feat: merge stuff

Signed-off-by: jupyterjazz <saba.sturua@jina.ai>

config.json

@@ -3,8 +3,12 @@
     "AutoConfig": "configuration_xlm_roberta.XLMRobertaFlashConfig",
     "AutoModel": "modeling_xlm_roberta.XLMRobertaModel",
     "AutoModelForPreTraining": "modeling_xlm_roberta.XLMRobertaForPreTraining",
-    "AutoModelForMaskedLM": "modeling_xlm_roberta.XLMRobertaForMaskedLM"
+    "AutoModelForMaskedLM": "modeling_xlm_roberta.XLMRobertaForMaskedLM",
+    "AutoModelForSequenceClassification":"modeling_xlm_roberta.XLMRobertaForSequenceClassification"
   },
+  "architectures": [
+    "XLMRobertaModel"
+  ],
   "attention_probs_dropout_prob": 0.1,
   "bos_token_id": 0,
   "eos_token_id": 2,

configuration_xlm_roberta.py

@@ -1,4 +1,5 @@
 from transformers import PretrainedConfig
+import torch
 
 class XLMRobertaFlashConfig(PretrainedConfig):
     def __init__(
@@ -23,6 +24,9 @@ class XLMRobertaFlashConfig(PretrainedConfig):
             classifier_dropout=None,
             num_loras=1,
             load_trained_adapters=False,
+            use_flash_attn=True,
+            torch_dtype=None,
+            emb_pooler=None,
             **kwargs,
     ):
         super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
@@ -45,3 +49,9 @@ class XLMRobertaFlashConfig(PretrainedConfig):
         self.classifier_dropout = classifier_dropout
         self.num_loras = num_loras
         self.load_trained_adapters = load_trained_adapters
+        self.use_flash_attn = use_flash_attn
+        self.emb_pooler = emb_pooler
+        if torch_dtype and hasattr(torch, torch_dtype) and type(getattr(torch, torch_dtype)) is torch.dtype:
+            self.torch_dtype = getattr(torch, torch_dtype)
+        else:
+            self.torch_dtype = torch_dtype

convert_roberta_weights_to_flash.py

@@ -1,10 +1,11 @@
 import re
 from collections import OrderedDict
 from transformers import PretrainedConfig
-from transformers import XLMRobertaForMaskedLM
+from transformers import XLMRobertaForMaskedLM, XLMRobertaForSequenceClassification
 
 from .configuration_xlm_roberta import XLMRobertaFlashConfig as BertConfig
-from .modeling_xlm_roberta import XLMRobertaForMaskedLM as BertModel
+from .modeling_xlm_roberta import XLMRobertaForMaskedLM as FlashXLMRobertaForMaskedLM
+from .modeling_xlm_roberta import XLMRobertaForSequenceClassification as FlashXLMRobertaForSequenceClassification
 import torch
 
 import click
@@ -137,14 +138,23 @@ def remap_state_dict(state_dict, config: PretrainedConfig):
 
 @click.command()
 @click.option('--model_name', default='FacebookAI/xlm-roberta-base', help='model name')
+@click.option('--revision', default='main', help='revision')
+@click.option('--task', default='masked_lm', help='task')
 @click.option('--output', default='converted_roberta_weights.bin', help='model name')
-def main(model_name, output):
-    roberta_model = XLMRobertaForMaskedLM.from_pretrained(model_name)
+def main(model_name, revision, task, output):
+    
+    if task == 'masked_lm':
+        roberta_model = XLMRobertaForMaskedLM.from_pretrained(model_name, revision=revision)
+    elif task == 'sequence_classification':
+        roberta_model = XLMRobertaForSequenceClassification.from_pretrained(model_name, revision=revision,num_labels=1)
     config = BertConfig.from_dict(roberta_model.config.to_dict())
     state_dict = roberta_model.state_dict()
     new_state_dict = remap_state_dict(state_dict, config)
-
-    flash_model = BertModel(config)
+    
+    if task == 'masked_lm':
+        flash_model = FlashXLMRobertaForMaskedLM(config)
+    elif task == 'sequence_classification':
+        flash_model = FlashXLMRobertaForSequenceClassification(config)
 
     for k, v in flash_model.state_dict().items():
         if k not in new_state_dict:

mha.py

@@ -10,8 +10,6 @@ import torch
 import torch.nn as nn
 from einops import rearrange, repeat
 
-from flash_attn.utils.distributed import get_dim_for_local_rank
-
 try:
     from flash_attn import (
         flash_attn_kvpacked_func,

modeling_xlm_roberta.py

@@ -1,6 +1,5 @@
 # This implementation was adopted from https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/models/bert.py
 # Commit id: abbc1311731867310635f9edc2a9ec18317c8c48
-
 # Copyright (c) 2022, Tri Dao.
 # This BERT implementation is based on our MLPerf 2.0 and MLPerf 2.1 BERT implementation.
 # https://github.com/mlcommons/training_results_v2.0/blob/main/HazyResearch/benchmarks/bert/implementations/pytorch/modeling.py
@@ -8,20 +7,23 @@
 
 # Inspired by https://github.com/huggingface/transformers/blob/main/src/transformers/models/bert/modeling_bert.py
 
+import importlib.util
 import logging
 import re
 from collections import OrderedDict
 from collections.abc import Sequence
 from functools import partial
+import numpy as np
 
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
 import torch.utils.checkpoint
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 from einops import rearrange
 from transformers import PretrainedConfig
 from transformers.modeling_utils import PreTrainedModel
-from transformers.modeling_outputs import MaskedLMOutput
+from transformers.modeling_outputs import MaskedLMOutput,SequenceClassifierOutput
 from transformers.models.xlm_roberta.modeling_xlm_roberta import XLMRobertaLMHead
 
 from transformers.models.bert.modeling_bert import (
@@ -29,7 +31,7 @@ from transformers.models.bert.modeling_bert import (
     BertForPreTrainingOutput,
 )
 
-from typing import Optional, Tuple, Union
+from typing import List, Optional, Tuple, Union
 
 from .xlm_padding import (
     index_first_axis,
@@ -61,12 +63,30 @@ try:
 except ImportError:
     CrossEntropyLoss = None
 
+try:
+    from tqdm.autonotebook import trange
+except ImportError:
+    trange = None
+
 
 logger = logging.getLogger(__name__)
 
 
+def get_use_flash_attn(config: XLMRobertaFlashConfig):
+    if not getattr(config, "use_flash_attn", False):
+        return False
+    if not torch.cuda.is_available():
+        return False
+    if importlib.util.find_spec("flash_attn") is None:
+        logger.warning(
+            'flash_attn is not installed. Using PyTorch native attention implementation.'
+        )
+        return False
+    return True
+
+
 def create_mixer_cls(config, cross_attn=False, return_residual=False):
-    use_flash_attn = getattr(config, "use_flash_attn", False)
+    use_flash_attn = get_use_flash_attn(config)
     fused_bias_fc = getattr(config, "fused_bias_fc", False)
     rotary_kwargs = {}
     if config.position_embedding_type == "rotary":
@@ -169,7 +189,7 @@ def _init_weights(module, initializer_range=0.02):
 class XLMRobertaEncoder(nn.Module):
     def __init__(self, config: XLMRobertaFlashConfig):
         super().__init__()
-        self.use_flash_attn = getattr(config, "use_flash_attn", False)
+        self.use_flash_attn = get_use_flash_attn(config)
         self.layers = nn.ModuleList(
             [create_block(config, layer_idx=i) for i in range(config.num_hidden_layers)]
         )
@@ -376,6 +396,17 @@ class XLMRobertaPreTrainedModel(PreTrainedModel):
         if isinstance(module, XLMRobertaEncoder):
             module.gradient_checkpointing = value
 
+    @classmethod
+    def from_pretrained(
+        cls,
+        *args,
+        **kwargs,
+    ):
+        if not 'torch_dtype' in kwargs:
+            kwargs['torch_dtype'] = 'auto'
+        return super().from_pretrained(*args, **kwargs)
+
+
 
 class XLMRobertaModel(XLMRobertaPreTrainedModel):
     def __init__(self, config: XLMRobertaFlashConfig, add_pooling_layer=True):
@@ -409,6 +440,169 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
 
         self.apply(partial(_init_weights, initializer_range=config.initializer_range))
 
+
+    @torch.inference_mode()
+    def encode(
+        self: 'XLMRobertaModel',
+        sentences: Union[str, List[str]],
+        batch_size: int = 32,
+        show_progress_bar: Optional[bool] = None,
+        output_value: str = 'sentence_embedding',
+        convert_to_numpy: bool = True,
+        convert_to_tensor: bool = False,
+        device: Optional[torch.device] = None,
+        normalize_embeddings: bool = False,
+        **tokenizer_kwargs,
+    ) -> Union[List[torch.Tensor], np.ndarray, torch.Tensor]:
+        """
+        Computes sentence embeddings
+        Args:
+            sentences(`str` or `List[str]`):
+                Sentence or sentences to be encoded
+            batch_size(`int`, *optional*, defaults to 32):
+                Batch size for the computation
+            show_progress_bar(`bool`, *optional*, defaults to None):
+                Show a progress bar when encoding sentences.
+                If set to None, progress bar is only shown when
+                `logger.level == logging.INFO` or `logger.level == logging.DEBUG`.
+            output_value(`str`, *optional*, defaults to 'sentence_embedding'):
+                Default sentence_embedding, to get sentence embeddings.
+                Can be set to token_embeddings to get wordpiece token embeddings.
+                Set to None, to get all output values
+            convert_to_numpy(`bool`, *optional*, defaults to True):
+                If true, the output is a list of numpy vectors.
+                Else, it is a list of pytorch tensors.
+            convert_to_tensor(`bool`, *optional*, defaults to False):
+                If true, you get one large tensor as return.
+                Overwrites any setting from convert_to_numpy
+            device(`torch.device`, *optional*, defaults to None):
+                Which torch.device to use for the computation
+            normalize_embeddings(`bool`, *optional*, defaults to False):
+                If set to true, returned vectors will have length 1. In that case, the
+                faster dot-product (util.dot_score) instead of cosine similarity can
+                be used.
+            tokenizer_kwargs(`Dict[str, Any]`, *optional*, defaults to {}):
+                Keyword arguments for the tokenizer
+        Returns:
+            By default, a list of tensors is returned.
+            If convert_to_tensor, a stacked tensor is returned.
+            If convert_to_numpy, a numpy matrix is returned.
+        """
+        from transformers import AutoTokenizer
+
+        self.tokenizer = AutoTokenizer.from_pretrained(
+            self.name_or_path, trust_remote_code=True
+        )
+
+        is_training = self.training
+        self.eval()
+
+        if show_progress_bar is None:
+            show_progress_bar = (
+                logger.getEffectiveLevel() == logging.INFO
+                or logger.getEffectiveLevel() == logging.DEBUG
+            )
+
+        if convert_to_tensor:
+            convert_to_numpy = False
+
+        if output_value != 'sentence_embedding':
+            convert_to_tensor = False
+            convert_to_numpy = False
+
+        input_was_string = False
+        if isinstance(sentences, str) or not hasattr(sentences, '__len__'):
+            sentences = [sentences]
+            input_was_string = True
+
+        if device is not None:
+            self.to(device)
+
+        permutation = np.argsort([-len(i) for i in sentences])
+        inverse_permutation = np.argsort(permutation)
+        sentences = [sentences[idx] for idx in permutation]
+
+        tokenizer_kwargs['padding'] = tokenizer_kwargs.get('padding', True)
+        tokenizer_kwargs['max_length'] = tokenizer_kwargs.get(
+            'max_length', self.tokenizer.init_kwargs.get('model_max_length', 8192)
+        )
+        tokenizer_kwargs['truncation'] = tokenizer_kwargs.get('truncation', True)
+
+        all_embeddings = []
+
+        if trange is not None:
+            range_iter = trange(
+                0,
+                len(sentences),
+                batch_size,
+                desc="Encoding",
+                disable=not show_progress_bar,
+            )
+        else:
+            range_iter = range(0, len(sentences), batch_size)
+
+        for i in range_iter:
+            encoded_input = self.tokenizer(
+                sentences[i : i + batch_size],
+                return_tensors='pt',
+                **tokenizer_kwargs,
+            ).to(self.device)
+            token_embs = self.forward(**encoded_input)[0]
+
+            # Accumulate in fp32 to avoid overflow
+            token_embs = token_embs.float()
+
+            if output_value == 'token_embeddings':
+                raise NotImplementedError
+            elif output_value is None:
+                raise NotImplementedError
+            else:
+                if self.config.emb_pooler == 'cls':
+                    embeddings = self.cls_pooling(
+                        token_embs, encoded_input['attention_mask']
+                    )
+                else:
+                    embeddings = self.mean_pooling(
+                        token_embs, encoded_input['attention_mask']
+                    )
+
+                if normalize_embeddings:
+                    embeddings = torch.nn.functional.normalize(embeddings, p=2, dim=1)
+
+                if convert_to_numpy:
+                    embeddings = embeddings.cpu()
+            all_embeddings.extend(embeddings)
+
+        all_embeddings = [all_embeddings[idx] for idx in inverse_permutation]
+
+        if convert_to_tensor:
+            all_embeddings = torch.stack(all_embeddings)
+        elif convert_to_numpy:
+            all_embeddings = np.asarray([emb.numpy() for emb in all_embeddings])
+
+        if input_was_string:
+            all_embeddings = all_embeddings[0]
+
+        self.train(is_training)
+        return all_embeddings
+
+    def mean_pooling(
+        self, token_embeddings: torch.Tensor, attention_mask: torch.Tensor
+    ):
+        input_mask_expanded = (
+            attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
+        )
+        return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(
+            input_mask_expanded.sum(1), min=1e-9
+        )
+
+
+    def cls_pooling(
+        self, token_embeddings: torch.Tensor, attention_mask: torch.Tensor
+    ):
+        return token_embeddings[:,0]
+
+
     def forward(
         self,
         input_ids,
@@ -946,3 +1140,117 @@ def inv_remap_state_dict(state_dict, config: PretrainedConfig):
     )
 
     return state_dict
+
+
+# Copied from transformers.models.roberta.modeling_roberta.RobertaClassificationHead with Roberta->XLMRoberta
+class XLMRobertaClassificationHead(nn.Module):
+    """Head for sentence-level classification tasks."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        classifier_dropout = (
+            config.classifier_dropout
+            if config.classifier_dropout is not None
+            else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
+
+    def forward(self, features, **kwargs):
+        x = features[:, 0, :]  # take <s> token (equiv. to [CLS])
+        x = self.dropout(x)
+        x = self.dense(x)
+        x = torch.tanh(x)
+        x = self.dropout(x)
+        x = self.out_proj(x)
+        return x
+
+
+# Copied from transformers.models.roberta.modeling_roberta.RobertaForSequenceClassification with Roberta->XLMRoberta, ROBERTA->XLM_ROBERTA
+class XLMRobertaForSequenceClassification(XLMRobertaPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.config = config
+
+        self.roberta = XLMRobertaModel(config, add_pooling_layer=False)
+        self.classifier = XLMRobertaClassificationHead(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], SequenceClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        outputs = self.roberta(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = outputs[0]
+        logits = self.classifier(sequence_output)
+
+        loss = None
+        if labels is not None:
+            # move labels to correct device to enable model parallelism
+            labels = labels.to(logits.device)
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (
+                    labels.dtype == torch.long or labels.dtype == torch.int
+                ):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

modeling_xlm_roberta_for_glue.py

@@ -0,0 +1,109 @@
+from typing import Optional, Union, Tuple
+
+import torch
+from torch import nn
+from torch.nn import CrossEntropyLoss, MSELoss, BCEWithLogitsLoss
+from transformers.modeling_outputs import SequenceClassifierOutput, QuestionAnsweringModelOutput, TokenClassifierOutput
+
+from .modeling_xlm_roberta import XLMRobertaPreTrainedModel, XLMRobertaModel
+from .configuration_xlm_roberta import XLMRobertaFlashConfig
+
+
+class XLMRobertaForSequenceClassification(XLMRobertaPreTrainedModel):
+    def __init__(self, config: XLMRobertaFlashConfig):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.config = config
+
+        self.roberta = XLMRobertaModel(config)
+        classifier_dropout = (
+            config.classifier_dropout
+            if config.classifier_dropout is not None
+            else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], SequenceClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        assert head_mask is None
+        assert inputs_embeds is None
+        assert output_attentions is None
+        assert output_hidden_states is None
+        assert return_dict
+        outputs = self.roberta(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (
+                    labels.dtype == torch.long or labels.dtype == torch.int
+                ):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )