s248",
+    "s249",
+    "s250",
+    "s251",
+    "s252",
+    "s253",
+    "s254",
+    "s255",
+    "s256",
+    "s257",
+    "s258",
+    "s259",
+    "s260",
+    "s261",
+    "s262",
+    "s263",
+    "s264",
+    "s265",
+    "s266",
+    "s267",
+    "s268",
+    "s269",
+    "s270",
+    "s271",
+    "s272",
+    "s273",
+    "s274",
+    "s275",
+    "s276",
+    "s277",
+    "s278",
+    "s279",
+    "s280",
+    "s281",
+    "s282",
+    "s283",
+    "s284",
+    "s285",
+    "s286",
+    "s287",
+    "s288",
+    "s289",
+    "s290",
+    "s291",
+    "s292",
+    "s293",
+    "s294",
+    "s295",
+    "s296",
+    "s297",
+    "s298",
+    "s299",
+    "s300",
+    "s301",
+    "s302",
+    "s303",
+    "s304",
+    "s305",
+    "s306",
+    "s307",
+    "s308",
+    "s309",
+    "s310",
+    "s311",
+    "s312",
+    "s313",
+    "s314",
+    "s315",
+    "s316",
+    "s317",
+    "s318",
+    "s319",
+    "s320",
+    "s321",
+    "s322",
+    "s323",
+    "s324",
+    "s325",
+    "s326",
+    "s327",
+    "s328",
+    "s329",
+    "s330",
+    "s331",
+    "s332",
+    "s333",
+    "s334",
+    "s335",
+    "s336",
+    "s337",
+    "s338",
+    "s339",
+    "s340",
+    "s341",
+    "s342",
+    "s343",
+    "s344",
+    "s345",
+    "s346",
+    "s347",
+    "s348",
+    "s349",
+    "s350",
+    "s351",
+    "s352",
+    "s353",
+    "s354",
+    "s355",
+    "s356",
+    "s357",
+    "s358",
+    "s359",
+    "s360",
+    "s361",
+    "s362",
+    "s363",
+    "s364",
+    "s365",
+    "s366",
+    "s367",
+    "s368",
+    "s369",
+    "s370",
+    "s371",
+    "s372",
+    "s373",
+    "s374",
+    "s375",
+    "s376",
+    "s377",
+    "s378",
+    "s379",
+    "s380",
+    "s381",
+    "s382",
+    "s383",
+    "s384",
+    "s385",
+    "s386",
+    "s387",
+    "s388",
+    "s389",
+    "s390",
+    "s391",
+    "s392",
+    "s393",
+    "s394",
+    "s395",
+    "s396",
+    "s397",
+    "s398",
+    "s399",
+    "s400",
+    "s401",
+    "s402",
+    "s403",
+    "s404",
+    "s405",
+    "s406",
+    "s407",
+    "s408",
+    "s409",
+    "s410",
+    "s411",
+    "s412",
+    "s413",
+    "s414",
+    "s415",
+    "s416",
+    "s417",
+    "s418",
+    "s419",
+    "s420",
+    "s421",
+    "s422",
+    "s423",
+    "s424",
+    "s425",
+    "s426",
+    "s427",
+    "s428",
+    "s429",
+    "s430",
+    "s431",
+    "s432",
+    "s433",
+    "s434",
+    "s435",
+    "s436",
+    "s437",
+    "s438",
+    "s439",
+    "s440",
+    "s441",
+    "s442",
+    "s443",
+    "s444",
+    "s445",
+    "s446",
+    "s447",
+    "s448",
+    "s449",
+    "s450",
+    "s451",
+    "s452",
+    "s453",
+    "s454",
+    "s455",
+    "s456",
+    "s457",
+    "s458",
+    "s459",
+    "s460",
+    "s461",
+    "s462",
+    "s463",
+    "s464",
+    "s465",
+    "s466",
+    "s467",
+    "s468",
+    "s469",
+    "s470",
+    "s471",
+    "s472",
+    "s473",
+    "s474",
+    "s475",
+    "s476",
+    "s477",
+    "s478",
+    "s479",
+    "s480",
+    "s481",
+    "s482",
+    "s483",
+    "s484",
+    "s485",
+    "s486",
+    "s487",
+    "s488",
+    "s489",
+    "s490",
+    "s491",
+    "s492",
+    "s493",
+    "s494",
+    "s495",
+    "s496",
+    "s497",
+    "s498",
+    "s499",
+    "s500",
+    "s501",
+    "s502",
+    "s503",
+    "s504",
+    "s505",
+    "s506",
+    "s507",
+    "s508",
+    "s509",
+    "s510",
+    "s511",
+    "s512",
+    "s513",
+    "s514",
+    "s515",
+    "s516",
+    "s517",
+    "s518",
+    "s519",
+    "s520",
+    "s521",
+    "s522",
+    "s523",
+    "s524",
+    "s525",
+    "s526",
+    "s527",
+    "s528",
+    "s529",
+    "s530",
+    "s531",
+    "s532",
+    "s533",
+    "s534",
+    "s535",
+    "s536",
+    "s537",
+    "s538",
+    "s539",
+    "s540",
+    "s541",
+    "s542",
+    "s543",
+    "s544",
+    "s545",
+    "s546",
+    "s547",
+    "s548",
+    "s549",
+    "s550",
+    "s551",
+    "s552",
+    "s553",
+    "s554",
+    "s555",
+    "s556",
+    "s557",
+    "s558",
+    "s559",
+    "s560",
+    "s561",
+    "s562",
+    "s563",
+    "s564",
+    "s565",
+    "s566",
+    "s567",
+    "s568",
+    "s569",
+    "s570",
+    "s571",
+    "s572",
+    "s573",
+    "s574",
+    "s575",
+    "s576",
+    "s577",
+    "s578",
+    "s579",
+    "s580",
+    "s581",
+    "s582",
+    "s583",
+    "s584",
+    "s585",
+    "s586",
+    "s587",
+    "s588",
+    "s589",
+    "s590",
+    "s591",
+    "s592",
+    "s593",
+    "s594",
+    "s595",
+    "s596",
+    "s597",
+    "s598",
+    "s599",
+    "s600",
+    "s601",
+    "s602",
+    "s603",
+    "s604",
+    "s605",
+    "s606",
+    "s607",
+    "s608",
+    "s609",
+    "s610",
+    "s611",
+    "s612",
+    "s613",
+    "s614",
+    "s615",
+    "s616",
+    "s617",
+    "s618",
+    "s619",
+    "s620",
+    "s621",
+    "s622",
+    "s623",
+    "s624",
+    "s625",
+    "s626",
+    "s627",
+    "s628",
+    "s629",
+    "s630",
+    "s631",
+    "s632",
+    "s633",
+    "s634",
+    "s635",
+    "s636",
+    "s637",
+    "s638",
+    "s639",
+    "s640",
+    "s641",
+    "s642",
+    "s643",
+    "s644",
+    "s645",
+    "s646",
+    "s647",
+    "s648",
+    "s649",
+    "s650",
+    "s651",
+    "s652",
+    "s653",
+    "s654",
+    "s655",
+    "s656",
+    "s657",
+    "s658",
+    "s659",
+    "s660",
+    "s661",
+    "s662",
+    "s663",
+    "s664",
+    "s665",
+    "s666",
+    "s667",
+    "s668",
+    "s669",
+    "s670",
+    "s671",
+    "s672",
+    "s673",
+    "s674",
+    "s675",
+    "s676",
+    "s677",
+    "s678",
+    "s679",
+    "s680",
+    "s681",
+    "s682",
+    "s683",
+    "s684",
+    "s685",
+    "s686",
+    "s687",
+    "s688",
+    "s689",
+    "s690",
+    "s691",
+    "s692",
+    "s693",
+    "s694",
+    "s695",
+    "s696",
+    "s697",
+    "s698",
+    "s699",
+    "s700",
+    "s701",
+    "s702",
+    "s703",
+    "s704",
+    "s705",
+    "s706",
+    "s707",
+    "s708",
+    "s709",
+    "s710",
+    "s711",
+    "s712",
+    "s713",
+    "s714",
+    "s715",
+    "s716",
+    "s717",
+    "s718",
+    "s719",
+    "s720",
+    "s721",
+    "s722",
+    "s723",
+    "s724",
+    "s725",
+    "s726",
+    "s727",
+    "s728",
+    "s729",
+    "s730",
+    "s731",
+    "s732",
+    "s733",
+    "s734",
+    "s735",
+    "s736",
+    "s737",
+    "s738",
+    "s739",
+    "s740",
+    "s741",
+    "s742",
+    "s743",
+    "s744",
+    "s745",
+    "s746",
+    "s747",
+    "s748",
+    "s749",
+    "s750",
+    "s751",
+    "s752",
+    "s753",
+    "s754",
+    "s755",
+    "s756",
+    "s757",
+    "s758",
+    "s759",
+    "s760",
+    "s761",
+    "s762",
+    "s763",
+    "s764",
+    "s765",
+    "s766",
+    "s767",
+    "s768",
+    "s769",
+    "s770",
+    "s771",
+    "s772",
+    "s773",
+    "s774",
+    "s775",
+    "s776",
+    "s777",
+    "s778",
+    "s779",
+    "s780",
+    "s781",
+    "s782",
+    "s783",
+    "s784",
+    "s785",
+    "s786",
+    "s787",
+    "s788",
+    "s789",
+    "s790",
+    "s791",
+    "s792",
+    "s793",
+    "s794",
+    "s795",
+    "s796",
+    "s797",
+    "s798",
+    "s799",
+    "s800",
+    "s801",
+    "s802",
+    "s803",
+    "s804",
+    "s805",
+    "s806",
+    "s807",
+    "s808",
+    "s809",
+    "s810",
+    "s811",
+    "s812",
+    "s813",
+    "s814",
+    "s815",
+    "s816",
+    "s817",
+    "s818",
+    "s819",
+    "s820",
+    "s821",
+    "s822",
+    "s823",
+    "s824",
+    "s825",
+    "s826",
+    "s827",
+    "s828",
+    "s829",
+    "s830",
+    "s831",
+    "s832",
+    "s833",
+    "s834",
+    "s835",
+    "s836",
+    "s837",
+    "s838",
+    "s839",
+    "s840",
+    "s841",
+    "s842",
+    "s843",
+    "s844",
+    "s845",
+    "s846",
+    "s847",
+    "s848",
+    "s849",
+    "s850",
+    "s851",
+    "s852",
+    "s853",
+    "s854",
+    "s855",
+    "s856",
+    "s857",
+    "s858",
+    "s859",
+    "s860",
+    "s861",
+    "s862",
+    "s863",
+    "s864",
+    "s865",
+    "s866",
+    "s867",
+    "s868",
+    "s869",
+    "s870",
+    "s871",
+    "s872",
+    "s873",
+    "s874",
+    "s875",
+    "s876",
+    "s877",
+    "s878",
+    "s879",
+    "s880",
+    "s881",
+    "s882",
+    "s883",
+    "s884",
+    "s885",
+    "s886",
+    "s887",
+    "s888",
+    "s889",
+    "s890",
+    "s891",
+    "s892",
+    "s893",
+    "s894",
+    "s895",
+    "s896",
+    "s897",
+    "s898",
+    "s899",
+    "s900",
+    "s901",
+    "s902",
+    "s903",
+    "s904",
+    "s905",
+    "s906",
+    "s907",
+    "s908",
+    "s909",
+    "s910",
+    "s911",
+    "s912",
+    "s913",
+    "s914",
+    "s915",
+    "s916",
+    "s917",
+    "s918",
+    "s919",
+    "s920",
+    "s921",
+    "s922",
+    "s923",
+    "s924",
+    "s925",
+    "s926",
+    "s927",
+    "s928",
+    "s929",
+    "s930",
+    "s931",
+    "s932",
+    "s933",
+    "s934",
+    "s935",
+    "s936",
+    "s937",
+    "s938",
+    "s939",
+    "s940",
+    "s941",
+    "s942",
+    "s943",
+    "s944",
+    "s945",
+    "s946",
+    "s947",
+    "s948",
+    "s949",
+    "s950",
+    "s951",
+    "s952",
+    "s953",
+    "s954",
+    "s955",
+    "s956",
+    "s957",
+    "s958",
+    "s959",
+    "s960",
+    "s961",
+    "s962",
+    "s963",
+    "s964",
+    "s965",
+    "s966",
+    "s967",
+    "s968",
+    "s969",
+    "s970",
+    "s971",
+    "s972",
+    "s973",
+    "s974",
+    "s975",
+    "s976",
+    "s977",
+    "s978",
+    "s979",
+    "s980",
+    "s981",
+    "s982",
+    "s983",
+    "s984",
+    "s985",
+    "s986",
+    "s987",
+    "s988",
+    "s989",
+    "s990",
+    "s991",
+    "s992",
+    "s993",
+    "s994",
+    "s995",
+    "s996",
+    "s997",
+    "s998",
+    "s999",
+    "<eop>",
+    "<eog>",
+    "<|begin_of_sentence|>",
+    "<|end_of_sentence|>",
+    "<|User|>",
+    "<|Assistant|>",
+    "<think>",
+    "</think>",
+    "<search_result>",
+    "</search_result>",
+    "<search_query>",
+    "</search_query>",
+    "<code_query>",
+    "</code_query>",
+    "<code_result>",
+    "</code_result>",
+    "<infer>",
+    "</infer>",
+    "<inferresult>",
+    "</inferresult>",
+    "<tool_calls>",
+    "</tool_calls>",
+    "<tool_response>",
+    "</tool_response>",
+    "<final_answer>",
+    "</final_answer>"
+  ],
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<eod>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<eod>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c84861a800c30e71099d63dca0963edbacf554586527ac037155a0560e2fb04
+size 14976038

tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:36f79e0c70f73cdd2a8dd0fbe7bfe290da158eea746778d289e4ad76c8b383d9
+size 2155861

utils.py

@@ -0,0 +1,144 @@
+from typing import (Callable, Dict, Iterable, List, Literal, Mapping, Optional, Protocol, Set, Tuple, Union, overload)
+import torch
+from torch.func import functional_call
+
+@overload
+def flatten_bn(x: torch.Tensor) -> torch.Tensor:
+    ...
+
+
+@overload
+def flatten_bn(x: List[torch.Tensor]) -> List[torch.Tensor]:
+    ...
+
+
+@overload
+def flatten_bn(
+    x: Union[List[torch.Tensor], torch.Tensor],
+    *,
+    concat: Literal[True],
+) -> torch.Tensor:
+    ...
+
+
+@overload
+def flatten_bn(
+    x: Union[List[torch.Tensor], torch.Tensor],
+    *,
+    concat: bool = False,
+) -> Union[List[torch.Tensor], torch.Tensor]:
+    ...
+
+
+def flatten_bn(
+    x: Union[List[torch.Tensor], torch.Tensor],
+    *,
+    concat: bool = False,
+) -> Union[List[torch.Tensor], torch.Tensor]:
+    """
+    Flatten the ``B`` and ``N`` dimensions of batched multimodal inputs.
+
+    The input tensor should have shape ``(B, N, ...)```.
+    """
+    if isinstance(x, torch.Tensor):
+        return x.flatten(0, 1)
+
+    if concat:
+        return torch.cat(x)
+
+    return [x_n for x_b in x for x_n in x_b]
+
+def _flatten_embeddings(embeddings: torch.Tensor) -> torch.Tensor:
+    """
+    Recursively flattens and concatenates NestedTensors on all but the last
+    dimension.
+    """
+
+    if isinstance(embeddings, torch.Tensor):
+        # Flatten all but the last dimension.
+        return embeddings.flatten(0, -2)
+
+    return torch.cat(tuple(_flatten_embeddings(t) for t in embeddings))
+
+def _embedding_count_expression(embeddings: torch.Tensor) -> str:
+    """
+    Constructs a debugging representation of the number of embeddings in the
+    Tensors.
+    """
+
+    if isinstance(embeddings, torch.Tensor):
+        return " x ".join([str(dim) for dim in embeddings.shape[:-1]])
+
+    return " + ".join(
+        _embedding_count_expression(inner) for inner in embeddings)
+
+def _merge_multimodal_embeddings(
+    inputs_embeds: torch.Tensor,
+    is_multimodal: torch.Tensor,
+    multimodal_embeddings: torch.Tensor,
+) -> torch.Tensor:
+    """
+    Merge ``multimodal_embeddings`` into ``inputs_embeds`` by overwriting the
+    positions in ``inputs_embeds`` corresponding to placeholder tokens in
+    ``input_ids``.
+
+    Note:
+        This updates ``inputs_embeds`` in place.
+    """
+    num_expected_tokens = is_multimodal.sum().item()
+    assert isinstance(num_expected_tokens, int)
+    # [total_patches, text_config.hidden_size]
+    flattened = _flatten_embeddings(multimodal_embeddings)
+    if flattened.shape[0] != num_expected_tokens:
+        expr = _embedding_count_expression(multimodal_embeddings)
+        raise ValueError(
+            f"Attempted to assign {expr} = {flattened.shape[0]} "
+            f"multimodal tokens to {num_expected_tokens} placeholders")
+
+    inputs_embeds[is_multimodal] = flattened
+    return inputs_embeds
+
+def merge_multimodal_embeddings(
+    input_ids: torch.Tensor,
+    inputs_embeds: torch.Tensor,
+    multimodal_embeddings: torch.Tensor,
+    placeholder_token_id: Union[int, List[int]],
+) -> torch.Tensor:
+    """
+    Merge ``multimodal_embeddings`` into ``inputs_embeds`` by overwriting the
+    positions in ``inputs_embeds`` corresponding to placeholder tokens in
+    ``input_ids``.
+
+    ``placeholder_token_id`` can be a list of token ids (e.g, token ids
+    of img_start, img_break, and img_end tokens) when needed: This means
+    the order of these tokens in the ``input_ids`` MUST MATCH the order of
+    their embeddings in ``multimodal_embeddings`` since we need to
+    slice-merge instead of individually scattering.
+
+    For example, if input_ids is "TTTTTSIIIBIIIBIIIETTT", where
+    - T is text token
+    - S is image start token
+    - I is image embedding token
+    - B is image break token
+    - E is image end token.
+
+    Then the image embeddings (that correspond to I's) from vision encoder
+    must be padded with embeddings of S, B, and E in the same order of
+    input_ids for a correct embedding merge.
+
+    Note:
+        This updates ``inputs_embeds`` in place.
+    """
+    if isinstance(placeholder_token_id, list):
+        placeholder_token_id = torch.tensor(placeholder_token_id,
+                                            device=input_ids.device)
+        return _merge_multimodal_embeddings(
+            inputs_embeds,
+            torch.isin(input_ids, placeholder_token_id),
+            multimodal_embeddings,
+        )
+    return _merge_multimodal_embeddings(
+        inputs_embeds,
+        (input_ids == placeholder_token_id),
+        multimodal_embeddings,
+    )