Upload folder using huggingface_hub

config.json

@@ -0,0 +1,58 @@
+{
+  "_name_or_path": "/content/d2c_hf_bin",
+  "architectures": [
+    "CogAgentForCausalLM"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_cogagent.CogAgentConfig",
+    "AutoModelForCausalLM": "modeling_cogagent.CogAgentForCausalLM"
+  },
+  "bos_token_id": 1,
+  "cross_compute_hidden_size": 1024,
+  "cross_hidden_size": 1024,
+  "cross_image_size": 1120,
+  "eos_token_id": 2,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "initializer_range": 0.02,
+  "intermediate_size": 11008,
+  "max_position_embeddings": 2048,
+  "num_attention_heads": 32,
+  "num_hidden_layers": 32,
+  "pad_token_id": 0,
+  "quantization_config": {
+    "_load_in_4bit": false,
+    "_load_in_8bit": true,
+    "bnb_4bit_compute_dtype": "float32",
+    "bnb_4bit_quant_storage": "uint8",
+    "bnb_4bit_quant_type": "fp4",
+    "bnb_4bit_use_double_quant": false,
+    "llm_int8_enable_fp32_cpu_offload": false,
+    "llm_int8_has_fp16_weight": false,
+    "llm_int8_skip_modules": null,
+    "llm_int8_threshold": 6.0,
+    "load_in_4bit": false,
+    "load_in_8bit": true,
+    "quant_method": "bitsandbytes"
+  },
+  "rms_norm_eps": 1e-05,
+  "template_version": "chat",
+  "tie_word_embeddings": false,
+  "torch_dtype": "float16",
+  "transformers_version": "4.41.0.dev0",
+  "use_cache": true,
+  "vision_config": {
+    "dropout_prob": 0.0,
+    "hidden_act": "gelu",
+    "hidden_size": 1792,
+    "image_size": 224,
+    "in_channels": 3,
+    "intermediate_size": 15360,
+    "layer_norm_eps": 1e-06,
+    "num_heads": 16,
+    "num_hidden_layers": 63,
+    "num_positions": 257,
+    "patch_size": 14
+  },
+  "vocab_size": 32000
+}

configuration_cogagent.py

@@ -0,0 +1,51 @@
+from typing import Literal
+from transformers import PretrainedConfig
+
+
+class CogAgentConfig(PretrainedConfig):
+    _auto_class = "AutoConfig"
+
+    def __init__(
+            self,
+            vocab_size=32000,
+            hidden_size=4096,
+            cross_hidden_size=1024,
+            cross_compute_hidden_size=1024,
+            cross_image_size=1120,
+            intermediate_size=11008,
+            num_hidden_layers=32,
+            num_attention_heads=32,
+            hidden_act='silu',
+            max_position_embeddings=2048,
+            initializer_range=0.02,
+            rms_norm_eps=1e-06,
+            template_version: Literal["base", "chat"] = "chat",
+
+            pad_token_id=0,
+            bos_token_id=1,
+            eos_token_id=2,
+            tie_word_embeddings=False,
+            use_cache=True,
+            **kwargs,
+    ):
+        self.hidden_size = hidden_size
+        self.cross_hidden_size = cross_hidden_size
+        self.cross_compute_hidden_size = cross_compute_hidden_size
+        self.cross_image_size = cross_image_size
+        self.intermediate_size = intermediate_size
+        self.num_attention_heads = num_attention_heads
+        self.max_position_embeddings = max_position_embeddings
+        self.rms_norm_eps = rms_norm_eps
+        self.initializer_range = initializer_range
+        self.vocab_size = vocab_size
+        self.num_hidden_layers = num_hidden_layers
+        self.hidden_act = hidden_act
+        self.template_version = template_version
+        self.use_cache = use_cache
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

cross_visual.py

@@ -0,0 +1,797 @@
+from math import pi
+import torch
+from torch import nn
+from einops import rearrange, repeat
+import logging
+
+def broadcat(tensors, dim = -1):
+    num_tensors = len(tensors)
+    shape_lens = set(list(map(lambda t: len(t.shape), tensors)))
+    assert len(shape_lens) == 1, 'tensors must all have the same number of dimensions'
+    shape_len = list(shape_lens)[0]
+    dim = (dim + shape_len) if dim < 0 else dim
+    dims = list(zip(*map(lambda t: list(t.shape), tensors)))
+    expandable_dims = [(i, val) for i, val in enumerate(dims) if i != dim]
+    assert all([*map(lambda t: len(set(t[1])) <= 2, expandable_dims)]), 'invalid dimensions for broadcastable concatentation'
+    max_dims = list(map(lambda t: (t[0], max(t[1])), expandable_dims))
+    expanded_dims = list(map(lambda t: (t[0], (t[1],) * num_tensors), max_dims))
+    expanded_dims.insert(dim, (dim, dims[dim]))
+    expandable_shapes = list(zip(*map(lambda t: t[1], expanded_dims)))
+    tensors = list(map(lambda t: t[0].expand(*t[1]), zip(tensors, expandable_shapes)))
+    return torch.cat(tensors, dim = dim)
+
+def rotate_half(x):
+    x = rearrange(x, '... (d r) -> ... d r', r = 2)
+    x1, x2 = x.unbind(dim = -1)
+    x = torch.stack((-x2, x1), dim = -1)
+    return rearrange(x, '... d r -> ... (d r)')
+
+class VisionRotaryEmbeddingFast(nn.Module):
+    def __init__(
+        self,
+        dim,
+        pt_seq_len,
+        ft_seq_len=None,
+        custom_freqs = None,
+        freqs_for = 'lang',
+        theta = 10000,
+        max_freq = 10,
+        num_freqs = 1,
+        patch_dropout = 0.
+    ):
+        super().__init__()
+        if custom_freqs:
+            freqs = custom_freqs
+        elif freqs_for == 'lang':
+            freqs = 1. / (theta ** (torch.arange(0, dim, 2)[:(dim // 2)].float() / dim))
+        elif freqs_for == 'pixel':
+            freqs = torch.linspace(1., max_freq / 2, dim // 2) * pi
+        elif freqs_for == 'constant':
+            freqs = torch.ones(num_freqs).float()
+        else:
+            raise ValueError(f'unknown modality {freqs_for}')
+
+        if ft_seq_len is None: ft_seq_len = pt_seq_len
+        t = torch.arange(ft_seq_len) / ft_seq_len * pt_seq_len
+
+        freqs = torch.einsum('..., f -> ... f', t, freqs)
+        freqs = repeat(freqs, '... n -> ... (n r)', r = 2)
+        freqs = broadcat((freqs[:, None, :], freqs[None, :, :]), dim = -1)
+
+        freqs_cos = freqs.cos().view(-1, freqs.shape[-1])
+        freqs_sin = freqs.sin().view(-1, freqs.shape[-1])
+
+        self.patch_dropout = patch_dropout
+
+        self.register_buffer("freqs_cos", freqs_cos)
+        self.register_buffer("freqs_sin", freqs_sin)
+
+        logging.info(f'Shape of rope freq: {self.freqs_cos.shape}')
+
+    def forward(self, t, patch_indices_keep=None):
+        if patch_indices_keep is not None:
+            batch = t.size()[0]
+            batch_indices = torch.arange(batch)
+            batch_indices = batch_indices[..., None]
+
+            freqs_cos = repeat(self.freqs_cos, 'i j -> n i m j', n=t.shape[0], m=t.shape[1])
+            freqs_sin = repeat(self.freqs_sin, 'i j -> n i m j', n=t.shape[0], m=t.shape[1])
+
+            freqs_cos = freqs_cos[batch_indices, patch_indices_keep]
+            freqs_cos = rearrange(freqs_cos, 'n i m j -> n m i j')
+            freqs_sin = freqs_sin[batch_indices, patch_indices_keep]
+            freqs_sin = rearrange(freqs_sin, 'n i m j -> n m i j')
+
+            return  t * freqs_cos + rotate_half(t) * freqs_sin
+
+        return  t * self.freqs_cos + rotate_half(t) * self.freqs_sin
+
+import torch.nn as nn
+import os
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+from functools import partial
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+# --------------------------------------------------------
+# Adapted from  https://github.com/microsoft/unilm/tree/master/beit
+# --------------------------------------------------------
+import math
+import os
+from functools import partial
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import logging
+try:
+    from timm.models.layers import drop_path, to_2tuple, trunc_normal_
+except:
+    from timm.layers import drop_path, to_2tuple, trunc_normal_
+    
+class PatchDropout(nn.Module):
+    """
+    https://arxiv.org/abs/2212.00794
+    """
+
+    def __init__(self, prob, exclude_first_token=True):
+        super().__init__()
+        assert 0 <= prob < 1.
+        self.prob = prob
+        self.exclude_first_token = exclude_first_token  # exclude CLS token
+        logging.info(f"os.getenv('RoPE')={os.getenv('RoPE')}")
+
+    def forward(self, x):
+        if not self.training or self.prob == 0.:
+            return x
+
+        if self.exclude_first_token:
+            cls_tokens, x = x[:, :1], x[:, 1:]
+        else:
+            cls_tokens = torch.jit.annotate(torch.Tensor, x[:, :1])
+
+        batch = x.size()[0]
+        num_tokens = x.size()[1]
+
+        batch_indices = torch.arange(batch)
+        batch_indices = batch_indices[..., None]
+
+        keep_prob = 1 - self.prob
+        num_patches_keep = max(1, int(num_tokens * keep_prob))
+
+        rand = torch.randn(batch, num_tokens)
+        patch_indices_keep = rand.topk(num_patches_keep, dim=-1).indices
+
+        x = x[batch_indices, patch_indices_keep]
+
+        if self.exclude_first_token:
+            x = torch.cat((cls_tokens, x), dim=1)
+
+        if self.training and os.getenv('RoPE') == '1':
+            return x, patch_indices_keep
+
+        return x
+
+if os.getenv('ENV_TYPE') == 'deepspeed':
+    try:
+        from deepspeed.runtime.activation_checkpointing.checkpointing import checkpoint
+    except:
+        from torch.utils.checkpoint import checkpoint
+else:
+    from torch.utils.checkpoint import checkpoint
+
+import xformers.ops as xops
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
+    """
+    def __init__(self, drop_prob=None):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, x):
+        return drop_path(x, self.drop_prob, self.training)
+    
+    def extra_repr(self) -> str:
+        return 'p={}'.format(self.drop_prob)
+
+
+class Mlp(nn.Module):
+    def __init__(
+        self, 
+        in_features, 
+        hidden_features=None, 
+        out_features=None, 
+        act_layer=nn.GELU, 
+        norm_layer=nn.LayerNorm, 
+        drop=0.,
+        subln=False,
+
+        ):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+
+        self.ffn_ln = norm_layer(hidden_features) if subln else nn.Identity()
+
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        # x = self.drop(x)
+        # commit this for the orignal BERT implement 
+        x = self.ffn_ln(x)
+
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+class SwiGLU(nn.Module):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.SiLU, drop=0., 
+                norm_layer=nn.LayerNorm, subln=False):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+
+        self.w1 = nn.Linear(in_features, hidden_features)
+        self.w2 = nn.Linear(in_features, hidden_features)
+
+        self.act = act_layer()
+        self.ffn_ln = norm_layer(hidden_features) if subln else nn.Identity()
+        self.w3 = nn.Linear(hidden_features, out_features)
+        
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x1 = self.w1(x)
+        x2 = self.w2(x)
+        hidden = self.act(x1) * x2
+        x = self.ffn_ln(hidden)
+        x = self.w3(x)
+        x = self.drop(x)
+        return x
+
+class Attention(nn.Module):
+    def __init__(
+            self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0.,
+            proj_drop=0., window_size=None, attn_head_dim=None, xattn=False, rope=None, subln=False, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        if attn_head_dim is not None:
+            head_dim = attn_head_dim
+        all_head_dim = head_dim * self.num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.subln = subln
+        if self.subln:
+            self.q_proj = nn.Linear(dim, all_head_dim, bias=False)
+            self.k_proj = nn.Linear(dim, all_head_dim, bias=False)
+            self.v_proj = nn.Linear(dim, all_head_dim, bias=False)
+        else:
+            self.qkv = nn.Linear(dim, all_head_dim * 3, bias=False)
+
+        if qkv_bias:
+            self.q_bias = nn.Parameter(torch.zeros(all_head_dim))
+            self.v_bias = nn.Parameter(torch.zeros(all_head_dim))
+        else:
+            self.q_bias = None
+            self.v_bias = None
+
+        if window_size:
+            self.window_size = window_size
+            self.num_relative_distance = (2 * window_size[0] - 1) * (2 * window_size[1] - 1) + 3
+            self.relative_position_bias_table = nn.Parameter(
+                torch.zeros(self.num_relative_distance, num_heads))  # 2*Wh-1 * 2*Ww-1, nH
+            # cls to token & token 2 cls & cls to cls
+
+            # get pair-wise relative position index for each token inside the window
+            coords_h = torch.arange(window_size[0])
+            coords_w = torch.arange(window_size[1])
+            coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+            coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+            relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+            relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+            relative_coords[:, :, 0] += window_size[0] - 1  # shift to start from 0
+            relative_coords[:, :, 1] += window_size[1] - 1
+            relative_coords[:, :, 0] *= 2 * window_size[1] - 1
+            relative_position_index = \
+                torch.zeros(size=(window_size[0] * window_size[1] + 1, ) * 2, dtype=relative_coords.dtype)
+            relative_position_index[1:, 1:] = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+            relative_position_index[0, 0:] = self.num_relative_distance - 3
+            relative_position_index[0:, 0] = self.num_relative_distance - 2
+            relative_position_index[0, 0] = self.num_relative_distance - 1
+
+            self.register_buffer("relative_position_index", relative_position_index)
+        else:
+            self.window_size = None
+            self.relative_position_bias_table = None
+            self.relative_position_index = None
+
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.inner_attn_ln = norm_layer(all_head_dim) if subln else nn.Identity()
+        # self.proj = nn.Linear(all_head_dim, all_head_dim)
+        self.proj = nn.Linear(all_head_dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+        self.xattn = xattn
+        self.xattn_drop = attn_drop
+
+        self.rope = rope
+
+    def forward(self, x, rel_pos_bias=None, attn_mask=None):
+        B, N, C = x.shape
+        if self.subln: 
+            if self.q_proj.weight.dtype == torch.uint8:
+                import bitsandbytes as bnb
+                q = bnb.matmul_4bit(x, self.q_proj.weight.t(), bias=self.q_bias, quant_state=self.q_proj.weight.quant_state)
+                k = bnb.matmul_4bit(x, self.k_proj.weight.t(), bias=None, quant_state=self.k_proj.weight.quant_state)
+                v = bnb.matmul_4bit(x, self.v_proj.weight.t(), bias=self.v_bias, quant_state=self.v_proj.weight.quant_state)
+            else:
+                q = F.linear(input=x, weight=self.q_proj.weight, bias=self.q_bias)
+                k = F.linear(input=x, weight=self.k_proj.weight, bias=None)
+                v = F.linear(input=x, weight=self.v_proj.weight, bias=self.v_bias)
+
+            q = q.reshape(B, N, self.num_heads, -1).permute(0, 2, 1, 3)     # B, num_heads, N, C
+            k = k.reshape(B, N, self.num_heads, -1).permute(0, 2, 1, 3)  
+            v = v.reshape(B, N, self.num_heads, -1).permute(0, 2, 1, 3) 
+        else: 
+
+            qkv_bias = None
+            if self.q_bias is not None:
+                qkv_bias = torch.cat((self.q_bias, torch.zeros_like(self.v_bias, requires_grad=False), self.v_bias))
+            
+            qkv = F.linear(input=x, weight=self.qkv.weight, bias=qkv_bias)
+            qkv = qkv.reshape(B, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)   # 3, B, num_heads, N, C
+            q, k, v = qkv[0], qkv[1], qkv[2]
+
+        if self.rope:
+            # slightly fast impl
+            q_t = q[:, :, 1:, :]
+            ro_q_t = self.rope(q_t)
+            q = torch.cat((q[:, :, :1, :], ro_q_t), -2).type_as(v)
+
+            k_t = k[:, :, 1:, :]
+            ro_k_t = self.rope(k_t)
+            k = torch.cat((k[:, :, :1, :], ro_k_t), -2).type_as(v)
+
+        if self.xattn:
+            q = q.permute(0, 2, 1, 3)   # B, num_heads, N, C -> B, N, num_heads, C
+            k = k.permute(0, 2, 1, 3)
+            v = v.permute(0, 2, 1, 3)
+
+            x = xops.memory_efficient_attention(
+                q, k, v,
+                p=self.xattn_drop,
+                scale=self.scale,
+                )
+            x = x.reshape(B, N, -1)
+            x = self.inner_attn_ln(x)
+            x = self.proj(x)
+            x = self.proj_drop(x)
+        else:
+            q = q * self.scale
+            attn = (q @ k.transpose(-2, -1))
+
+            if self.relative_position_bias_table is not None:
+                relative_position_bias = \
+                    self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
+                        self.window_size[0] * self.window_size[1] + 1,
+                        self.window_size[0] * self.window_size[1] + 1, -1)  # Wh*Ww,Wh*Ww,nH
+                relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+                attn = attn + relative_position_bias.unsqueeze(0).type_as(attn)
+
+            if rel_pos_bias is not None:
+                attn = attn + rel_pos_bias.type_as(attn)
+
+            if attn_mask is not None:
+                attn_mask = attn_mask.bool()
+                attn = attn.masked_fill(~attn_mask[:, None, None, :], float("-inf"))
+            
+            attn = attn.softmax(dim=-1)
+            attn = self.attn_drop(attn)
+
+            x = (attn @ v).transpose(1, 2).reshape(B, N, -1)
+            x = self.inner_attn_ln(x)
+            x = self.proj(x)
+            x = self.proj_drop(x)
+        return x
+
+
+class Block(nn.Module):
+
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., init_values=None, act_layer=nn.GELU, norm_layer=nn.LayerNorm,
+                 window_size=None, attn_head_dim=None, xattn=False, rope=None, postnorm=False,
+                 subln=False, naiveswiglu=False):
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
+            attn_drop=attn_drop, proj_drop=drop, window_size=window_size, attn_head_dim=attn_head_dim,
+            xattn=xattn, rope=rope, subln=subln, norm_layer=norm_layer)
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+
+        if naiveswiglu:
+            self.mlp = SwiGLU(
+                in_features=dim, 
+                hidden_features=mlp_hidden_dim, 
+                subln=subln,
+                norm_layer=norm_layer,
+            )
+        else:
+            self.mlp = Mlp(
+                in_features=dim, 
+                hidden_features=mlp_hidden_dim, 
+                act_layer=act_layer,
+                subln=subln,
+                drop=drop
+            )
+
+        if init_values is not None and init_values > 0:
+            self.gamma_1 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)
+            self.gamma_2 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)
+        else:
+            self.gamma_1, self.gamma_2 = None, None
+
+        self.postnorm = postnorm
+
+    def forward(self, x, rel_pos_bias=None, attn_mask=None):
+        if self.gamma_1 is None:
+            if self.postnorm:
+                x = x + self.drop_path(self.norm1(self.attn(x, rel_pos_bias=rel_pos_bias, attn_mask=attn_mask)))
+                x = x + self.drop_path(self.norm2(self.mlp(x)))
+            else:
+                x = x + self.drop_path(self.attn(self.norm1(x), rel_pos_bias=rel_pos_bias, attn_mask=attn_mask))
+                x = x + self.drop_path(self.mlp(self.norm2(x)))
+        else:
+            if self.postnorm:
+                x = x + self.drop_path(self.gamma_1 * self.norm1(self.attn(x, rel_pos_bias=rel_pos_bias, attn_mask=attn_mask)))
+                x = x + self.drop_path(self.gamma_2 * self.norm2(self.mlp(x)))
+            else:
+                x = x + self.drop_path(self.gamma_1 * self.attn(self.norm1(x), rel_pos_bias=rel_pos_bias, attn_mask=attn_mask))
+                x = x + self.drop_path(self.gamma_2 * self.mlp(self.norm2(x)))
+        return x
+
+
+class PatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+    """
+    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])
+        self.patch_shape = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.num_patches = num_patches
+
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
+
+    def forward(self, x, **kwargs):
+        B, C, H, W = x.shape
+        # FIXME look at relaxing size constraints
+        assert H == self.img_size[0] and W == self.img_size[1], \
+            f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
+        x = self.proj(x).flatten(2).transpose(1, 2)
+        return x
+
+
+class RelativePositionBias(nn.Module):
+
+    def __init__(self, window_size, num_heads):
+        super().__init__()
+        self.window_size = window_size
+        self.num_relative_distance = (2 * window_size[0] - 1) * (2 * window_size[1] - 1) + 3
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros(self.num_relative_distance, num_heads))  # 2*Wh-1 * 2*Ww-1, nH
+        # cls to token & token 2 cls & cls to cls
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(window_size[0])
+        coords_w = torch.arange(window_size[1])
+        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * window_size[1] - 1
+        relative_position_index = \
+            torch.zeros(size=(window_size[0] * window_size[1] + 1,) * 2, dtype=relative_coords.dtype)
+        relative_position_index[1:, 1:] = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        relative_position_index[0, 0:] = self.num_relative_distance - 3
+        relative_position_index[0:, 0] = self.num_relative_distance - 2
+        relative_position_index[0, 0] = self.num_relative_distance - 1
+
+        self.register_buffer("relative_position_index", relative_position_index)
+
+    def forward(self):
+        relative_position_bias = \
+            self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
+                self.window_size[0] * self.window_size[1] + 1,
+                self.window_size[0] * self.window_size[1] + 1, -1)  # Wh*Ww,Wh*Ww,nH
+        return relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+
+
+class EVAVisionTransformer(nn.Module):
+    """ Vision Transformer with support for patch or hybrid CNN input stage
+    """
+    def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dim=768, depth=12,
+                 num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0.,
+                 drop_path_rate=0., norm_layer=nn.LayerNorm, init_values=None, patch_dropout=0.,
+                 use_abs_pos_emb=True, use_rel_pos_bias=False, use_shared_rel_pos_bias=False, rope=False,
+                 use_mean_pooling=True, init_scale=0.001, grad_checkpointing=False, xattn=False, postnorm=False,
+                 pt_hw_seq_len=16, intp_freq=False, naiveswiglu=False, subln=False):
+        super().__init__()
+        self.image_size = img_size
+        self.num_classes = num_classes
+        self.num_features = self.embed_dim = embed_dim  # num_features for consistency with other models
+
+        self.patch_embed = PatchEmbed(
+            img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim)
+        num_patches = self.patch_embed.num_patches
+
+        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
+        # self.mask_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
+        if use_abs_pos_emb:
+            self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))
+        else:
+            self.pos_embed = None
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        if use_shared_rel_pos_bias:
+            self.rel_pos_bias = RelativePositionBias(window_size=self.patch_embed.patch_shape, num_heads=num_heads)
+        else:
+            self.rel_pos_bias = None
+
+        if rope:
+            half_head_dim = embed_dim // num_heads // 2
+            hw_seq_len = img_size // patch_size
+            self.rope = VisionRotaryEmbeddingFast(
+                dim=half_head_dim,
+                pt_seq_len=pt_hw_seq_len,
+                ft_seq_len=hw_seq_len if intp_freq else None,
+                # patch_dropout=patch_dropout
+            )
+        else: 
+            self.rope = None
+
+        self.naiveswiglu = naiveswiglu
+
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
+        self.use_rel_pos_bias = use_rel_pos_bias
+        self.blocks = nn.ModuleList([
+            Block(
+                dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
+                drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer,
+                init_values=init_values, window_size=self.patch_embed.patch_shape if use_rel_pos_bias else None,
+                xattn=xattn, rope=self.rope, postnorm=postnorm, subln=subln, naiveswiglu=naiveswiglu)
+            for i in range(depth)])
+        self.norm = nn.Identity() if use_mean_pooling else norm_layer(embed_dim)
+        self.fc_norm = norm_layer(embed_dim) if use_mean_pooling else None
+        self.head = nn.Linear(embed_dim, num_classes) if num_classes > 0 else nn.Identity()
+
+        if self.pos_embed is not None:
+            trunc_normal_(self.pos_embed, std=.02)
+
+        trunc_normal_(self.cls_token, std=.02)
+        # trunc_normal_(self.mask_token, std=.02)
+
+        self.apply(self._init_weights)
+        self.fix_init_weight()
+
+        if isinstance(self.head, nn.Linear):
+            trunc_normal_(self.head.weight, std=.02)
+            self.head.weight.data.mul_(init_scale)
+            self.head.bias.data.mul_(init_scale)
+
+        # setting a patch_dropout of 0. would mean it is disabled and this function would be the identity fn
+        self.patch_dropout = PatchDropout(patch_dropout) if patch_dropout > 0. else nn.Identity()
+
+        self.grad_checkpointing = grad_checkpointing
+
+    def fix_init_weight(self):
+        def rescale(param, layer_id):
+            param.div_(math.sqrt(2.0 * layer_id))
+
+        for layer_id, layer in enumerate(self.blocks):
+            rescale(layer.attn.proj.weight.data, layer_id + 1)
+            if self.naiveswiglu:
+                rescale(layer.mlp.w3.weight.data, layer_id + 1)
+            else:
+                rescale(layer.mlp.fc2.weight.data, layer_id + 1)
+
+    def get_cast_dtype(self) -> torch.dtype:
+        return self.blocks[0].mlp.fc2.weight.dtype
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    def get_num_layers(self):
+        return len(self.blocks)
+    
+    def lock(self, unlocked_groups=0, freeze_bn_stats=False):
+        assert unlocked_groups == 0, 'partial locking not currently supported for this model'
+        for param in self.parameters():
+            param.requires_grad = False
+
+    @torch.jit.ignore
+    def set_grad_checkpointing(self, enable=True):
+        self.grad_checkpointing = enable
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'pos_embed', 'cls_token'}
+
+    def get_classifier(self):
+        return self.head
+
+    def reset_classifier(self, num_classes, global_pool=''):
+        self.num_classes = num_classes
+        self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
+
+    def forward_features(self, x, return_all_features=False):
+        
+        x = self.patch_embed(x)
+        batch_size, seq_len, _ = x.size()
+
+        cls_tokens = self.cls_token.expand(batch_size, -1, -1)  # stole cls_tokens impl from Phil Wang, thanks
+        x = torch.cat((cls_tokens, x), dim=1)
+        if self.pos_embed is not None:
+            x = x + self.pos_embed
+        x = self.pos_drop(x)
+
+        # a patch_dropout of 0. would mean it is disabled and this function would do nothing but return what was passed in
+        if os.getenv('RoPE') == '1':
+            if self.training and not isinstance(self.patch_dropout, nn.Identity):
+                x, patch_indices_keep = self.patch_dropout(x)
+                self.rope.forward = partial(self.rope.forward, patch_indices_keep=patch_indices_keep)
+            else:
+                self.rope.forward = partial(self.rope.forward, patch_indices_keep=None)
+                x = self.patch_dropout(x)
+        else:
+            x = self.patch_dropout(x)
+
+        rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None
+        for i, blk in enumerate(self.blocks):
+            if i == len(self.blocks)-1:
+                continue
+            if self.grad_checkpointing:
+                x = checkpoint(blk, x, (rel_pos_bias,))
+            else:
+                x = blk(x, rel_pos_bias=rel_pos_bias)
+
+        if not return_all_features:
+            x = self.norm(x)
+            if self.fc_norm is not None:
+                return self.fc_norm(x.mean(1))
+            else:
+                return x[:, 0]
+        return x
+
+    def forward(self, x, return_all_features=False):
+        if return_all_features:
+            return self.forward_features(x, return_all_features)
+        x = self.forward_features(x)
+        x = self.head(x)
+        return x
+
+class LayerNorm(nn.LayerNorm):
+    """Subclass torch's LayerNorm (with cast back to input dtype)."""
+
+    def forward(self, x: torch.Tensor):
+        orig_type = x.dtype
+        x = F.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps)
+        return x.to(orig_type)
+
+try:
+    from apex.normalization import FusedLayerNorm
+except:
+    FusedLayerNorm = LayerNorm
+    print("Please 'pip install apex'")
+
+
+@dataclass
+class CLIPVisionCfg:
+    layers: Union[Tuple[int, int, int, int], int] = 12
+    width: int = 768
+    head_width: int = 64
+    mlp_ratio: float = 4.0
+    patch_size: int = 16
+    image_size: Union[Tuple[int, int], int] = 224
+    ls_init_value: Optional[float] = None  # layer scale initial value
+    patch_dropout: float = 0. # what fraction of patches to dropout during training (0 would mean disabled and no patches dropped) - 0.5 to 0.75 recommended in the paper for optimal results
+    global_average_pool: bool = False # whether to global average pool the last embedding layer, instead of using CLS token (https://arxiv.org/abs/2205.01580)
+    drop_path_rate: Optional[float] = None  # drop path rate
+    timm_model_name: str = None  # a valid model name overrides layers, width, patch_size
+    timm_model_pretrained: bool = False  # use (imagenet) pretrained weights for named model
+    timm_pool: str = 'avg'  # feature pooling for timm model ('abs_attn', 'rot_attn', 'avg', '')
+    timm_proj: str = 'linear'  # linear projection for timm model output ('linear', 'mlp', '')
+    timm_proj_bias: bool = False  # enable bias final projection
+    eva_model_name: str = None # a valid eva model name overrides layers, width, patch_size
+    qkv_bias: bool = True
+    fusedLN: bool = False
+    xattn: bool = False
+    postnorm: bool = False
+    rope: bool = False
+    pt_hw_seq_len: int = 16   # 224/14
+    intp_freq: bool = False
+    naiveswiglu: bool = False
+    subln: bool = False
+
+
+def _build_vision_tower(
+        embed_dim: int,
+        vision_cfg: CLIPVisionCfg
+):
+    if isinstance(vision_cfg, dict):
+        vision_cfg = CLIPVisionCfg(**vision_cfg)
+
+    if vision_cfg.eva_model_name:
+        vision_heads = vision_cfg.width // vision_cfg.head_width
+        norm_layer = LayerNorm
+        visual = EVAVisionTransformer(
+            img_size=vision_cfg.image_size,
+            patch_size=vision_cfg.patch_size,
+            num_classes=embed_dim,
+            use_mean_pooling=vision_cfg.global_average_pool, #False
+            init_values=vision_cfg.ls_init_value,
+            patch_dropout=vision_cfg.patch_dropout,
+            embed_dim=vision_cfg.width,
+            depth=vision_cfg.layers,
+            num_heads=vision_heads,
+            mlp_ratio=vision_cfg.mlp_ratio,
+            qkv_bias=vision_cfg.qkv_bias,
+            drop_path_rate=vision_cfg.drop_path_rate,
+            norm_layer= partial(FusedLayerNorm, eps=1e-6) if vision_cfg.fusedLN else partial(norm_layer, eps=1e-6),
+            xattn=vision_cfg.xattn,
+            rope=vision_cfg.rope,
+            postnorm=vision_cfg.postnorm,
+            pt_hw_seq_len= vision_cfg.pt_hw_seq_len,   # 224/14
+            intp_freq= vision_cfg.intp_freq,
+            naiveswiglu= vision_cfg.naiveswiglu,
+            subln= vision_cfg.subln
+        )
+
+    return visual
+
+class Eva2LargeEncoder(nn.Module):
+    def __init__(self, image_size=224):
+        super(Eva2LargeEncoder, self).__init__()
+        self.config = {
+            "embed_dim": 768,
+            "vision_cfg": {
+                "image_size": 336,
+                "layers": 24,
+                "width": 1024,
+                "drop_path_rate": 0,
+                "head_width": 64,
+                "mlp_ratio": 2.6667,
+                "patch_size": 14,
+                "eva_model_name": "eva-clip-l-14-336",
+                "xattn": True,
+                "fusedLN": True,
+                "rope": True,
+                "pt_hw_seq_len": 16,
+                "intp_freq": True,
+                "naiveswiglu": True,
+                "subln": True
+            }
+        }
+        self.config['vision_cfg']['image_size'] = image_size
+        
+        import os
+        os.environ['delRoPE'] = '1' # to avoid error in rope params when changing image size
+        self.model = _build_vision_tower(**self.config)
+
+
+    def forward(self, images):
+        encode = self.model(images, return_all_features=True)[:, 1:, :]
+        return encode
+
+class CrossVisionModel(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.vit = Eva2LargeEncoder(image_size=config.cross_image_size)
+        self.pos_embed = nn.Parameter(torch.zeros((self.vit.config['vision_cfg']['image_size'] // self.vit.config['vision_cfg']['patch_size']) ** 2, self.vit.config['vision_cfg']['width']))
+
+    def forward(self, images):
+        enc = self.vit(images)
+        return enc + self.pos_embed.to(enc.device).unsqueeze(0)

generation_config.json

@@ -0,0 +1,7 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "pad_token_id": 0,
+  "transformers_version": "4.41.0.dev0"
+}

modeling_cogagent.py

@@ -0,0 +1,974 @@
+"""largely copy from llama and adapt for CogAgent"""
+import warnings
+from typing import TYPE_CHECKING, Optional, Tuple, List, Union, Literal, Dict, Any
+
+import math
+import torch
+from torch import nn
+from torch.nn import CrossEntropyLoss
+from torchvision import transforms
+from einops import rearrange
+
+from transformers import PreTrainedModel, PreTrainedTokenizer
+from transformers.utils.logging import get_logger
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+
+from .configuration_cogagent import CogAgentConfig
+# from .util import FastRotaryEmbedding
+from torch.nn import functional as F
+from .visual import EVA2CLIPModel
+from .cross_visual import CrossVisionModel
+
+if TYPE_CHECKING:
+    from transformers.utils import ModelOutput
+
+logger = get_logger(__name__)
+
+LANGUAGE_TOKEN_TYPE = 0
+VISION_TOKEN_TYPE = 1
+
+
+# Copied from transformers.models.bart.modeling_bart._make_causal_mask
+def _make_causal_mask(
+        input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
+):
+    """
+    Make causal mask used for bi-directional self-attention.
+    """
+    bsz, tgt_len = input_ids_shape
+    mask = torch.full((tgt_len, tgt_len), torch.finfo(dtype).min, device=device)
+    mask_cond = torch.arange(mask.size(-1), device=device)
+    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
+    mask = mask.to(dtype)
+
+    if past_key_values_length > 0:
+        mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
+    return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
+
+
+# Copied from transformers.models.bart.modeling_bart._expand_mask
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    """
+    bsz, src_len = mask.size()
+    tgt_len = tgt_len if tgt_len is not None else src_len
+
+    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
+
+    inverted_mask = 1.0 - expanded_mask
+
+    return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
+
+
+class RMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-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)
+
+
+class MLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+
+def get_expert_mask(token_type_ids: "torch.LongTensor(B, L)") -> "[torch.BoolTensor(B, L), torch.BoolTensor(B, L)]":
+    vision_token_mask = torch.zeros_like(token_type_ids, dtype=torch.bool)
+    vision_token_mask[:, :-1] = (token_type_ids[:, :-1] == VISION_TOKEN_TYPE) & (token_type_ids[:, 1:] == VISION_TOKEN_TYPE)
+    language_token_mask = ~vision_token_mask
+    return vision_token_mask, language_token_mask
+
+
+class VisionExpertMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.language_mlp = MLP(config)
+        self.vision_mlp = MLP(config)
+
+    def forward(self, hidden_states: "torch.Tensor(B, L, D)", token_type_ids: "torch.LongTensor(B, L)"):
+        output = torch.empty(hidden_states.shape, dtype=hidden_states.dtype, device=hidden_states.device)
+        vision_token_mask, language_token_mask = get_expert_mask(token_type_ids)
+        output[vision_token_mask] = self.vision_mlp(hidden_states[vision_token_mask])
+        output[language_token_mask] = self.language_mlp(hidden_states[language_token_mask])
+        return output
+
+
+def attention_fn(
+        query_layer: "torch.tensor(B, H, L, HD)",
+        key_layer: "torch.tensor(B, H, L, HD)",
+        value_layer: "torch.tensor(B, H, L, HD)",
+        attention_mask: "torch.tensor(B, H, L, HD)",
+        *,
+        scaling_attention_score: bool = True,
+        attention_dropout: nn.Module = None
+):
+    attention_mask_bool = (attention_mask == 0)
+    is_low_triangle = (attention_mask_bool == torch.ones_like(attention_mask_bool, dtype=torch.float).tril()).all()
+    is_full = (attention_mask_bool > 0).all()
+    if not (int(torch.__version__.split('.')[0]) >= 2):
+        warnings.warn("It's recommended to use torch2.0 or higher.")
+    if int(torch.__version__.split('.')[0]) >= 2 and scaling_attention_score and (is_full or is_low_triangle):
+        dropout_p = 0. if attention_dropout is None or not attention_dropout.training else attention_dropout.p
+        return torch.nn.functional.scaled_dot_product_attention(
+            query_layer, key_layer, value_layer,
+            attn_mask=None,
+            dropout_p=dropout_p,
+            is_causal=not is_full
+        )
+    else:
+        if scaling_attention_score:
+            query_layer = query_layer / math.sqrt(query_layer.shape[-1])
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+        attention_scores = attention_scores + attention_mask
+        attention_scores = nn.functional.softmax(attention_scores, dim=-1, dtype=torch.float32).to(query_layer.dtype)
+        if attention_dropout is not None:
+            attention_scores = attention_dropout(attention_scores)
+        context_layer = torch.matmul(attention_scores, value_layer)
+        return context_layer
+
+class RotaryEmbedding(torch.nn.Module):
+    def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
+        super().__init__()
+
+        self.dim = dim
+        self.max_position_embeddings = max_position_embeddings
+        self.base = base
+        inv_freq = self._compute_inv_freq(device)
+        self.register_buffer("inv_freq", inv_freq)
+        self.max_seq_len_cached = 0
+
+    def _compute_inv_freq(self, device=None):
+        return 1.0 / (
+                self.base
+                ** (torch.arange(0, self.dim, 2, device=device) / self.dim)
+        )
+
+    def _set_cos_sin_cache(self, seq_len, device, dtype):
+        self.max_seq_len_cached = seq_len
+        t = torch.arange(self.max_seq_len_cached, device=device, dtype=self.inv_freq.dtype)
+
+        freqs = torch.einsum("i,j->ij", t, self.inv_freq)
+        # Different from paper, but it uses a different permutation in order to obtain the same calculation
+        emb = torch.cat((freqs, freqs), dim=-1)
+        self.register_buffer("cos_cached", emb.cos()[:, None, :].to(dtype), persistent=False)
+        self.register_buffer("sin_cached", emb.sin()[:, None, :].to(dtype), persistent=False)
+
+    def forward(self, x, seq_len):
+        # x: [bs, num_attention_heads, seq_len, head_size]
+        if seq_len > self.max_seq_len_cached:
+            self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)
+
+        return (
+            self.cos_cached[:seq_len, ...].to(dtype=x.dtype),
+            self.sin_cached[:seq_len, ...].to(dtype=x.dtype),
+        )
+
+
+def rotate_half(x):
+    x1, x2 = x[..., :x.shape[-1] // 2], x[..., x.shape[-1] // 2:]
+    return torch.cat((-x2, x1), dim=x1.ndim - 1)
+
+
+def apply_rotary_pos_emb_index_bhs(q, k, cos, sin, position_id):
+    # batch_size, num_head, seq_len, hidden_size
+    cos, sin = F.embedding(position_id, cos.squeeze(1)).unsqueeze(1), \
+        F.embedding(position_id, sin.squeeze(1)).unsqueeze(1)
+    q, k = (q * cos) + (rotate_half(q) * sin), (k * cos) + (rotate_half(k) * sin)
+    return q, k
+
+class VisionExpertAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.max_position_embeddings = config.max_position_embeddings
+
+        self.rotary_emb = RotaryEmbedding(self.head_dim)
+        self.vision_expert_query_key_value = nn.Linear(self.hidden_size, self.hidden_size * 3, bias=False)
+        self.vision_expert_dense = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
+        self.language_expert_query_key_value = nn.Linear(self.hidden_size, self.hidden_size * 3, bias=False)
+        self.language_expert_dense = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
+
+    def _transpose_for_scores(self, tensor):
+        """Transpose a 3D tensor [B, L, H*HD] into a 4D tensor with size [B H L HD]."""
+        new_tensor_shape = tensor.size()[:-1] + (self.num_heads, self.head_dim)
+        tensor = tensor.view(*new_tensor_shape)
+        return tensor.permute(0, 2, 1, 3)
+
+    def forward(
+            self,
+            hidden_states: torch.Tensor,
+            token_type_ids: torch.LongTensor,
+            position_ids: torch.LongTensor,
+            attention_mask: Optional[torch.Tensor] = None,
+            past_key_value: Optional[Tuple[torch.Tensor]] = None,
+            output_attentions: bool = False,
+            use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+        vision_token_mask, language_token_mask = get_expert_mask(token_type_ids)
+
+        shape = list(hidden_states.shape)
+        shape[-1] = shape[-1] * 3
+        mixed_raw_layer = torch.empty(shape, dtype=hidden_states.dtype, device=hidden_states.device)
+        mixed_raw_layer[vision_token_mask] = self.vision_expert_query_key_value(hidden_states[vision_token_mask])
+        mixed_raw_layer[language_token_mask] = self.language_expert_query_key_value(hidden_states[language_token_mask])
+
+        query_states, key_states, value_states = torch.split(mixed_raw_layer, self.hidden_size, dim=-1)
+        query_states = self._transpose_for_scores(query_states)  # B, H, L, HD
+        key_states = self._transpose_for_scores(key_states)  # B, H, L, HD
+        value_states = self._transpose_for_scores(value_states)  # B, H, L, HD
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value[0].shape[-2]
+
+        cos, sin = self.rotary_emb(value_states, seq_len=position_ids.max() + 1)
+        query_states, key_states = apply_rotary_pos_emb_index_bhs(query_states, key_states, cos, sin, position_ids)
+
+        if past_key_value is not None:
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+
+        past_key_value = (key_states, value_states) if use_cache else None
+
+        context_layer = attention_fn(
+            query_layer=query_states, key_layer=key_states, value_layer=value_states, attention_mask=attention_mask,
+            scaling_attention_score=True, attention_dropout=None)
+        if context_layer.size() != (bsz, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {context_layer.size()}"
+            )
+        context_layer = context_layer.transpose(1, 2).contiguous().reshape(bsz, q_len, self.hidden_size)
+
+        attn_output = torch.empty(context_layer.shape, dtype=hidden_states.dtype, device=hidden_states.device)
+        attn_output[vision_token_mask] = self.vision_expert_dense(context_layer[vision_token_mask])
+        attn_output[language_token_mask] = self.language_expert_dense(context_layer[language_token_mask])
+
+        if output_attentions:
+            warnings.warn("output_attentions is not implemented.")
+
+        return attn_output, None, past_key_value
+
+class CrossAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.cross_hidden_size = config.cross_hidden_size
+        self.cross_compute_hidden_size = config.cross_compute_hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.cross_head_dim = self.cross_compute_hidden_size // self.num_heads
+        self.max_position_embeddings = config.max_position_embeddings
+
+        self.query = nn.Linear(self.hidden_size, self.cross_compute_hidden_size, bias=False)
+        self.key_value = nn.Linear(self.cross_hidden_size, self.cross_compute_hidden_size * 2, bias=False)
+        self.dense = nn.Linear(self.cross_compute_hidden_size, self.hidden_size, bias=False)
+
+    def _transpose_for_scores(self, tensor):
+        """Transpose a 3D tensor [B, L, H*HD] into a 4D tensor with size [B H L HD]."""
+        new_tensor_shape = tensor.size()[:-1] + (self.num_heads, self.cross_head_dim)
+        tensor = tensor.view(*new_tensor_shape)
+        return tensor.permute(0, 2, 1, 3)
+
+    def forward(
+            self,
+            hidden_states: torch.Tensor,
+            encoder_outputs: torch.LongTensor,
+            attention_mask: Optional[torch.Tensor] = None,
+            past_key_value: Optional[Tuple[torch.Tensor]] = None,
+            output_attentions: bool = False,
+            use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        shape = list(hidden_states.shape)
+        shape[-1] = shape[-1] * 3
+
+        mixed_query_layer = self.query(hidden_states)
+        if past_key_value is None:
+            mixed_x_layer = self.key_value(encoder_outputs)
+            mixed_key_layer, mixed_value_layer = torch.split(mixed_x_layer, self.cross_compute_hidden_size, dim=-1)
+            key_states = self._transpose_for_scores(mixed_key_layer)  # B, H, L, HD
+            value_states = self._transpose_for_scores(mixed_value_layer)  # B, H, L, HD
+        else:
+            key_states, value_states = past_key_value
+
+        query_states = self._transpose_for_scores(mixed_query_layer)  # B, H, L, HD
+
+        past_key_value = (key_states, value_states) if use_cache else None
+
+        context_layer = attention_fn(
+            query_layer=query_states, key_layer=key_states, value_layer=value_states, attention_mask=attention_mask,
+            scaling_attention_score=True, attention_dropout=None)
+        if context_layer.size() != (bsz, self.num_heads, q_len, self.cross_head_dim):
+            raise ValueError(
+                f"`cross_attn_output` should be of size {(bsz, self.num_heads, q_len, self.cross_head_dim)}, but is"
+                f" {context_layer.size()}"
+            )
+        context_layer = context_layer.transpose(1, 2).contiguous().reshape(bsz, q_len, self.cross_hidden_size)
+
+        attn_output = self.dense(context_layer)
+
+        if output_attentions:
+            warnings.warn("output_attentions is not implemented.")
+
+        return attn_output, None, past_key_value
+
+class CogAgentDecoderLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = VisionExpertAttention(config=config)
+        self.cross_attn = CrossAttention(config=config)
+        self.mlp = VisionExpertMLP(config)
+        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_cross_attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+            self,
+            hidden_states: torch.Tensor,
+            encoder_outputs: torch.Tensor,
+            token_type_ids: torch.LongTensor,
+            position_ids: torch.LongTensor,
+            attention_mask: Optional[torch.Tensor] = None,
+            cross_attention_mask: Optional[torch.Tensor] = None,
+            past_key_value: Optional[Tuple[torch.Tensor]] = None,
+            output_attentions: Optional[bool] = False,
+            use_cache: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            attention_mask=attention_mask,
+            past_key_value=past_key_value[:2] if past_key_value is not None else None,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+        )
+        hidden_states = residual + hidden_states
+
+        cross_input = self.post_cross_attention_layernorm(hidden_states)
+        # Fully Connected
+        attention_output, self_cross_attn_weights, present_cross_key_value = self.cross_attn(
+            hidden_states=cross_input,
+            encoder_outputs=encoder_outputs,
+            attention_mask=cross_attention_mask,
+            past_key_value=past_key_value[-2:] if past_key_value is not None else None,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            )
+        hidden_states = hidden_states + attention_output
+        mlp_input = self.post_attention_layernorm(hidden_states)
+        mlp_output = self.mlp(mlp_input, token_type_ids=token_type_ids)
+        hidden_states = mlp_output + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value+present_cross_key_value,)
+
+        return outputs  # type: ignore
+
+
+class CogAgentPreTrainedModel(PreTrainedModel):
+    config_class = CogAgentConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = False
+    _no_split_modules = ["CogAgentDecoderLayer", 'TransformerLayer', 'Block']
+    _skip_keys_device_placement = "past_key_values"
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+
+def is_empty(images_list: Optional[List[List[torch.Tensor]]]):
+    if images_list is None or len(images_list) == 0:
+        return True
+    for image_list in images_list:
+        if len(image_list):
+            return False
+    return True
+
+
+def build_position_ids(x: "torch.BoolTensor(B, L)", attention_mask: Optional["torch.BoolTensor(B, L)"] = None) -> "torch.LongTensor(B, L)":
+    if attention_mask is not None:
+        tmp = x.clone()
+        tmp[~(attention_mask.bool())] = -1
+    else:
+        tmp = x.clone()
+    # image boi eoi token as LANGUAGE_TOKEN_TYPE
+    is_boi_eoi = torch.zeros_like(x, dtype=torch.bool)
+    is_boi_eoi[:, 1:] |= (tmp[:, 1:] == VISION_TOKEN_TYPE) & (tmp[:, :-1] == LANGUAGE_TOKEN_TYPE)
+    is_boi_eoi[:, 0] |= (tmp[:, 0] == VISION_TOKEN_TYPE)
+    is_boi_eoi[:, :-1] |= (tmp[:, :-1] == VISION_TOKEN_TYPE) & (tmp[:, 1:] == LANGUAGE_TOKEN_TYPE)
+    is_boi_eoi[:, -1] |= (tmp[:, -1] == VISION_TOKEN_TYPE)
+    tmp[is_boi_eoi] = LANGUAGE_TOKEN_TYPE
+    # final position ids
+    y = torch.zeros_like(x, dtype=torch.long)
+    y[:, 1:] = (tmp[:, 1:] == LANGUAGE_TOKEN_TYPE) | ((tmp[:, 1:] == VISION_TOKEN_TYPE) & (tmp[:, :-1] == LANGUAGE_TOKEN_TYPE))
+    y = y.cumsum(dim=-1)
+    return y
+
+
+class CogAgentModel(CogAgentPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList([CogAgentDecoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        self.vision = EVA2CLIPModel(config)
+        self.cross_vision = CrossVisionModel(config)
+
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def encode_images(self, images: List[List[torch.Tensor]]) -> torch.Tensor:
+        images_list, images = images, []
+
+        images = []
+        for image_list in images_list:
+            for image in image_list:
+                images.append(image)
+
+        images = torch.stack(images)
+        images_features = self.vision(images)
+        return images_features
+
+    def encode_cross_images(self, images: List[List[torch.Tensor]]) -> torch.Tensor:
+        images_list, images = images, []
+
+        images = []
+        for image_list in images_list:
+            for image in image_list:
+                images.append(image)
+
+        images = torch.stack(images)
+        encoder_outputs = self.cross_vision(images)
+        return encoder_outputs
+    
+    def forward(
+            self,
+            input_ids: torch.LongTensor = None,
+            images: List[List[torch.Tensor]] = None,
+            cross_images: List[List[torch.Tensor]] = None,
+            token_type_ids: Optional[torch.LongTensor] = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            cross_attention_mask: Optional[torch.Tensor] = None,
+            position_ids: Optional[torch.LongTensor] = None,
+            past_key_values: Optional[List[torch.FloatTensor]] = None,
+            inputs_embeds: Optional[torch.FloatTensor] = None,
+            use_cache: Optional[bool] = None,
+            output_attentions: Optional[bool] = None,
+            output_hidden_states: Optional[bool] = None,
+            return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        """take care of image_encode, token_type_ids, position_ids and (attention_mask = None is fine)"""
+
+        if past_key_values is not None:
+            encoder_outputs = None
+            # generate mode with past_key_values. the image features are already mapped
+        else:
+            # not allow for inputs_embeds, because we want to process image feature
+            assert input_ids is not None and inputs_embeds is None, f"{input_ids} {inputs_embeds}"
+            if not is_empty(images):  # multi-modality
+                assert token_type_ids is not None, f"multi-modality requires `token_type_ids`!"
+                assert len(input_ids) == len(images), f"{len(input_ids)} {len(images)}"
+                inputs_embeds = self.embed_tokens(input_ids)
+                images_features = self.encode_images(images)
+                encoder_outputs = self.encode_cross_images(cross_images)
+                images_features = rearrange(images_features, 'b n d -> (b n) d')
+                images_features = images_features.to(dtype=inputs_embeds.dtype, device=inputs_embeds.device)
+                inputs_embeds = inputs_embeds.index_put([token_type_ids == VISION_TOKEN_TYPE], images_features)
+            else:  # single-modality
+                if token_type_ids is None:
+                    token_type_ids = torch.ones_like(input_ids, dtype=torch.long, device=input_ids.device) * LANGUAGE_TOKEN_TYPE
+                assert not (token_type_ids == VISION_TOKEN_TYPE).any(), f"{(token_type_ids == VISION_TOKEN_TYPE).sum()}"
+                inputs_embeds = self.embed_tokens(input_ids)
+                encoder_outputs = None
+
+            if position_ids is None:
+                position_ids = build_position_ids(token_type_ids, attention_mask)
+            input_ids = None
+
+        return self.llm_forward(
+            input_ids=input_ids,
+            encoder_outputs=encoder_outputs,
+            token_type_ids=token_type_ids,
+            attention_mask=attention_mask,
+            cross_attention_mask=cross_attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+    def llm_forward(
+            self,
+            input_ids: torch.LongTensor = None,
+            encoder_outputs: torch.LongTensor = None,
+            token_type_ids: torch.LongTensor = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            cross_attention_mask: Optional[torch.Tensor] = None,
+            position_ids: Optional[torch.LongTensor] = None,
+            past_key_values: Optional[List[torch.FloatTensor]] = None,
+            inputs_embeds: Optional[torch.FloatTensor] = None,
+            use_cache: Optional[bool] = None,
+            output_attentions: Optional[bool] = None,
+            output_hidden_states: Optional[bool] = None,
+            return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        """largely copy from llama forward and adapt for CogAgent with `token_type_ids`"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+        elif input_ids is not None:
+            batch_size, seq_length = input_ids.shape
+        elif inputs_embeds is not None:
+            batch_size, seq_length, _ = inputs_embeds.shape
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        seq_length_with_past = seq_length
+        past_key_values_length = 0
+
+        if past_key_values is not None:
+            past_key_values_length = past_key_values[0][0].shape[2]
+            seq_length_with_past = seq_length_with_past + past_key_values_length
+
+        if position_ids is None:
+            device = input_ids.device if input_ids is not None else inputs_embeds.device
+            position_ids = torch.arange(
+                past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
+            )
+            position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
+        else:
+            position_ids = position_ids.view(-1, seq_length).long()
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+        # embed positions
+        if attention_mask is None:
+            attention_mask = torch.ones(
+                (batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device
+            )
+        if cross_attention_mask is None:
+            cross_attention_mask = torch.ones(
+                (batch_size, 1), dtype=torch.bool, device=inputs_embeds.device
+            )
+        attention_mask = self._prepare_decoder_attention_mask(
+            attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
+        )
+
+        hidden_states = inputs_embeds
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = () if use_cache else None
+
+        for idx, decoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+            layer_outputs = decoder_layer(
+                hidden_states,
+                encoder_outputs=encoder_outputs,
+                token_type_ids=token_type_ids,
+                attention_mask=attention_mask,
+                cross_attention_mask=cross_attention_mask,
+                position_ids=position_ids,
+                past_key_value=past_key_value,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+            )
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    # noinspection PyMethodMayBeStatic
+    # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
+    def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
+        # create causal mask
+        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+        combined_attention_mask = None
+        if input_shape[-1] > 1:
+            combined_attention_mask = _make_causal_mask(
+                input_shape,
+                inputs_embeds.dtype,
+                device=inputs_embeds.device,
+                past_key_values_length=past_key_values_length,
+            )
+
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
+                inputs_embeds.device
+            )
+            combined_attention_mask = (
+                expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
+            )
+
+        return combined_attention_mask
+
+def vqa_history_to_prompt(history, query):
+    # Only support single round chat in vqa mode
+    prompt = "<EOI>Question: "
+    # for i, (old_query, response) in enumerate(history):
+    #     prompt += old_query + " Short answer: " + response + " Question: "
+    prompt += query + " Short answer:"
+    return prompt
+
+def chat_old_history_to_prompt(history, query):
+    prompt = "<EOI>Question: "
+    for i, (old_query, response) in enumerate(history):
+        prompt += old_query + " Answer: " + response + "\nQuestion: "
+    prompt += query + " Answer:"
+    return prompt
+
+def chat_history_to_prompt(history, query):
+    prompt = " [INST] "
+    for i, (old_query, response) in enumerate(history):
+        prompt += old_query + " [/INST] " + response + " [INST] "
+    prompt += query + " [/INST] "
+    return prompt
+
+
+def base_history_to_prompt(history, query):
+    prompt = query
+    return prompt
+
+
+_history_to_prompt = {
+    "base": base_history_to_prompt,
+    "chat": chat_history_to_prompt,
+    "chat_old": chat_old_history_to_prompt,
+    "vqa": vqa_history_to_prompt
+}
+
+
+class CogAgentForCausalLM(CogAgentPreTrainedModel):
+    _auto_class = "AutoModelForCausalLM"
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = CogAgentModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    def forward(
+            self,
+            input_ids: torch.LongTensor = None,
+            images: List[List[torch.Tensor]] = None,
+            cross_images: List[List[torch.Tensor]] = None,
+            token_type_ids: Optional[torch.LongTensor] = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            position_ids: Optional[torch.LongTensor] = None,
+            past_key_values: Optional[List[torch.FloatTensor]] = None,
+            inputs_embeds: Optional[torch.FloatTensor] = None,
+            use_cache: Optional[bool] = None,
+            output_attentions: Optional[bool] = None,
+            output_hidden_states: Optional[bool] = None,
+            return_dict: Optional[bool] = None,
+            labels: Optional[torch.LongTensor] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            images=images,
+            cross_images=cross_images,
+            token_type_ids=token_type_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+        logits = self.lm_head(hidden_states)
+        logits = logits.float()
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def _prepare_attention_mask_for_generation(
+            self,
+            inputs: torch.Tensor,
+            pad_token_id: Optional[int],
+            eos_token_id: Optional[Union[int, List[int]]],
+    ) -> torch.LongTensor:
+        return torch.ones(inputs.shape[:2], dtype=torch.long, device=inputs.device)  # type: ignore
+
+    def prepare_inputs_for_generation(
+            self, input_ids, token_type_ids, images=None, cross_images=None, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
+    ):
+        # build position_ids if needed
+        position_ids = kwargs.get("position_ids", None)
+        if position_ids is None:
+            position_ids = build_position_ids(token_type_ids, attention_mask)
+
+        if past_key_values:
+            input_ids = input_ids[:, -1:]
+            token_type_ids = token_type_ids[:, -1:]
+            position_ids = position_ids[:, -1:]
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "token_type_ids": token_type_ids,
+                "images": images,
+                "cross_images": cross_images,
+                "position_ids": position_ids,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+
+    def _update_model_kwargs_for_generation(
+            self,
+            outputs: "ModelOutput",
+            model_kwargs: Dict[str, Any],
+            is_encoder_decoder: bool = False,
+            standardize_cache_format: bool = False,
+    ) -> Dict[str, Any]:
+        # update past_key_values
+        model_kwargs["past_key_values"] = self._extract_past_from_model_output(
+            outputs, standardize_cache_format=standardize_cache_format
+        )
+        if getattr(outputs, "state", None) is not None:
+            model_kwargs["state"] = outputs.state
+
+        # update token_type_ids with last value
+        if "token_type_ids" in model_kwargs:
+            token_type_ids = model_kwargs["token_type_ids"]
+            new_token_type_ids = torch.ones(size=(token_type_ids.shape[0], 1), dtype=token_type_ids.dtype, device=token_type_ids.device) * LANGUAGE_TOKEN_TYPE
+            model_kwargs["token_type_ids"] = torch.cat([token_type_ids, new_token_type_ids], dim=-1)
+
+        if not is_encoder_decoder:
+            # update attention mask
+            if "attention_mask" in model_kwargs:
+                attention_mask = model_kwargs["attention_mask"]
+                model_kwargs["attention_mask"] = torch.cat(
+                    [attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1
+                )
+        else:
+            # update decoder attention mask
+            if "decoder_attention_mask" in model_kwargs:
+                decoder_attention_mask = model_kwargs["decoder_attention_mask"]
+                model_kwargs["decoder_attention_mask"] = torch.cat(
+                    [decoder_attention_mask, decoder_attention_mask.new_ones((decoder_attention_mask.shape[0], 1))],
+                    dim=-1,
+                )
+
+        return model_kwargs
+
+    def _reorder_cache(self, past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (
+                tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
+            )
+        return reordered_past
+
+    def build_conversation_input_ids(
+            self,
+            tokenizer: "PreTrainedTokenizer",
+            *,
+            query: str,
+            history: Optional[List[Tuple[str, str]]] = None,
+            images: Optional[List["PIL.Image"]] = None,
+            template_version: Optional[Literal["base", "chat", "vqa"]] = None,
+    ):
+        image_size: int = self.config.vision_config['image_size']
+        cross_image_size: int = self.config.cross_image_size
+        patch_size: int = self.config.vision_config['patch_size']
+        template_version = template_version or self.config.template_version
+        assert images is None or len(images) <= 1, f"not support multi images by now."
+        history = history or []
+        text = _history_to_prompt[template_version](history, query)
+
+        input_ids = [tokenizer.bos_token_id]
+        token_type_ids = [LANGUAGE_TOKEN_TYPE]
+        if images is not None and len(images) == 1:
+            ori = images
+            # vision
+            transform = transforms.Compose(
+                [
+                    transforms.Resize(
+                        (image_size, image_size), interpolation=transforms.InterpolationMode.BICUBIC
+                    ),
+                    transforms.ToTensor(),
+                    transforms.Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)),
+                ]
+            )
+            images = [transform(ori[0])]
+            cross_transform = transforms.Compose(
+                [
+                    transforms.Resize(
+                        (cross_image_size, cross_image_size), interpolation=transforms.InterpolationMode.BICUBIC
+                    ),
+                    transforms.ToTensor(),
+                    transforms.Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)),
+                ]
+            )
+            cross_images = [cross_transform(ori[0])]
+            # language
+            vision_token_num = (image_size // patch_size) * (image_size // patch_size) + 2
+            input_ids += [tokenizer.pad_token_id] * vision_token_num
+            token_type_ids += [VISION_TOKEN_TYPE] * vision_token_num
+        text_ids = tokenizer.encode(text, add_special_tokens=False)
+
+        input_ids += text_ids
+        token_type_ids += [LANGUAGE_TOKEN_TYPE] * len(text_ids)
+        attention_mask = [1] * len(input_ids)
+
+        return {
+            'input_ids': torch.tensor(input_ids, dtype=torch.long),
+            'token_type_ids': torch.tensor(token_type_ids, dtype=torch.long),
+            'attention_mask': torch.tensor(attention_mask, dtype=torch.long),
+            'images': images,
+            'cross_images': cross_images
+        }

pytorch_model-00001-of-00004.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8d09e6d7a4f2620d2d32c839747af292cbfbdf6cc9c35ea5a6619b30ef820856
+size 4979106642

pytorch_model-00002-of-00004.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6dbaddf8f16b064c0f6d55f97cd492fd44cbc3e20bb812d27f4e968d4010b63f
+size 4987736182

pytorch_model-00003-of-00004.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96aa4207396b0794637891ffee4d900af383c74c2e7b4f00c0115c0eb6630f24
+size 4974066815

pytorch_model-00004-of-00004.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5d8172706e1fcf3be3e3c9f96baac7f1cf42c984a8360423d58373f8d5f3afec
+size 3654473430

special_tokens_map.json

@@ -0,0 +1,30 @@
+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347
+size 499723

tokenizer_config.json

@@ -0,0 +1,43 @@
+{
+  "add_bos_token": true,
+  "add_eos_token": false,
+  "add_prefix_space": true,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<s>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "</s>",
+  "legacy": false,
+  "model_max_length": 4096,
+  "pad_token": "<unk>",
+  "padding_side": "right",
+  "sp_model_kwargs": {},
+  "spaces_between_special_tokens": false,
+  "tokenizer_class": "LlamaTokenizer",
+  "unk_token": "<unk>",
+  "use_default_system_prompt": false
+}

visual.py

@@ -0,0 +1,136 @@
+import torch
+from torch import nn
+from argparse import Namespace
+import xformers.ops as xops
+from transformers.activations import ACT2FN
+
+
+class PatchEmbedding(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.proj = nn.Conv2d(config.in_channels, config.hidden_size, kernel_size=config.patch_size, stride=config.patch_size)
+        self.cls_embedding = nn.Parameter(torch.zeros(1, config.hidden_size))
+        self.position_embedding = nn.Embedding(config.num_positions, config.hidden_size)
+
+    def forward(self, images: "tensor(B, C, H, W)") -> "tensor(B, L, D)":
+        x = self.proj(images)
+        x = x.flatten(2).transpose(1, 2)
+        cls_token = self.cls_embedding.expand(x.shape[0], -1, -1)
+        x = torch.cat((cls_token, x), dim=1)
+        x += self.position_embedding.weight.unsqueeze(0)
+        return x
+
+
+class Attention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.num_heads = config.num_heads
+        head_dim = config.hidden_size // config.num_heads
+        self.scale = head_dim ** -0.5
+        self.query_key_value = nn.Linear(config.hidden_size, config.hidden_size * 3)
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.output_dropout = torch.nn.Dropout(config.dropout_prob)
+
+    def forward(self, x: "tensor(B, L, D)") -> "tensor(B, L, D)":
+        B, L, _ = x.shape
+        qkv = self.query_key_value(x)
+        qkv = qkv.reshape(B, L, 3, self.num_heads, -1).permute(2, 0, 1, 3, 4)  # 3, B, L, H, D
+        q, k, v = qkv[0], qkv[1], qkv[2]
+
+        out = xops.memory_efficient_attention(
+            q, k, v, scale=self.scale,
+        )
+        output = self.dense(out.view(B, L, -1))
+        output = self.output_dropout(output)
+        return output
+
+    def attention(self, q, k, v):
+        attn_weights = torch.matmul(q * self.scale, k.transpose(-2, -1))
+        attn_weights = attn_weights.softmax(dim=-1)
+        output = torch.matmul(attn_weights, v)
+        return output
+
+
+class MLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.activation_fn = ACT2FN[config.hidden_act]
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.fc1(x)
+        x = self.activation_fn(x)
+        x = self.fc2(x)
+        return x
+
+
+class TransformerLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.attention = Attention(config)
+        self.mlp = MLP(config)
+        self.post_attention_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+    def forward(self, hidden_states):
+        attention_input = hidden_states
+        attention_output = self.input_layernorm(self.attention(attention_input))
+        hidden_states = attention_input + attention_output
+        mlp_input = hidden_states
+        mlp_output = self.post_attention_layernorm(self.mlp(mlp_input))
+        output = mlp_input + mlp_output
+        return output
+
+
+class Transformer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.layers = nn.ModuleList([TransformerLayer(config) for _ in range(config.num_hidden_layers)])
+
+    def forward(self, hidden_states):
+        for layer_module in self.layers:
+            hidden_states = layer_module(hidden_states)
+        return hidden_states
+
+
+class GLU(nn.Module):
+    def __init__(self, config, in_features):
+        super().__init__()
+        self.linear_proj = nn.Linear(in_features, config.hidden_size, bias=False)
+        self.norm1 = nn.LayerNorm(config.hidden_size)
+        self.act1 = nn.GELU()
+        self.act2 = nn.functional.silu
+        self.dense_h_to_4h = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
+        self.gate_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
+        self.dense_4h_to_h = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
+
+    def forward(self, x):
+        x = self.linear_proj(x)
+        x = self.act1(self.norm1(x))
+        x = self.act2(self.gate_proj(x)) * self.dense_h_to_4h(x)
+        x = self.dense_4h_to_h(x)
+        return x
+
+
+class EVA2CLIPModel(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        vision_config = Namespace(**config.vision_config)
+        self.patch_embedding = PatchEmbedding(vision_config)
+        self.transformer = Transformer(vision_config)
+        self.linear_proj = GLU(config, in_features=vision_config.hidden_size)
+        self.boi = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
+        self.eoi = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
+        self.pos_embed = nn.Parameter(torch.zeros((vision_config.image_size // vision_config.patch_size) ** 2, vision_config.hidden_size))
+
+    def forward(self, images: "tensor(B, C, H, W)") -> "tensor(B, L, D)":
+        x = self.patch_embedding(images)
+        x = self.transformer(x)
+        x = x[:, 1:]
+        x = self.linear_proj(x + self.pos_embed.to(x.device).unsqueeze(0))
+        boi = self.boi.to(x.device).expand(x.shape[0], -1, -1)
+        eoi = self.eoi.to(x.device).expand(x.shape[0], -1, -1)
+        x = torch.cat((boi, x, eoi), dim=1)
+        return x