N8Programs/ProGen2-base-bf16

1# coding=utf-82# Copyright 2021 The EleutherAI and HuggingFace Teams. All rights reserved.3#4# Licensed under the Apache License, Version 2.0 (the "License");5# you may not use this file except in compliance with the License.6# You may obtain a copy of the License atí7#8#     http://www.apache.org/licenses/LICENSE-2.09#10# Unless required by applicable law or agreed to in writing, software11# distributed under the License is distributed on an "AS IS" BASIS,12# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.13# See the License for the specific language governing permissions and14# limitations under the License.15 16# Modified forward-pass implementation based on https://github.com/huggingface/transformers/blob/main/src/transformers/models/gptj/modeling_gptj.py17 18from typing import Tuple19 20import numpy as np21 22import torch23import torch.utils.checkpoint24from torch import nn25from torch.nn import CrossEntropyLoss26import torch.nn.functional as F27 28from transformers.activations import ACT2FN29from transformers.modeling_outputs import (30    BaseModelOutputWithPast,31    CausalLMOutputWithPast,32)33from transformers.modeling_utils import PreTrainedModel34from transformers.utils import logging35from .configuration_progen import ProGenConfig36 37 38logger = logging.get_logger(__name__)39 40 41def fixed_pos_embedding(x, seq_dim=1, seq_len=None):42    dim = x.shape[-1]43    if seq_len is None:44        seq_len = x.shape[seq_dim]45    inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2) / dim))46    sinusoid_inp = (47        torch.einsum("i , j -> i j", torch.arange(seq_len), inv_freq)48        .to(x.device)49        .float()50    )51    return torch.sin(sinusoid_inp), torch.cos(sinusoid_inp)52 53 54def rotate_every_two(x: torch.Tensor):55    x1 = x[:, :, :, ::2]56    x2 = x[:, :, :, 1::2]57    x = torch.stack((-x2, x1), axis=-1)58    return x.flatten(-2)59 60def apply_rotary_pos_emb(x, sincos, offset=0):61    sin, cos = map(62        lambda t: t[None, offset : x.shape[1] + offset, None, :].repeat_interleave(63            2, 364        ),65        sincos,66    )67    # einsum notation for lambda t: repeat(t[offset:x.shape[1]+offset,:], "n d -> () n () (d j)", j=2)68    return (x * cos) + (rotate_every_two(x) * sin)69 70 71class ProGenAttention(nn.Module):72    def __init__(self, config):73        super().__init__()74 75        max_positions = config.n_positions76        self.register_buffer(77            "bias",78            torch.tril(79                torch.ones((max_positions, max_positions), dtype=torch.bool)80            ).view(1, 1, max_positions, max_positions),81            persistent=False82        )83        self.register_buffer("masked_bias", torch.tensor(-1e9), persistent=False)     # approx. -inf84 85        self.attn_dropout = nn.Dropout(config.attn_pdrop)86        self.resid_dropout = nn.Dropout(config.resid_pdrop)87 88        self.embed_dim = config.embed_dim89        self.num_attention_heads = config.n_head90        self.head_dim = self.embed_dim // self.num_attention_heads91        if self.head_dim * self.num_attention_heads != self.embed_dim:92            raise ValueError(93                f"embed_dim must be divisible by num_attention_heads (got `embed_dim`: {self.embed_dim} and `num_attention_heads`: {self.num_attention_heads})."94            )95        self.scale_attn = torch.sqrt(96            torch.tensor(self.head_dim, dtype=torch.float32)97        ).to(torch.get_default_dtype())98        self.qkv_proj = nn.Linear(self.embed_dim, self.embed_dim * 3, bias=False)99 100        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False)101        self.rotary_dim = None102        if config.rotary_dim is not None:103            self.rotary_dim = config.rotary_dim104 105    def _split_heads(self, x: torch.Tensor, n_head, dim_head) -> torch.Tensor:106        x = x.reshape(x.shape[:-2] + (-1,))                             # (B, T, 8 * E // 8)107        x = x.reshape(x.shape[:-1] + (n_head, dim_head))             # (B, T, n_heads, dim_head)108        return x109 110    def _merge_heads(self, tensor, num_attention_heads, attn_head_size) -> torch.Tensor:111        """112        Merges attn_head_size dim and num_attn_heads dim into n_positions113        """114        if len(tensor.shape) == 5:115            tensor = tensor.permute(0, 1, 3, 2, 4).contiguous()116        elif len(tensor.shape) == 4:117            tensor = tensor.permute(0, 2, 1, 3).contiguous()118        else:119            raise ValueError(120                f"Input tensor rank should be one of [4, 5], but is: {len(tensor.shape)}"121            )122        new_shape = tensor.size()[:-2] + (num_attention_heads * attn_head_size,)123        return tensor.view(new_shape)124 125    def _attn(126        self,127        query,128        key,129        value,130        attention_mask=None,131        head_mask=None,132    ):133        # compute causal mask from causal mask buffer134        query_length, key_length = query.size(-2), key.size(-2)135        causal_mask = self.bias[136            :, :, key_length - query_length : key_length, :key_length137        ]138 139        # Keep the attention weights computation in fp32 to avoid overflow issues140        query = query.to(torch.float32)141        key = key.to(torch.float32)142 143        attn_weights = query @ key.transpose(-1, -2)                 # (B, n_heads, T, T)144 145        attn_weights = attn_weights / self.scale_attn146 147        # attend only to previous positions148        attn_weights = torch.where(149            causal_mask, attn_weights, self.masked_bias.to(attn_weights.dtype)150        )151 152        if attention_mask is not None:153            attn_weights = attn_weights + attention_mask154 155        attn_weights = F.softmax(attn_weights, dim=-1)156        attn_weights = attn_weights.to(value.dtype)157        attn_weights = self.attn_dropout(attn_weights)158 159        if head_mask is not None:160            attn_weights = attn_weights * head_mask161 162        attn_output = attn_weights @ value                # (B, n_heads, T, dim_head)163 164        return attn_output, attn_weights165 166    def forward(167        self,168        hidden_states,169        attention_mask=None,170        layer_past=None,171        head_mask=None,172        use_cache=False,173        output_attentions=False,174    ):175        qkv = self.qkv_proj(hidden_states)                                         # (B, T, 3 * E)176 177        mp_num = 8178        qkv_split = qkv.reshape(qkv.shape[:-1] + (mp_num, -1))                     # (B, T, 8, 3 * E // 8)179 180        query, value, key = torch.split(qkv_split, self.embed_dim // mp_num, dim=-1) # 3 * (B, T, 8, E // 8)181 182        query = self._split_heads(query, self.num_attention_heads, self.head_dim)  # (B, T, n_heads, dim_head)183        key = self._split_heads(key, self.num_attention_heads, self.head_dim)      # (B, T, n_heads, dim_head)184        value = self._split_heads(value, self.num_attention_heads, self.head_dim)  # (B, T, n_heads, dim_head)185        value = value.permute(0, 2, 1, 3)186 187        seq_len = key.shape[1]188        offset = 0189 190        if layer_past is not None:191            offset = layer_past[0].shape[-2]192            seq_len += offset193 194        if self.rotary_dim is not None:195            k_rot = key[:, :, :, : self.rotary_dim]196            k_pass = key[:, :, :, self.rotary_dim :]197 198            q_rot = query[:, :, :, : self.rotary_dim]199            q_pass = query[:, :, :, self.rotary_dim :]200 201            sincos = fixed_pos_embedding(k_rot, 1, seq_len=seq_len)202            k_rot = apply_rotary_pos_emb(k_rot, sincos, offset=offset)203            q_rot = apply_rotary_pos_emb(q_rot, sincos, offset=offset)204 205            key = torch.cat([k_rot, k_pass], dim=-1)206            query = torch.cat([q_rot, q_pass], dim=-1)207        else:208            sincos = fixed_pos_embedding(key, 1, seq_len=seq_len)209            key = apply_rotary_pos_emb(key, sincos, offset=offset)210            query = apply_rotary_pos_emb(query, sincos, offset=offset)211 212        key = key.permute(0, 2, 1, 3)213        query = query.permute(0, 2, 1, 3)214 215        if layer_past is not None:216            past_key = layer_past[0]217            past_value = layer_past[1]218            key = torch.cat((past_key, key), dim=-2)219            value = torch.cat((past_value, value), dim=-2)220 221        if use_cache is True:222            present = (key, value)223        else:224            present = None225 226        # compute self-attention: softmax((Q @ K.T) / sqrt(dim_head)) @ V227        attn_output, attn_weights = self._attn(228            query, key, value, attention_mask, head_mask229        )230 231        attn_output = self._merge_heads(                               # (B, T, E)  232            attn_output, self.num_attention_heads, self.head_dim233        )234 235        attn_output = self.out_proj(attn_output)236        attn_output = self.resid_dropout(attn_output)237 238        outputs = (attn_output, present)239        if output_attentions:240            outputs += (attn_weights,)241 242        return outputs  # a, present, (attentions)243 244 245class ProGenMLP(nn.Module):246    def __init__(247        self, intermediate_size, config248    ):  # in MLP: intermediate_size= 4 * embed_dim249        super().__init__()250        embed_dim = config.embed_dim251 252        self.fc_in = nn.Linear(embed_dim, intermediate_size)253        self.fc_out = nn.Linear(intermediate_size, embed_dim)254 255        self.act = ACT2FN[config.activation_function]256        self.dropout = nn.Dropout(config.resid_pdrop)257 258    def forward(self, hidden_states):259        hidden_states = self.fc_in(hidden_states)260        hidden_states = self.act(hidden_states)261        hidden_states = self.fc_out(hidden_states)262        hidden_states = self.dropout(hidden_states)263        return hidden_states264 265 266class ProGenBlock(nn.Module):267    def __init__(self, config):268        super().__init__()269        inner_dim = config.n_inner if config.n_inner is not None else 4 * config.embed_dim270        self.ln_1 = nn.LayerNorm(config.embed_dim, eps=config.layer_norm_epsilon)271        self.attn = ProGenAttention(config)272        self.mlp = ProGenMLP(inner_dim, config)273 274    def forward(275        self,276        hidden_states,277        layer_past=None,278        attention_mask=None,279        head_mask=None,280        use_cache=False,281        output_attentions=False,282    ):283        residual = hidden_states284        hidden_states = self.ln_1(hidden_states)285        attn_outputs = self.attn(286            hidden_states,287            layer_past=layer_past,288            attention_mask=attention_mask,289            head_mask=head_mask,290            use_cache=use_cache,291            output_attentions=output_attentions,292        )293        attn_output = attn_outputs[0]  # output_attn: a, present, (attentions)294        outputs = attn_outputs[1:]295 296        feed_forward_hidden_states = self.mlp(hidden_states)            # (B, T, E)297        hidden_states = attn_output + feed_forward_hidden_states + residual298 299        if use_cache:300            outputs = (hidden_states,) + outputs301        else:302            outputs = (hidden_states,) + outputs[1:]303 304        return outputs  # hidden_states, present, (attentions)305 306 307class ProGenPreTrainedModel(PreTrainedModel):308    """309    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained310    models.311    """312 313    config_class = ProGenConfig314    base_model_prefix = "transformer"315    is_parallelizable = False316 317    def __init__(self, *inputs, **kwargs):318        super().__init__(*inputs, **kwargs)319 320    def _init_weights(self, module):321        """Initialize the weights."""322        if isinstance(module, (nn.Linear,)):323            # Slightly different from Mesh Transformer JAX which uses truncated_normal for initialization324            # cf https://github.com/pytorch/pytorch/pull/5617325            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)326            if module.bias is not None:327                module.bias.data.zero_()328        elif isinstance(module, nn.Embedding):329            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)330            if module.padding_idx is not None:331                module.weight.data[module.padding_idx].zero_()332        elif isinstance(module, nn.LayerNorm):333            module.bias.data.zero_()334            module.weight.data.fill_(1.0)335 336 337class ProGenModel(ProGenPreTrainedModel):338    def __init__(self, config):339        super().__init__(config)340        self.vocab_size_emb = config.vocab_size_emb341        self.embed_dim = config.embed_dim342        self.wte = nn.Embedding(config.vocab_size_emb, self.embed_dim)343        self.drop = nn.Dropout(config.embd_pdrop)344        self.h = nn.ModuleList([ProGenBlock(config) for _ in range(config.n_layer)])345        self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)346        self.rotary_dim = min(347            config.rotary_dim, config.n_positions // config.n_head348        )349        self.init_weights()350 351    def forward(352        self,353        input_ids=None,354        past_key_values=None,355        attention_mask=None,356        token_type_ids=None,357        position_ids=None,358        head_mask=None,359        inputs_embeds=None,360        use_cache=None,361        output_attentions=None,362        output_hidden_states=None,363        return_dict=None,364    ):365        output_attentions = (366            output_attentions367            if output_attentions is not None368            else self.config.output_attentions369        )370        output_hidden_states = (371            output_hidden_states372            if output_hidden_states is not None373            else self.config.output_hidden_states374        )375        use_cache = use_cache if use_cache is not None else self.config.use_cache376        return_dict = (377            return_dict if return_dict is not None else self.config.use_return_dict378        )379 380        if input_ids is not None and inputs_embeds is not None:381            raise ValueError(382                "You cannot specify both input_ids and inputs_embeds at the same time"383            )384        elif input_ids is not None:385            input_shape = input_ids.size()386            input_ids = input_ids.view(-1, input_shape[-1])387            batch_size = input_ids.shape[0]388        elif inputs_embeds is not None:389            input_shape = inputs_embeds.size()[:-1]390            batch_size = inputs_embeds.shape[0]391        else:392            raise ValueError("You have to specify either input_ids or inputs_embeds")393 394        device = input_ids.device if input_ids is not None else inputs_embeds.device395 396        if token_type_ids is not None:397            token_type_ids = token_type_ids.view(-1, input_shape[-1])398 399        if position_ids is not None:400            position_ids = position_ids.view(-1, input_shape[-1])401 402        if past_key_values is None:403            past_length = 0404            past_key_values = tuple([None] * len(self.h))405        else:406            past_length = past_key_values[0][0].size(-2)407 408        if position_ids is None:409            position_ids = torch.arange(410                past_length,411                input_shape[-1] + past_length,412                dtype=torch.long,413                device=device,414            )415            position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])416 417        # Attention mask.418        if attention_mask is not None:419            assert batch_size > 0, "batch_size has to be defined and > 0"420            attention_mask = attention_mask.view(batch_size, -1)421            # We create a 3D attention mask from a 2D tensor mask.422            # Sizes are [batch_size, 1, 1, to_seq_length]423            # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]424            # this attention mask is more simple than the triangular masking of causal attention425            # used in OpenAI GPT, we just need to prepare the broadcast dimension here.426            attention_mask = attention_mask[:, None, None, :]427 428            # Since attention_mask is 1.0 for positions we want to attend and 0.0 for429            # masked positions, this operation will create a tensor which is 0.0 for430            # positions we want to attend and -10000.0 for masked positions.431            # Since we are adding it to the raw scores before the softmax, this is432            # effectively the same as removing these entirely.433            attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility434            attention_mask = (1.0 - attention_mask) * -10000.0435 436        # Prepare head mask if needed437        # 1.0 in head_mask indicate we keep the head438        # attention_probs has shape bsz x num_attention_heads x N x N439        # head_mask has shape n_layer x batch x num_attention_heads x N x N440        head_mask = self.get_head_mask(head_mask, self.config.n_layer)441 442        if inputs_embeds is None:443            inputs_embeds = self.wte(input_ids)444 445        hidden_states = inputs_embeds446 447        if token_type_ids is not None:448            token_type_embeds = self.wte(token_type_ids)449            hidden_states = hidden_states + token_type_embeds450 451        hidden_states = self.drop(hidden_states)452 453        output_shape = input_shape + (hidden_states.size(-1),)454 455        presents = () if use_cache else None456        all_self_attentions = () if output_attentions else None457        all_hidden_states = () if output_hidden_states else None458        for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):459            if output_hidden_states:460                all_hidden_states = all_hidden_states + (hidden_states,)461 462            if getattr(self.config, "gradient_checkpointing", False) and self.training:463                if use_cache:464                    logger.warning(465                        "`use_cache=True` is incompatible with `config.gradient_checkpointing=True`. Setting "466                        "`use_cache=False`..."467                    )468                    use_cache = False469 470                def create_custom_forward(module):471                    def custom_forward(*inputs):472                        # None for past_key_value473                        return module(*inputs, use_cache, output_attentions)474 475                    return custom_forward476 477                outputs = torch.utils.checkpoint.checkpoint(478                    create_custom_forward(block),479                    hidden_states,480                    None,481                    attention_mask,482                    head_mask[i],483                )484            else:485                outputs = block(486                    hidden_states,487                    layer_past=layer_past,488                    attention_mask=attention_mask,489                    head_mask=head_mask[i],490                    use_cache=use_cache,491                    output_attentions=output_attentions,492                )493 494            hidden_states = outputs[0]495            if use_cache is True:496                presents = presents + (outputs[1],)497 498            if output_attentions:499                all_self_attentions = all_self_attentions + (500                    outputs[2 if use_cache else 1],501                )502 503        hidden_states = self.ln_f(hidden_states)504 505        hidden_states = hidden_states.view(*output_shape)506        # Add last hidden state507        if output_hidden_states:508            all_hidden_states = all_hidden_states + (hidden_states,)509 510        if not return_dict:511            return tuple(512                v513                for v in [514                    hidden_states,515                    presents,516                    all_hidden_states,517                    all_self_attentions,518                ]519                if v is not None520            )521 522        return BaseModelOutputWithPast(523            last_hidden_state=hidden_states,524            past_key_values=presents,525            hidden_states=all_hidden_states,526            attentions=all_self_attentions,527        )528 529 530class ProGenForCausalLM(ProGenPreTrainedModel):531    _keys_to_ignore_on_load_missing = [532        r"h\.\d+\.attn\.masked_bias",533        r"h\.\d+\.attn\.bias",534        r"lm_head\.weight",535    ]536 537    def __init__(self, config):538        super().__init__(config)539        self.transformer = ProGenModel(config)540        self.lm_head = nn.Linear(config.embed_dim, config.vocab_size_lm_head)541        self.init_weights()542 543    def prepare_inputs_for_generation(self, input_ids, past=None, **kwargs):544        token_type_ids = kwargs.get("token_type_ids", None)545        # only last token for inputs_ids if past is defined in kwargs546        if past:547            input_ids = input_ids[:, -1].unsqueeze(-1)548            if token_type_ids is not None:549                token_type_ids = token_type_ids[:, -1].unsqueeze(-1)550 551        attention_mask = kwargs.get("attention_mask", None)552        position_ids = kwargs.get("position_ids", None)553 554        if attention_mask is not None and position_ids is None:555            # create position_ids on the fly for batch generation556            position_ids = attention_mask.long().cumsum(-1) - 1557            position_ids.masked_fill_(attention_mask == 0, 1)558            if past:559                position_ids = position_ids[:, -1].unsqueeze(-1)560        else:561            position_ids = None562        return {563            "input_ids": input_ids,564            "past_key_values": past,565            "use_cache": kwargs.get("use_cache"),566            "position_ids": position_ids,567            "attention_mask": attention_mask,568            "token_type_ids": token_type_ids,569        }570 571    def forward(572        self,573        input_ids=None,574        past_key_values=None,575        attention_mask=None,576        token_type_ids=None,577        position_ids=None,578        head_mask=None,579        inputs_embeds=None,580        labels=None,581        use_cache=None,582        output_attentions=None,583        output_hidden_states=None,584        return_dict=None,585    ):586        r"""587        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):588            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set589            ``labels = input_ids`` Indices are selected in ``[-100, 0, ..., config.vocab_size]`` All labels set to590            ``-100`` are ignored (masked), the loss is only computed for labels in ``[0, ..., config.vocab_size]``591        """592        return_dict = (593            return_dict if return_dict is not None else self.config.use_return_dict594        )595 596        transformer_outputs = self.transformer(597            input_ids,598            past_key_values=past_key_values,599            attention_mask=attention_mask,600            token_type_ids=token_type_ids,601            position_ids=position_ids,602            head_mask=head_mask,603            inputs_embeds=inputs_embeds,604            use_cache=use_cache,605            output_attentions=output_attentions,606            output_hidden_states=output_hidden_states,607            return_dict=return_dict,608        )609        hidden_states = transformer_outputs[0]610 611        # make sure sampling in fp16 works correctly and612        # compute loss in fp32 to match with mesh-tf version613        # https://github.com/EleutherAI/gpt-neo/blob/89ce74164da2fb16179106f54e2269b5da8db333/models/gpt2/gpt2.py#L179614        lm_logits = self.lm_head(hidden_states).to(torch.float32)615 616        loss = None617        if labels is not None:618            # Shift so that tokens < n predict n619            shift_logits = lm_logits[..., :-1, :].contiguous()620            shift_labels = labels[..., 1:].contiguous()621            loss_fct = CrossEntropyLoss()622            loss = loss_fct(623                shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)624            )625            loss = loss.to(hidden_states.dtype)626 627        if not return_dict:628            output = (lm_logits,) + transformer_outputs[1:]629            return ((loss,) + output) if loss is not None else output630 631        return CausalLMOutputWithPast(632            loss=loss,633            logits=lm_logits,634            past_key_values=transformer_outputs.past_key_values,635            hidden_states=transformer_outputs.hidden_states,636            attentions=transformer_outputs.attentions,637        )638 639    @staticmethod640    def _reorder_cache(641        past: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor642    ) -> Tuple[Tuple[torch.Tensor]]:643        """644        This function is used to re-order the :obj:`past_key_values` cache if645        :meth:`~transformers.PretrainedModel.beam_search` or :meth:`~transformers.PretrainedModel.beam_sample` is646        called. This is required to match :obj:`past_key_values` with the correct beam_idx at every generation step.647        """648        return tuple(649            tuple(650                past_state.index_select(0, beam_idx.to(past_state.device))651                for past_state in layer_past652            )653            for layer_past in past654        )655