OpenGVLab/VisualPRM-8B

1# --------------------------------------------------------2# InternVL3# Copyright (c) 2024 OpenGVLab4# Licensed under The MIT License [see LICENSE for details]5# --------------------------------------------------------6 7from typing import Optional, Tuple, Union8 9import torch10import torch.nn.functional as F11import torch.utils.checkpoint12from einops import rearrange13from timm.models.layers import DropPath14from torch import nn15from transformers.activations import ACT2FN16from transformers.modeling_outputs import (BaseModelOutput,17                                           BaseModelOutputWithPooling)18from transformers.modeling_utils import PreTrainedModel19from transformers.utils import logging20 21from .configuration_intern_vit import InternVisionConfig22 23try:24    from flash_attn.bert_padding import pad_input, unpad_input25    from flash_attn.flash_attn_interface import \26        flash_attn_varlen_qkvpacked_func27    has_flash_attn = True28except:29    print('FlashAttention2 is not installed.')30    has_flash_attn = False31 32logger = logging.get_logger(__name__)33 34 35class FlashAttention(nn.Module):36    """Implement the scaled dot product attention with softmax.37    Arguments38    ---------39        softmax_scale: The temperature to use for the softmax attention.40                      (default: 1/sqrt(d_keys) where d_keys is computed at41                      runtime)42        attention_dropout: The dropout rate to apply to the attention43                           (default: 0.0)44    """45 46    def __init__(self, softmax_scale=None, attention_dropout=0.0, device=None, dtype=None):47        super().__init__()48        self.softmax_scale = softmax_scale49        self.dropout_p = attention_dropout50 51    def forward(self, qkv, key_padding_mask=None, causal=False, cu_seqlens=None,52                max_s=None, need_weights=False):53        """Implements the multihead softmax attention.54        Arguments55        ---------56            qkv: The tensor containing the query, key, and value. (B, S, 3, H, D) if key_padding_mask is None57                if unpadded: (nnz, 3, h, d)58            key_padding_mask: a bool tensor of shape (B, S)59        """60        assert not need_weights61        assert qkv.dtype in [torch.float16, torch.bfloat16]62        assert qkv.is_cuda63 64        if cu_seqlens is None:65            batch_size = qkv.shape[0]66            seqlen = qkv.shape[1]67            if key_padding_mask is None:68                qkv = rearrange(qkv, 'b s ... -> (b s) ...')69                max_s = seqlen70                cu_seqlens = torch.arange(0, (batch_size + 1) * seqlen, step=seqlen, dtype=torch.int32,71                                          device=qkv.device)72                output = flash_attn_varlen_qkvpacked_func(73                    qkv, cu_seqlens, max_s, self.dropout_p if self.training else 0.0,74                    softmax_scale=self.softmax_scale, causal=causal75                )76                output = rearrange(output, '(b s) ... -> b s ...', b=batch_size)77            else:78                nheads = qkv.shape[-2]79                x = rearrange(qkv, 'b s three h d -> b s (three h d)')80                x_unpad, indices, cu_seqlens, max_s = unpad_input(x, key_padding_mask)81                x_unpad = rearrange(x_unpad, 'nnz (three h d) -> nnz three h d', three=3, h=nheads)82                output_unpad = flash_attn_varlen_qkvpacked_func(83                    x_unpad, cu_seqlens, max_s, self.dropout_p if self.training else 0.0,84                    softmax_scale=self.softmax_scale, causal=causal85                )86                output = rearrange(pad_input(rearrange(output_unpad, 'nnz h d -> nnz (h d)'),87                                             indices, batch_size, seqlen),88                                   'b s (h d) -> b s h d', h=nheads)89        else:90            assert max_s is not None91            output = flash_attn_varlen_qkvpacked_func(92                qkv, cu_seqlens, max_s, self.dropout_p if self.training else 0.0,93                softmax_scale=self.softmax_scale, causal=causal94            )95 96        return output, None97 98 99class InternRMSNorm(nn.Module):100    def __init__(self, hidden_size, eps=1e-6):101        super().__init__()102        self.weight = nn.Parameter(torch.ones(hidden_size))103        self.variance_epsilon = eps104 105    def forward(self, hidden_states):106        input_dtype = hidden_states.dtype107        hidden_states = hidden_states.to(torch.float32)108        variance = hidden_states.pow(2).mean(-1, keepdim=True)109        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)110        return self.weight * hidden_states.to(input_dtype)111 112 113try:114    from apex.normalization import FusedRMSNorm115 116    InternRMSNorm = FusedRMSNorm  # noqa117 118    logger.info('Discovered apex.normalization.FusedRMSNorm - will use it instead of InternRMSNorm')119except ImportError:120    # using the normal InternRMSNorm121    pass122except Exception:123    logger.warning('discovered apex but it failed to load, falling back to InternRMSNorm')124    pass125 126 127NORM2FN = {128    'rms_norm': InternRMSNorm,129    'layer_norm': nn.LayerNorm,130}131 132 133class InternVisionEmbeddings(nn.Module):134    def __init__(self, config: InternVisionConfig):135        super().__init__()136        self.config = config137        self.embed_dim = config.hidden_size138        self.image_size = config.image_size139        self.patch_size = config.patch_size140 141        self.class_embedding = nn.Parameter(142            torch.randn(1, 1, self.embed_dim),143        )144 145        self.patch_embedding = nn.Conv2d(146            in_channels=3, out_channels=self.embed_dim, kernel_size=self.patch_size, stride=self.patch_size147        )148 149        self.num_patches = (self.image_size // self.patch_size) ** 2150        self.num_positions = self.num_patches + 1151 152        self.position_embedding = nn.Parameter(torch.randn(1, self.num_positions, self.embed_dim))153 154    def _get_pos_embed(self, pos_embed, H, W):155        target_dtype = pos_embed.dtype156        pos_embed = pos_embed.float().reshape(157            1, self.image_size // self.patch_size, self.image_size // self.patch_size, -1).permute(0, 3, 1, 2)158        pos_embed = F.interpolate(pos_embed, size=(H, W), mode='bicubic', align_corners=False). \159            reshape(1, -1, H * W).permute(0, 2, 1).to(target_dtype)160        return pos_embed161 162    def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:163        target_dtype = self.patch_embedding.weight.dtype164        patch_embeds = self.patch_embedding(pixel_values)  # shape = [*, channel, width, height]165        batch_size, _, height, width = patch_embeds.shape166        patch_embeds = patch_embeds.flatten(2).transpose(1, 2)167        class_embeds = self.class_embedding.expand(batch_size, 1, -1).to(target_dtype)168        embeddings = torch.cat([class_embeds, patch_embeds], dim=1)169        position_embedding = torch.cat([170            self.position_embedding[:, :1, :],171            self._get_pos_embed(self.position_embedding[:, 1:, :], height, width)172        ], dim=1)173        embeddings = embeddings + position_embedding.to(target_dtype)174        return embeddings175 176 177class InternAttention(nn.Module):178    """Multi-headed attention from 'Attention Is All You Need' paper"""179 180    def __init__(self, config: InternVisionConfig):181        super().__init__()182        self.config = config183        self.embed_dim = config.hidden_size184        self.num_heads = config.num_attention_heads185        self.use_flash_attn = config.use_flash_attn and has_flash_attn186        if config.use_flash_attn and not has_flash_attn:187            print('Warning: Flash Attention is not available, use_flash_attn is set to False.')188        self.head_dim = self.embed_dim // self.num_heads189        if self.head_dim * self.num_heads != self.embed_dim:190            raise ValueError(191                f'embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:'192                f' {self.num_heads}).'193            )194 195        self.scale = self.head_dim ** -0.5196        self.qkv = nn.Linear(self.embed_dim, 3 * self.embed_dim, bias=config.qkv_bias)197        self.attn_drop = nn.Dropout(config.attention_dropout)198        self.proj_drop = nn.Dropout(config.dropout)199 200        self.qk_normalization = config.qk_normalization201 202        if self.qk_normalization:203            self.q_norm = InternRMSNorm(self.embed_dim, eps=config.layer_norm_eps)204            self.k_norm = InternRMSNorm(self.embed_dim, eps=config.layer_norm_eps)205 206        if self.use_flash_attn:207            self.inner_attn = FlashAttention(attention_dropout=config.attention_dropout)208        self.proj = nn.Linear(self.embed_dim, self.embed_dim)209 210    def _naive_attn(self, x):211        B, N, C = x.shape212        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)213        q, k, v = qkv.unbind(0)  # make torchscript happy (cannot use tensor as tuple)214 215        if self.qk_normalization:216            B_, H_, N_, D_ = q.shape217            q = self.q_norm(q.transpose(1, 2).flatten(-2, -1)).view(B_, N_, H_, D_).transpose(1, 2)218            k = self.k_norm(k.transpose(1, 2).flatten(-2, -1)).view(B_, N_, H_, D_).transpose(1, 2)219 220        attn = ((q * self.scale) @ k.transpose(-2, -1))221        attn = attn.softmax(dim=-1)222        attn = self.attn_drop(attn)223 224        x = (attn @ v).transpose(1, 2).reshape(B, N, C)225        x = self.proj(x)226        x = self.proj_drop(x)227        return x228 229    def _flash_attn(self, x, key_padding_mask=None, need_weights=False):230        qkv = self.qkv(x)231        qkv = rearrange(qkv, 'b s (three h d) -> b s three h d', three=3, h=self.num_heads)232 233        if self.qk_normalization:234            q, k, v = qkv.unbind(2)235            q = self.q_norm(q.flatten(-2, -1)).view(q.shape)236            k = self.k_norm(k.flatten(-2, -1)).view(k.shape)237            qkv = torch.stack([q, k, v], dim=2)238 239        context, _ = self.inner_attn(240            qkv, key_padding_mask=key_padding_mask, need_weights=need_weights, causal=False241        )242        outs = self.proj(rearrange(context, 'b s h d -> b s (h d)'))243        outs = self.proj_drop(outs)244        return outs245 246    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:247        x = self._naive_attn(hidden_states) if not self.use_flash_attn else self._flash_attn(hidden_states)248        return x249 250 251class InternMLP(nn.Module):252    def __init__(self, config: InternVisionConfig):253        super().__init__()254        self.config = config255        self.act = ACT2FN[config.hidden_act]256        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)257        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)258 259    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:260        hidden_states = self.fc1(hidden_states)261        hidden_states = self.act(hidden_states)262        hidden_states = self.fc2(hidden_states)263        return hidden_states264 265 266class InternVisionEncoderLayer(nn.Module):267    def __init__(self, config: InternVisionConfig, drop_path_rate: float):268        super().__init__()269        self.embed_dim = config.hidden_size270        self.intermediate_size = config.intermediate_size271        self.norm_type = config.norm_type272 273        self.attn = InternAttention(config)274        self.mlp = InternMLP(config)275        self.norm1 = NORM2FN[self.norm_type](self.embed_dim, eps=config.layer_norm_eps)276        self.norm2 = NORM2FN[self.norm_type](self.embed_dim, eps=config.layer_norm_eps)277 278        self.ls1 = nn.Parameter(config.initializer_factor * torch.ones(self.embed_dim))279        self.ls2 = nn.Parameter(config.initializer_factor * torch.ones(self.embed_dim))280        self.drop_path1 = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity()281        self.drop_path2 = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity()282 283    def forward(284            self,285            hidden_states: torch.Tensor,286    ) -> Tuple[torch.FloatTensor, Optional[torch.FloatTensor], Optional[Tuple[torch.FloatTensor]]]:287        """288        Args:289            hidden_states (`Tuple[torch.FloatTensor, Optional[torch.FloatTensor]]`): input to the layer of shape `(batch, seq_len, embed_dim)`290        """291        hidden_states = hidden_states + self.drop_path1(self.attn(self.norm1(hidden_states).to(hidden_states.dtype)) * self.ls1)292 293        hidden_states = hidden_states + self.drop_path2(self.mlp(self.norm2(hidden_states).to(hidden_states.dtype)) * self.ls2)294 295        return hidden_states296 297 298class InternVisionEncoder(nn.Module):299    """300    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a301    [`InternEncoderLayer`].302 303    Args:304        config (`InternConfig`):305            The corresponding vision configuration for the `InternEncoder`.306    """307 308    def __init__(self, config: InternVisionConfig):309        super().__init__()310        self.config = config311        # stochastic depth decay rule312        dpr = [x.item() for x in torch.linspace(0, config.drop_path_rate, config.num_hidden_layers)]313        self.layers = nn.ModuleList([314            InternVisionEncoderLayer(config, dpr[idx]) for idx in range(config.num_hidden_layers)])315        self.gradient_checkpointing = True316 317    def forward(318            self,319            inputs_embeds,320            output_hidden_states: Optional[bool] = None,321            return_dict: Optional[bool] = None,322    ) -> Union[Tuple, BaseModelOutput]:323        r"""324        Args:325            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):326                Embedded representation of the inputs. Should be float, not int tokens.327            output_hidden_states (`bool`, *optional*):328                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors329                for more detail.330            return_dict (`bool`, *optional*):331                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.332        """333        output_hidden_states = (334            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states335        )336        return_dict = return_dict if return_dict is not None else self.config.use_return_dict337 338        encoder_states = () if output_hidden_states else None339        hidden_states = inputs_embeds340 341        for idx, encoder_layer in enumerate(self.layers):342            if output_hidden_states:343                encoder_states = encoder_states + (hidden_states,)344            if self.gradient_checkpointing and self.training:345                layer_outputs = torch.utils.checkpoint.checkpoint(346                    encoder_layer,347                    hidden_states)348            else:349                layer_outputs = encoder_layer(350                    hidden_states,351                )352            hidden_states = layer_outputs353 354        if output_hidden_states:355            encoder_states = encoder_states + (hidden_states,)356 357        if not return_dict:358            return tuple(v for v in [hidden_states, encoder_states] if v is not None)359        return BaseModelOutput(360            last_hidden_state=hidden_states, hidden_states=encoder_states361        )362 363 364class InternVisionModel(PreTrainedModel):365    main_input_name = 'pixel_values'366    _supports_flash_attn_2 = True367    config_class = InternVisionConfig368    _no_split_modules = ['InternVisionEncoderLayer']369 370    def __init__(self, config: InternVisionConfig):371        super().__init__(config)372        self.config = config373 374        self.embeddings = InternVisionEmbeddings(config)375        self.encoder = InternVisionEncoder(config)376 377    def resize_pos_embeddings(self, old_size, new_size, patch_size):378        pos_emb = self.embeddings.position_embedding379        _, num_positions, embed_dim = pos_emb.shape380        cls_emb = pos_emb[:, :1, :]381        pos_emb = pos_emb[:, 1:, :].reshape(1, old_size // patch_size, old_size // patch_size, -1).permute(0, 3, 1, 2)382        pos_emb = F.interpolate(pos_emb.float(), size=new_size // patch_size, mode='bicubic', align_corners=False)383        pos_emb = pos_emb.to(cls_emb.dtype).reshape(1, embed_dim, -1).permute(0, 2, 1)384        pos_emb = torch.cat([cls_emb, pos_emb], dim=1)385        self.embeddings.position_embedding = nn.Parameter(pos_emb)386        self.embeddings.image_size = new_size387        logger.info('Resized position embeddings from {} to {}'.format(old_size, new_size))388 389    def get_input_embeddings(self):390        return self.embeddings391 392    def forward(393            self,394            pixel_values: Optional[torch.FloatTensor] = None,395            output_hidden_states: Optional[bool] = None,396            return_dict: Optional[bool] = None,397            pixel_embeds: Optional[torch.FloatTensor] = None,398    ) -> Union[Tuple, BaseModelOutputWithPooling]:399        output_hidden_states = (400            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states401        )402        return_dict = return_dict if return_dict is not None else self.config.use_return_dict403 404        if pixel_values is None and pixel_embeds is None:405            raise ValueError('You have to specify pixel_values or pixel_embeds')406 407        if pixel_embeds is not None:408            hidden_states = pixel_embeds409        else:410            if len(pixel_values.shape) == 4:411                hidden_states = self.embeddings(pixel_values)412            else:413                raise ValueError(f'wrong pixel_values size: {pixel_values.shape}')414        encoder_outputs = self.encoder(415            inputs_embeds=hidden_states,416            output_hidden_states=output_hidden_states,417            return_dict=return_dict,418        )419        last_hidden_state = encoder_outputs.last_hidden_state420        pooled_output = last_hidden_state[:, 0, :]421 422        if not return_dict:423            return (last_hidden_state, pooled_output) + encoder_outputs[1:]424 425        return BaseModelOutputWithPooling(426            last_hidden_state=last_hidden_state,427            pooler_output=pooled_output,428            hidden_states=encoder_outputs.hidden_states,429            attentions=encoder_outputs.attentions,430        )431