first

README.md

@@ -0,0 +1,109 @@
+---
+language:
+- en
+tags:
+- summarization
+datasets:
+- scientific_papers
+metrics:
+- rouge
+model-index:
+- name: ccdv/lsg-bart-base-16384-mediasum
+  results: []
+---
+
+<!-- This model card has been generated automatically according to the information the Trainer had access to. You
+should probably proofread and complete it, then remove this comment. -->
+
+**This model relies on a custom modeling file, you need to add trust_remote_code=True**\
+**See [\#13467](https://github.com/huggingface/transformers/pull/13467)**
+
+```python
+from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline
+
+tokenizer = AutoTokenizer.from_pretrained("ccdv/lsg-bart-base-16384-mediasum", trust_remote_code=True)
+model = AutoModelForSeq2SeqLM.from_pretrained("ccdv/lsg-bart-base-16384-mediasum", trust_remote_code=True)
+
+text = "Replace by what you want."
+pipe = pipeline("text2text-generation", model=model, tokenizer=tokenizer, device=0)
+generated_text = pipe(
+  text, 
+  truncation=True, 
+  max_length=64, 
+  no_repeat_ngram_size=7,
+  num_beams=2,
+  early_stopping=True
+  )
+```
+
+# ccdv/lsg-bart-base-16384-mediasum
+
+This model is a fine-tuned version of [ccdv/lsg-bart-base-4096-mediasum](https://huggingface.co/ccdv/lsg-bart-base-4096-mediasum) on the [ccdv/mediasum roberta_prepended mediasum](https://huggingface.co/datasets/ccdv/mediasum) dataset. \
+The model is converted to handle 16384 long sequences and fine-tuned accordingly during 1 epoch. \
+It achieves the following results on the test set:
+
+| Length | Global tokens | Fine-tuning | Block Size | Sparsity | Connexions | R1    | R2    | RL    | RLsum |
+|:------ |:------------- |:----------- |:---------- |:-------- | :--------- |:----- |:----- |:----- |:----- |
+| 16384  | 64            | Full        | 256        | 0        | 768        | 35.31 | 18.35 | 31.81 | 32.47 |
+| 16384  | 1             | Full        | 256        | 0        | 768        | 35.21 | 18.20 | 31.73 | 32.37 |
+| 16384  | 64            | Global only | 256        | 0        | 768        | 35.22 | 18.08 | 31.54 | 32.21 |
+| 16384  | 1             | None        | 256        | 0        | 768        | 35.17 | 18.13 | 31.54 | 32.20 |
+
+Reference model:
+
+| Length | Global tokens | Fine-tuning | Block Size | Sparsity | Connexions | R1    | R2    | RL    | RLsum |
+|:------ |:------------- |:----------- |:---------- |:-------- | :--------- |:----- |:----- |:----- |:----- |
+| 4096   | 1             | -           | 256        | 0        | 768        | 35.16 | 18.13 | 31.54 | 32.20
+
+## Model description
+The model relies on Local-Sparse-Global attention to handle long sequences:
+![attn](attn.png)
+
+The model has about ~145 millions parameters (6 encoder layers - 6 decoder layers). \
+The model is warm started from [ccdv/lsg-bart-base-4096-mediasum](https://huggingface.co/ccdv/lsg-bart-base-4096-mediasum), converted to handle long sequences (encoder only) and fine tuned.
+
+## Intended uses & limitations
+
+More information needed
+
+## Training and evaluation data
+
+More information needed
+
+## Training procedure
+
+### Training hyperparameters
+
+The following hyperparameters were used during training:
+- learning_rate: 8e-05
+- train_batch_size: 8
+- seed: 42
+- gradient_accumulation_steps: 4
+- total_train_batch_size: 32
+- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
+- lr_scheduler_type: linear
+- lr_scheduler_warmup_ratio: 0.1
+- num_epochs: 1.0
+
+### Generate hyperparameters
+
+The following hyperparameters were used during generation:
+- dataset_name: ccdv/mediasum
+- dataset_config_name: roberta_prepended
+- eval_batch_size: 8
+- eval_samples: 10000
+- early_stopping: True
+- ignore_pad_token_for_loss: True
+- length_penalty: 2.0
+- max_length: 128
+- min_length: 3
+- num_beams: 5
+- no_repeat_ngram_size: None
+- seed: 123
+
+### Framework versions
+
+- Transformers 4.18.0
+- Pytorch 1.10.1+cu102
+- Datasets 2.1.0
+- Tokenizers 0.11.6

all_results.json

@@ -0,0 +1,8 @@
+{
+    "epoch": 1.0,
+    "train_loss": 1.5347596183515975,
+    "train_runtime": 69651.7637,
+    "train_samples": 443596,
+    "train_samples_per_second": 6.369,
+    "train_steps_per_second": 0.199
+}

config.json

@@ -0,0 +1,94 @@
+{
+  "_name_or_path": "ccdv/lsg-bart-base-16384-mediasum",
+  "activation_dropout": 0.1,
+  "activation_function": "gelu",
+  "adaptive": true,
+  "add_bias_logits": false,
+  "add_final_layer_norm": false,
+  "architectures": [
+    "LSGBartForConditionalGeneration"
+  ],
+  "attention_dropout": 0.1,
+  "auto_map": {
+    "AutoConfig": "modeling_lsg_bart.LSGBartConfig",
+    "AutoModel": "modeling_lsg_bart.LSGBartModel",
+    "AutoModelForCausalLM": "modeling_lsg_bart.LSGBartForCausalLM",
+    "AutoModelForQuestionAnswering": "modeling_lsg_bart.LSGBartForQuestionAnswering",
+    "AutoModelForSeq2SeqLM": "modeling_lsg_bart.LSGBartForConditionalGeneration",
+    "AutoModelForSequenceClassification": "modeling_lsg_bart.LSGBartForSequenceClassification"
+  },
+  "base_model_prefix": "lsg",
+  "block_size": 256,
+  "bos_token_id": 0,
+  "classif_dropout": 0.1,
+  "classifier_dropout": 0.0,
+  "d_model": 768,
+  "decoder_attention_heads": 12,
+  "decoder_ffn_dim": 3072,
+  "decoder_layerdrop": 0.0,
+  "decoder_layers": 6,
+  "decoder_start_token_id": 2,
+  "dropout": 0.1,
+  "early_stopping": true,
+  "encoder_attention_heads": 12,
+  "encoder_ffn_dim": 3072,
+  "encoder_layerdrop": 0.0,
+  "encoder_layers": 6,
+  "eos_token_id": 2,
+  "forced_bos_token_id": 0,
+  "forced_eos_token_id": 2,
+  "gradient_checkpointing": false,
+  "id2label": {
+    "0": "LABEL_0",
+    "1": "LABEL_1",
+    "2": "LABEL_2"
+  },
+  "init_std": 0.02,
+  "is_encoder_decoder": true,
+  "label2id": {
+    "LABEL_0": 0,
+    "LABEL_1": 1,
+    "LABEL_2": 2
+  },
+  "lsh_num_pre_rounds": 1,
+  "mask_first_token": false,
+  "max_position_embeddings": 16384,
+  "model_type": "bart",
+  "no_repeat_ngram_size": 3,
+  "normalize_before": false,
+  "normalize_embedding": true,
+  "num_beams": 4,
+  "num_global_tokens": 64,
+  "num_hidden_layers": 6,
+  "pad_token_id": 1,
+  "pass_global_tokens_to_decoder": true,
+  "pool_with_global": true,
+  "scale_embedding": false,
+  "sparse_block_size": 0,
+  "sparsity_factor": 4,
+  "sparsity_type": "none",
+  "task_specific_params": {
+    "summarization": {
+      "length_penalty": 1.0,
+      "max_length": 128,
+      "min_length": 12,
+      "num_beams": 4
+    },
+    "summarization_cnn": {
+      "length_penalty": 2.0,
+      "max_length": 142,
+      "min_length": 56,
+      "num_beams": 4
+    },
+    "summarization_xsum": {
+      "length_penalty": 1.0,
+      "max_length": 62,
+      "min_length": 11,
+      "num_beams": 6
+    }
+  },
+  "torch_dtype": "float32",
+  "transformers_version": "4.19.2",
+  "use_cache": true,
+  "vocab_size": 50265
+}

modeling_lsg_bart.py

@@ -0,0 +1,1122 @@
+from logging import warn
+import torch
+from transformers.models.bart.modeling_bart import *
+from transformers.models.bart.modeling_bart import _expand_mask
+import torch.nn as nn
+from torch.nn import BCEWithLogitsLoss
+import sys
+
+AUTO_MAP = {
+        "AutoModel": "modeling_lsg_bart.LSGBartModel",
+        "AutoModelForCausalLM": "modeling_lsg_bart.LSGBartForCausalLM",
+        "AutoModelForQuestionAnswering": "modeling_lsg_bart.LSGBartForQuestionAnswering",
+        "AutoModelForSequenceClassification": "modeling_lsg_bart.LSGBartForSequenceClassification",
+        "AutoModelForSeq2SeqLM": "modeling_lsg_bart.LSGBartForConditionalGeneration"
+    }
+
+class LSGBartConfig(BartConfig):
+    """
+    This class overrides :class:`~transformers.RobertaConfig`. Please check the superclass for the appropriate
+    documentation alongside usage examples.
+    """
+
+    base_model_prefix = "lsg"
+    model_type = "bart"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    attribute_map = {"num_attention_heads": "encoder_attention_heads", "hidden_size": "d_model"}
+
+    def __init__(
+        self,
+        adaptive=True,
+        base_model_prefix="lsg",
+        block_size=128,
+        lsh_num_pre_rounds=1,
+        mask_first_token=False,
+        num_global_tokens=1,
+        pass_global_tokens_to_decoder=True,
+        pool_with_global=True,
+        sparse_block_size=128,
+        sparsity_factor=2,
+        sparsity_type="norm",
+        **kwargs
+        ):
+        """Constructs LSGConfig."""
+        super().__init__(**kwargs)
+        
+        self.adaptive = adaptive
+        self.auto_map = AUTO_MAP
+        self.base_model_prefix = base_model_prefix
+        self.block_size = block_size
+        self.lsh_num_pre_rounds = lsh_num_pre_rounds
+        self.mask_first_token = mask_first_token
+        self.num_global_tokens = num_global_tokens
+        self.pass_global_tokens_to_decoder = pass_global_tokens_to_decoder
+        self.pool_with_global = pool_with_global
+        self.sparse_block_size = sparse_block_size
+        self.sparsity_factor = sparsity_factor
+        self.sparsity_type = sparsity_type
+
+        if sparsity_type not in [None, "none", "norm", "lsh", "pooling", "stride", "block_stride"]:
+            logger.warning(
+                "[WARNING CONFIG]: sparsity_mode not in [None, 'none', 'norm', 'lsh', 'pooling', 'stride', 'block_stride'], setting sparsity_type=None, computation will skip sparse attention")
+            self.sparsity_type = None
+
+        if self.sparsity_type in ["stride", "block_stride"]:
+            if self.sparsity_factor > self.encoder_attention_heads:
+                logger.warning(
+                "[WARNING CONFIG]: sparsity_factor > encoder_attention_heads is not recommended for stride/block_stride sparsity"
+            )
+        
+        if self.num_global_tokens < 1:
+            logger.warning(
+                "[WARNING CONFIG]: num_global_tokens < 1 is not compatible, setting num_global_tokens=1"
+            )
+            self.num_global_tokens = 1
+        elif self.num_global_tokens > 512:
+            logger.warning(
+                "[WARNING CONFIG]: num_global_tokens > 512 is not compatible, setting num_global_tokens=512"
+            )
+            self.num_global_tokens = 512
+        
+        if self.sparsity_factor > 0:
+            assert self.block_size % self.sparsity_factor == 0, "[ERROR CONFIG]: block_size must be divisible by sparsity_factor"
+            assert self.block_size//self.sparsity_factor >= 1, "[ERROR CONFIG]: make sure block_size >= sparsity_factor"
+            
+        
+class BaseSelfAttention(nn.Module):
+
+    def __init__(
+        self,
+        embed_dim,
+        num_heads,
+        dropout=0.0,
+        is_decoder=False,
+        bias=True,
+        ):
+
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+
+        if (self.head_dim * num_heads) != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
+                f" and `num_heads`: {num_heads})."
+            )
+        self.scaling = self.head_dim ** -0.5
+        self.is_decoder = is_decoder
+
+        self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+
+    def transpose_for_scores(self, x):
+        new_x_shape = x.size()[:-1] + (
+            self.num_heads,
+            self.head_dim,
+        )
+        x = x.view(*new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def reshape_output(self, context_layer):
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.embed_dim,)
+        return context_layer.view(*new_context_layer_shape)
+
+    def project_QKV(self, hidden_states):
+
+        query_layer = self.transpose_for_scores(self.q_proj(hidden_states))
+        key_layer = self.transpose_for_scores(self.k_proj(hidden_states))
+        value_layer = self.transpose_for_scores(self.v_proj(hidden_states))
+        return query_layer, key_layer, value_layer
+
+    
+class BaseAttentionProduct(nn.Module):
+
+    def __init__(self, config):
+        """
+        Compute attention: softmax(Q @ K.T) @ V
+        """
+        super().__init__()
+        self.dropout = nn.Dropout(config.attention_dropout)
+
+    def forward(self, query_layer, key_layer, value_layer, attention_mask=None):
+        
+        d = query_layer.shape[-1]
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = query_layer @ key_layer.transpose(-1, -2) / math.sqrt(d)
+
+        del query_layer
+        del key_layer
+
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in RobertaModel forward() function)
+            attention_scores = attention_scores + attention_mask
+            del attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        context_layer = self.dropout(attention_probs) @ value_layer
+
+        return context_layer
+
+
+class LSGAttentionProduct(nn.Module):
+
+    def __init__(self, config, block_size=None, sparse_block_size=None, sparsity_factor=4):
+        """
+        Compute block or overlapping blocks attention products
+        """
+        super().__init__()
+ 
+        self.block_size = block_size
+        self.sparse_block_size = sparse_block_size
+        self.sparsity_factor = sparsity_factor
+
+        if self.block_size is None:
+            self.block_size = config.block_size
+
+        if self.sparse_block_size is None:
+            self.sparse_block_size = config.sparse_block_size
+
+        # Shape of blocks
+        self.local_shapes = (self.block_size*3, self.block_size)
+        if self.sparse_block_size and self.sparsity_factor > 0:
+            self.sparse_shapes = (self.sparse_block_size*3, self.block_size//self.sparsity_factor)
+
+        self.attention = BaseAttentionProduct(config)
+        
+    def build_lsg_inputs(self, hidden_states, sparse_hidden_states, global_hidden_states, is_attn_mask=False):
+        
+        # Build local tokens
+        local_hidden_states = self.reshape_to_local_block(hidden_states, is_attn_mask)
+        del hidden_states
+
+        # Build sparse tokens
+        if sparse_hidden_states is not None:
+            sparse_hidden_states = self.reshape_to_sparse_block(sparse_hidden_states, is_attn_mask)
+        
+        return self.cat_global_sparse_local_tokens(global_hidden_states, sparse_hidden_states, local_hidden_states)
+
+    def forward(
+        self, 
+        query_layer, 
+        key_layer, 
+        value_layer, 
+        attention_mask=None, 
+        sparse_key=None,
+        sparse_value=None, 
+        sparse_mask=None, 
+        global_key=None, 
+        global_value=None, 
+        global_mask=None
+        ):
+
+        # Input batch, heads, length, hidden_size
+        n, h, t, d = query_layer.size()
+        n_blocks = t // self.block_size
+        assert t % self.block_size == 0
+
+        key_layer = self.build_lsg_inputs(
+            key_layer, 
+            sparse_key, 
+            global_key
+            )
+        del sparse_key
+        del global_key
+
+        value_layer = self.build_lsg_inputs(
+            value_layer, 
+            sparse_value, 
+            global_value
+            )
+        del sparse_value
+        del global_value
+
+        attention_mask = self.build_lsg_inputs(
+            attention_mask, 
+            sparse_mask, 
+            global_mask.transpose(-1, -2), 
+            is_attn_mask=True
+            ).transpose(-1, -2)
+        del sparse_mask
+        del global_mask
+
+        # expect (..., t, d) shape
+        # Compute attention
+        context_layer = self.attention(
+                query_layer=self.chunk(query_layer, n_blocks), 
+                key_layer=key_layer,
+                value_layer=value_layer,
+                attention_mask=attention_mask
+                )
+                
+        return context_layer.reshape(n, h, -1, d)
+    
+    def reshape_to_local_block(self, hidden_states, is_attn_mask=False):
+        
+        size, step = self.local_shapes
+        s = (size - step) // 2
+
+        # Pad before block reshaping
+        if is_attn_mask:
+            pad_value = -10000  
+            hidden_states = hidden_states.transpose(-1, -2)
+        else: 
+            pad_value = 0
+
+        hidden_states = torch.nn.functional.pad(
+            hidden_states.transpose(-1, -2), 
+            pad=(s, s),
+            value=pad_value
+            ).transpose(-1, -2)
+
+        # Make blocks
+        hidden_states = hidden_states.unfold(-2, size=size, step=step).transpose(-1, -2)
+
+        return hidden_states
+
+    def reshape_to_sparse_block(self, hidden_states, is_attn_mask=False):
+        
+        size, step = self.sparse_shapes
+
+        # In case of odd case
+        odd_offset = (step % 2)
+
+        # n, h, t, d*2 + 1
+        size = size*2 
+        s = (size - step) // 2 + odd_offset
+
+        # Pad before block reshaping
+        if is_attn_mask:
+            pad_value = -10000  
+            hidden_states = hidden_states.transpose(-1, -2)
+        else: 
+            pad_value = 0
+
+        hidden_states = torch.nn.functional.pad(
+            hidden_states.transpose(-1, -2), 
+            pad=(s, s),
+            value=pad_value
+            ).transpose(-1, -2)
+
+        # Make blocks
+        hidden_states = hidden_states.unfold(-2, size=size, step=step).transpose(-1, -2)
+
+        # Fix case where block_size == sparsify_factor
+        if odd_offset: 
+            hidden_states = hidden_states[..., :-1, :, :]
+
+        # Indexes for selection
+        u = (size - self.block_size * 3 // self.sparsity_factor) // 2 + odd_offset
+        s = self.sparse_block_size
+
+        u_ = u + odd_offset
+        return torch.cat([hidden_states[..., u-s:u, :], hidden_states[..., -u_:-u_+s, :]], dim=-2)
+
+    def cat_global_sparse_local_tokens(self, x_global, x_sparse=None, x_local=None, dim=-2):
+
+        n, h, b, t, d = x_local.size()
+        x_global = x_global.unsqueeze(-3).expand(-1, -1, b, -1, -1)
+        if x_sparse is not None:
+            return torch.cat([x_global, x_sparse, x_local], dim=dim)
+        return torch.cat([x_global, x_local], dim=dim)
+
+    def chunk(self, x, n_blocks):
+
+        t, d = x.size()[-2:]
+        return x.reshape(*x.size()[:-2], n_blocks, -1, d)
+
+
+class LSGBartEncoderAttention(BaseSelfAttention):
+    '''
+    Compute local attention with overlapping blocs
+    Use global attention for tokens with highest norm
+    '''
+    def __init__(
+        self, 
+        config, 
+        embed_dim,
+        num_heads,
+        dropout
+        ):
+
+        super().__init__(embed_dim, num_heads, dropout)
+
+        self.block_size = config.block_size
+        self.sparse_block_size = config.sparse_block_size
+        self.num_global_tokens = config.num_global_tokens
+        self.sparsity_factor = config.sparsity_factor
+
+        self.attention = LSGAttentionProduct(
+            config, 
+            block_size=config.block_size, 
+            sparse_block_size=config.sparse_block_size, 
+            sparsity_factor=self.sparsity_factor, 
+            )
+
+        self.full_attention = BaseAttentionProduct(config)
+
+        sparse_functions = {
+            "norm": self.get_sparse_tokens_with_norm, 
+            "pooling": self.get_sparse_tokens_with_pooling,
+            "lsh": self.get_sparse_tokens_with_lsh,
+            "stride": self.get_sparse_tokens_with_stride,
+            "block_stride": self.get_sparse_tokens_with_block_stride,
+            }
+        
+        self.sparsity_type = config.sparsity_type
+        self.get_sparse_elements = sparse_functions.get(self.sparsity_type, lambda x, y, z: (None, None, None))
+            
+        if config.sparsity_type == "lsh":
+            self.lsh_num_pre_rounds = config.lsh_num_pre_rounds
+        
+    def get_sparse_tokens_with_norm(self, keys, values, mask):
+        
+        if self.sparsity_factor == 1:
+            return keys, values, mask.expand(-1, keys.size()[1], -1, -1)
+
+        with torch.no_grad():
+
+            block_size = min(self.block_size, self.sparse_block_size)
+            key_norm = keys.detach().norm(dim=-1, keepdim=True)
+            key_norm = key_norm * ~mask.transpose(-1, -2).bool()
+            key_norm = self.chunk(key_norm, block_size)
+
+            n, h, b, t, d = key_norm.size()
+            
+            idx = key_norm.argsort(dim=-2) 
+            del key_norm
+            idx += (torch.arange(b, device=keys.device)*t).reshape(1, 1, b, 1, 1)
+
+            split = (t - block_size // self.sparsity_factor, block_size // self.sparsity_factor)
+            sparse_idx = idx.split(split, -2)[-1].reshape(n, h, -1, 1)
+        
+        d = keys.size()[-1]
+        keys = keys.gather(dim=-2, index=sparse_idx.expand(-1, -1, -1, d))
+        values = values.gather(dim=-2, index=sparse_idx.expand(-1, -1, -1, d))
+        mask = mask.expand(-1, h, -1, -1).transpose(-1, -2).gather(dim=-2, index=sparse_idx).transpose(-1, -2)
+
+        return keys, values, mask
+
+    def get_sparse_tokens_with_pooling(self, keys, values, mask):
+        
+        if self.sparsity_factor == 1:
+            return keys, values, mask.expand(-1, keys.size()[1], -1, -1)
+
+        keys = self.chunk(keys, self.sparsity_factor)
+        values = self.chunk(values, self.sparsity_factor)
+
+        n, h, b, t, d = keys.size()
+        mask = mask.reshape(n, 1, b, 1, t)
+        mask = ~mask.transpose(-1, -2).bool()
+
+        keys = keys * mask
+        values = values * mask
+
+        mask = mask.sum(dim=-2)
+        keys = keys.sum(dim=-2) / (mask + 1e-6)
+        values = values.sum(dim=-2) / (mask + 1e-6)
+
+        mask = - (1. - mask.clamp(0, 1)) * 1e4
+        return keys.reshape(n, h, -1, d), values.reshape(n, h, -1, d), mask.expand(-1, h, -1, -1).transpose(-1, -2)
+
+    def get_sparse_tokens_with_stride(self, keys, values, mask):
+
+        if self.sparsity_factor == 1:
+            return keys, values, mask.expand(-1, keys.size()[1], -1, -1)
+
+        n, h, t, d = keys.size()
+        sparse_idx = torch.arange(t // self.sparsity_factor, device=keys.device) * self.sparsity_factor
+        sparse_idx = sparse_idx.reshape(1, 1, -1, 1) + (torch.arange(h, device=keys.device) % self.sparsity_factor).reshape(1, h, 1, 1)
+        sparse_idx = sparse_idx.expand(n, h, -1, 1)
+
+        keys = keys.gather(dim=-2, index=sparse_idx.expand(-1, -1, -1, d))
+        values = values.gather(dim=-2, index=sparse_idx.expand(-1, -1, -1, d))
+        mask = mask.expand(-1, h, -1, -1).transpose(-1, -2).gather(dim=-2, index=sparse_idx).transpose(-1, -2)
+
+        return keys, values, mask
+
+    def get_sparse_tokens_with_block_stride(self, keys, values, mask):
+
+        if self.sparsity_factor == 1:
+            return keys, values, mask.expand(-1, keys.size()[1], -1, -1)
+
+        n, h, t, d = keys.size()
+
+        t, b = self.block_size, t // self.block_size
+        sparse_idx = torch.arange(t // self.sparsity_factor, device=keys.device)
+        sparse_idx = sparse_idx.reshape(1, 1, 1, -1, 1) + torch.arange(h, device=keys.device).reshape(1, h, 1, 1, 1) * (t // self.sparsity_factor)
+        sparse_idx = (sparse_idx % t) 
+        sparse_idx = sparse_idx + torch.arange(b, device=keys.device).reshape(1, 1, -1, 1, 1) * t
+        sparse_idx = sparse_idx.reshape(1, h, -1, 1).expand(n, h, -1, 1)
+
+        keys = keys.gather(dim=-2, index=sparse_idx.expand(-1, -1, -1, d))
+        values = values.gather(dim=-2, index=sparse_idx.expand(-1, -1, -1, d))
+        mask = mask.expand(-1, h, -1, -1).transpose(-1, -2).gather(dim=-2, index=sparse_idx).transpose(-1, -2)
+
+        return keys, values, mask
+        
+    def get_sparse_tokens_with_lsh(self, keys, values, mask):
+        
+        if self.sparsity_factor == 1:
+            return keys, values, mask.expand(-1, keys.size()[1], -1, -1)
+
+        block_size = min(self.block_size, self.sparse_block_size)
+        keys = self.chunk(keys, block_size)
+        values = self.chunk(values, block_size)
+
+        n, h, b, t, d = keys.size()
+        mask = mask.reshape(n, 1, b, 1, t)
+        mask = ~mask.transpose(-1, -2).bool()
+
+        keys = keys * mask
+        values = values * mask
+        mask = mask.expand(-1, h, -1, -1, -1).float()
+
+        extra_factor = 1
+        
+        for _ in range(self.lsh_num_pre_rounds):
+            keys, values, mask = self.lsh_round(keys, values, mask, t*extra_factor)
+
+        keys, values, mask = self.lsh_round(keys, values, mask, t//self.sparsity_factor)
+        keys /= mask + 1e-8
+        values /= mask + 1e-8
+
+        mask = -10000 * (1. - mask.clamp(0, 1))
+
+        return keys.reshape(n, h, -1, d), values.reshape(n, h, -1, d), mask.transpose(-1, -2).reshape(n, h, 1, -1)
+
+    def lsh_round(self, keys, values, mask, output_size):
+
+        with torch.no_grad():
+
+            n_hashes = output_size // 2
+            n, h, b, t, d = keys.size()
+            binary_mask = mask.clamp(0, 1)
+
+            indexes = (torch.nn.functional.normalize(keys, dim=-1) * binary_mask) @ torch.randn(1, h, 1, d, n_hashes, device=keys.device)
+            indexes = torch.cat([indexes, -indexes], dim=-1).argmax(dim=-1, keepdim=True)
+
+        n, h, b, t, d = keys.size()
+        
+        x_ = torch.zeros(n, h, b, output_size, d, device=keys.device)
+        mask_ = torch.zeros(n, h, b, output_size, 1, device=keys.device)
+        keys = torch.scatter_add(x_, dim=-2, index=indexes.expand(-1, -1, -1, -1, d), src=keys)
+        values = torch.scatter_add(x_, dim=-2, index=indexes.expand(-1, -1, -1, -1, d), src=values)
+        mask = torch.scatter_add(mask_, dim=-2, index=indexes, src=mask)
+
+        return keys[..., :output_size, :], values[..., :output_size, :], mask[..., :output_size, :]
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        layer_head_mask=None,
+        output_attentions=False
+        ):
+
+        query_layer, key_layer, value_layer = self.project_QKV(hidden_states)
+        outputs = self.not_causal_forward(
+            query_layer,
+            key_layer,
+            value_layer, 
+            attention_mask=attention_mask[:, :, :1, :], 
+            head_mask=layer_head_mask, 
+            output_attentions=output_attentions
+            )
+        
+        return self.out_proj(outputs), None, None
+
+    def not_causal_forward(
+        self,
+        query_layer,
+        key_layer,
+        value_layer,
+        attention_mask=None,
+        head_mask=None,
+        output_attentions=False,
+        ):
+
+        n, h, t, d = query_layer.size()
+
+        # Cat global mask
+        attention_mask = torch.nn.functional.pad(attention_mask, (self.num_global_tokens, 0), value=0)
+        
+        # Use normal attention if local attention covers every tokens
+        if t <= 2 * self.block_size + self.num_global_tokens:
+            context_layer = self.full_attention(
+                query_layer=query_layer, 
+                key_layer=key_layer, 
+                value_layer=value_layer, 
+                attention_mask=attention_mask
+                )
+
+            if head_mask is not None:
+                context_layer = context_layer * head_mask[:, :, :1, :1]
+            return self.reshape_output(context_layer)
+
+        # Split input into global tokens and other tokens
+        split = (self.num_global_tokens, t - self.num_global_tokens)
+        global_query, query_layer = query_layer.split(split, dim=-2)
+        
+        # Get global_attention
+        bos = self.full_attention(
+            query_layer=global_query, 
+            key_layer=key_layer, 
+            value_layer=value_layer, 
+            attention_mask=attention_mask
+            )
+        
+        # Split K Q M on global and non global
+        global_key, key_layer = key_layer.split(split, dim=-2)
+        global_value, value_layer = value_layer.split(split, dim=-2)
+        global_mask, attention_mask = attention_mask.split(split, dim=-1)
+        
+        n, h, t, d = key_layer.size()
+
+        # Get sparse idx
+        sparse_key, sparse_value, sparse_mask = (None, None, None)
+
+        if self.sparse_block_size and self.sparsity_factor > 0:
+            sparse_key, sparse_value, sparse_mask = self.get_sparse_elements(key_layer, value_layer, attention_mask)
+        
+        # Expand masks on heads
+        attention_mask = attention_mask.expand(-1, h, -1, -1)
+        global_mask = global_mask.expand(-1, h, -1, -1)
+
+        # Compute dot product attention
+        context_layer = self.attention(
+            query_layer, 
+            key_layer, 
+            value_layer, 
+            attention_mask,
+            sparse_key=sparse_key, 
+            sparse_value=sparse_value, 
+            sparse_mask=sparse_mask,
+            global_key=global_key,
+            global_value=global_value,
+            global_mask=global_mask
+            )
+
+        # Merge global and local-sparse tokens
+        context_layer = torch.cat([bos, context_layer], dim=-2)
+        if head_mask is not None:
+            context_layer = context_layer * head_mask[:, :, :1, :1]
+        context_layer = self.reshape_output(context_layer)
+        
+        return context_layer
+
+    def chunk(self, x, chunk_size):
+
+        n, h, t, d = x.size()
+        return x.reshape(n, h, -1, chunk_size, d)
+
+
+class LSGBartEncoderLayer(BartEncoderLayer):
+
+    def __init__(self, config):
+
+        super().__init__(config)
+        self.self_attn = LSGBartEncoderAttention(
+            config=config,
+            embed_dim=self.embed_dim,
+            num_heads=config.encoder_attention_heads,
+            dropout=config.attention_dropout,
+        )
+
+
+class LSGBartDecoderLayer(BartDecoderLayer):
+
+    def __init__(self, config):
+
+        super().__init__(config)
+        
+
+class LSGBartClassificationHead(BartClassificationHead):
+    """Head for sentence-level classification tasks."""
+
+    def __init__(
+        self,
+        input_dim,
+        inner_dim,
+        num_classes,
+        pooler_dropout,
+        ):
+
+        super().__init__(input_dim, inner_dim, num_classes, pooler_dropout)
+
+
+class LSGBartPretrainedModel(BartPretrainedModel):
+
+    config_class = LSGBartConfig
+
+    def _set_gradient_checkpointing(self, module, value=False):
+
+        if isinstance(module, (BartDecoder, BartEncoder, LSGBartDecoder, LSGBartEncoder)):
+            module.gradient_checkpointing = value
+
+
+class PretrainedLSGBartModel(LSGBartPretrainedModel):
+
+    def __init_subclass__(self):
+        warnings.warn(
+            "The class `PretrainedBartModel` has been depreciated, please use `LSGBartPretrainedModel` instead.",
+            FutureWarning,
+        )
+
+
+class LSGBartEncoder(LSGBartPretrainedModel, BartEncoder):
+    """
+    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
+    :class:`BartEncoderLayer`.
+    Args:
+        config: BartConfig
+        embed_tokens (nn.Embedding): output embedding
+    """
+
+    def __init__(self, config, embed_tokens=None):
+
+        super().__init__(config)
+        self.dropout = config.dropout
+        self.layerdrop = config.encoder_layerdrop
+
+        embed_dim = config.d_model
+        self.padding_idx = config.pad_token_id
+        self.max_source_positions = config.max_position_embeddings
+        self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0
+
+        if embed_tokens is not None:
+            self.embed_tokens = embed_tokens
+        else:
+            self.embed_tokens = nn.Embedding(config.vocab_size, embed_dim, self.padding_idx)
+
+        self.embed_positions = BartLearnedPositionalEmbedding(
+            config.max_position_embeddings,
+            embed_dim,
+        )
+        self.layers = nn.ModuleList([LSGBartEncoderLayer(config) for _ in range(config.encoder_layers)])
+        self.layernorm_embedding = nn.LayerNorm(embed_dim)
+
+        # 
+        assert hasattr(config, "num_global_tokens")
+        self.num_global_tokens = config.num_global_tokens
+        self.pad_idx = config.pad_token_id
+
+        assert hasattr(config, "block_size") and hasattr(config, "adaptive")
+        self.block_size = config.block_size
+        self.adaptive = config.adaptive
+        self.mask_first_token = config.mask_first_token
+        self.pool_with_global = config.pool_with_global
+        self.pass_global_tokens_to_decoder = config.pass_global_tokens_to_decoder
+
+        self.global_embeddings = nn.Embedding(512, embedding_dim=config.d_model)
+
+        self.gradient_checkpointing = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def forward(self,
+        input_ids=None,
+        attention_mask=None,
+        head_mask=None,
+        inputs_embeds=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None
+        ):
+
+        
+        inputs_ = input_ids if input_ids is not None else inputs_embeds
+        n, t = inputs_.size()[:2]
+
+        if attention_mask is None:
+            attention_mask = torch.ones(n, t, device=inputs_.device)
+        if self.mask_first_token:
+            attention_mask[:, 0] = 0
+
+        b = self.block_size * 2
+        pad = t % self.block_size
+        
+        # Check if t is multiple of block_size and pad
+        if self.adaptive and t > b and pad > 0:
+            pad_length = self.block_size - pad
+            if input_ids is not None:
+                input_ids = torch.nn.functional.pad(input_ids, (0, pad_length), value=self.pad_idx)
+            else:
+                inputs_embeds = torch.nn.functional.pad(inputs_embeds.transpose(-1, -2), (0, pad_length), value=0.).transpose(-1, -2)
+            attention_mask = torch.nn.functional.pad(attention_mask, (0, pad_length), value=0)
+        
+        n, t_ = attention_mask.size()
+        
+        encoder_outputs = self.forward_with_adaptive(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            )
+        
+        context = encoder_outputs[0]
+        diff = t - t_
+
+        if self.pass_global_tokens_to_decoder:
+            offset = self.num_global_tokens
+        else:
+            if self.pool_with_global:
+                context[:, self.num_global_tokens] = context[:, 0]
+            context = context[..., self.num_global_tokens:, :]
+            offset = 0
+
+        # Adapt sequence to initial shape
+        if diff < 0:
+            context = context[:, :t + offset]
+        
+        if return_dict:
+            encoder_outputs.last_hidden_state = context
+        else:
+            encoder_outputs = (context, ) + encoder_outputs[1:]
+        
+        return encoder_outputs
+
+    def forward_with_adaptive(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        head_mask=None,
+        inputs_embeds=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        ):
+        
+        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
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale
+
+        embed_pos = self.embed_positions(input_shape)
+        hidden_states = inputs_embeds + embed_pos
+
+        # Add global tokens
+        n, t, d = hidden_states.size()
+        global_idx = torch.arange(self.num_global_tokens, device=hidden_states.device).reshape(1, -1)
+        hidden_states = torch.cat([self.global_embeddings(global_idx).expand(n, -1, -1), hidden_states], dim=-2)
+
+        hidden_states = self.layernorm_embedding(hidden_states)
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+
+        # expand attention_mask
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            attention_mask = _expand_mask(attention_mask, inputs_embeds.dtype)
+
+        encoder_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        # check if head_mask has a correct number of layers specified if desired
+        if head_mask is not None:
+            if head_mask.size()[0] != (len(self.layers)):
+                raise ValueError(
+                    f"The head_mask should be specified for {len(self.layers)} layers, but it is for {head_mask.size()[0]}."
+                )
+
+        for idx, encoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            dropout_probability = random.uniform(0, 1)
+            if self.training and (dropout_probability < self.layerdrop):  # skip the layer
+                layer_outputs = (None, None)
+            else:
+                if self.gradient_checkpointing and self.training:
+
+                    def create_custom_forward(module):
+                        def custom_forward(*inputs):
+                            return module(*inputs, output_attentions)
+
+                        return custom_forward
+
+                    layer_outputs = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(encoder_layer),
+                        hidden_states,
+                        attention_mask,
+                        (head_mask[idx] if head_mask is not None else None),
+                    )
+                else:
+                    layer_outputs = encoder_layer(
+                        hidden_states,
+                        attention_mask,
+                        layer_head_mask=(head_mask[idx] if head_mask is not None else None),
+                        output_attentions=output_attentions,
+                    )
+
+                hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            encoder_states = encoder_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
+        )
+
+
+class LSGBartDecoder(BartDecoder, LSGBartPretrainedModel):
+    """
+    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a :class:`LSGBartDecoderLayer`
+    Args:
+        config: BartConfig
+        embed_tokens (nn.Embedding): output embedding
+    """
+
+    def __init__(self, config, embed_tokens=None):
+        
+        LSGBartPretrainedModel.__init__(self, config)
+        
+        self.dropout = config.dropout
+        self.layerdrop = config.decoder_layerdrop
+        self.padding_idx = config.pad_token_id
+        self.max_target_positions = config.max_position_embeddings
+        self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0
+        self.adaptive = config.adaptive
+
+        if embed_tokens is not None:
+            self.embed_tokens = embed_tokens
+        else:
+            self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model, self.padding_idx)
+
+        self.embed_positions = BartLearnedPositionalEmbedding(
+            config.max_position_embeddings,
+            config.d_model,
+        )
+        self.layers = nn.ModuleList([LSGBartDecoderLayer(config) for _ in range(config.decoder_layers)])
+        self.layernorm_embedding = nn.LayerNorm(config.d_model)
+
+        self.gradient_checkpointing = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+
+class LSGBartModel(LSGBartPretrainedModel, BartModel):
+
+    def __init__(self, config):
+
+        LSGBartPretrainedModel.__init__(self, config)
+
+        padding_idx, vocab_size = config.pad_token_id, config.vocab_size
+        self.shared = nn.Embedding(vocab_size, config.d_model, padding_idx)
+
+        self.pass_global_tokens_to_decoder = config.pass_global_tokens_to_decoder
+        self.num_global_tokens = config.num_global_tokens
+
+        self.encoder = LSGBartEncoder(config, self.shared)
+        self.decoder = LSGBartDecoder(config, self.shared)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        decoder_input_ids=None,
+        decoder_attention_mask=None,
+        head_mask=None,
+        decoder_head_mask=None,
+        cross_attn_head_mask=None,
+        encoder_outputs=None,
+        past_key_values=None,
+        inputs_embeds=None,
+        decoder_inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        ):
+
+        # different to other models, Bart automatically creates decoder_input_ids from
+        # input_ids if no decoder_input_ids are provided
+        if decoder_input_ids is None and decoder_inputs_embeds is None:
+            decoder_input_ids = shift_tokens_right(
+                input_ids, self.config.pad_token_id, self.config.decoder_start_token_id
+            )
+
+        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
+
+        if encoder_outputs is None:
+            encoder_outputs = self.encoder(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                head_mask=head_mask,
+                inputs_embeds=inputs_embeds,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+        # If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput when return_dict=True
+        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
+            encoder_outputs = BaseModelOutput(
+                last_hidden_state=encoder_outputs[0],
+                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
+                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
+            )
+
+        # Pad mask for global tokens
+        if self.pass_global_tokens_to_decoder and attention_mask is not None:
+            attention_mask = torch.nn.functional.pad(attention_mask, pad=(self.num_global_tokens, 0), value=1)
+            
+        # decoder outputs consists of (dec_features, past_key_value, dec_hidden, dec_attn)
+        decoder_outputs = self.decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            encoder_hidden_states=encoder_outputs[0],
+            encoder_attention_mask=attention_mask,
+            head_mask=decoder_head_mask,
+            cross_attn_head_mask=cross_attn_head_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        if not return_dict:
+            return decoder_outputs + encoder_outputs
+
+        return Seq2SeqModelOutput(
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            past_key_values=decoder_outputs.past_key_values,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+        )
+
+
+class LSGBartForConditionalGeneration(BartForConditionalGeneration, LSGBartPretrainedModel):
+    
+    base_model_prefix = "model"
+    _keys_to_ignore_on_load_missing = [r"final_logits_bias", r"lm_head\.weight"]
+
+    def __init__(self, config):
+
+        LSGBartPretrainedModel.__init__(self, config)
+        self.model = LSGBartModel(config)
+        self.register_buffer("final_logits_bias", torch.zeros((1, self.model.shared.num_embeddings)))
+        self.lm_head = nn.Linear(config.d_model, self.model.shared.num_embeddings, bias=False)
+    
+        # Initialize weights and apply final processing
+        self.post_init()
+
+
+class LSGBartForSequenceClassification(BartForSequenceClassification, LSGBartPretrainedModel):
+
+    def __init__(self, config: LSGBartConfig, **kwargs):
+
+        LSGBartPretrainedModel.__init__(self, config, **kwargs)
+        self.model = LSGBartModel(config)
+        self.classification_head = LSGBartClassificationHead(
+            config.d_model,
+            config.d_model,
+            config.num_labels,
+            config.classifier_dropout,
+        )
+        self.model._init_weights(self.classification_head.dense)
+        self.model._init_weights(self.classification_head.out_proj)
+
+
+class LSGBartForQuestionAnswering(BartForQuestionAnswering, LSGBartPretrainedModel):
+
+    def __init__(self, config: LSGBartConfig):
+
+        LSGBartPretrainedModel.__init__(self, config)
+
+        config.num_labels = 2
+        self.num_labels = config.num_labels
+
+        self.model = LSGBartModel(config)
+        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.model._init_weights(self.qa_outputs)
+
+
+class LSGBartDecoderWrapper(LSGBartPretrainedModel):
+    """
+    This wrapper class is a helper class to correctly load pretrained checkpoints when the causal language model is
+    used in combination with the :class:`~transformers.EncoderDecoderModel` framework.
+    """
+
+    def __init__(self, config: LSGBartConfig):
+        super().__init__(config)
+        self.decoder = LSGBartDecoder(config)
+
+    def forward(self, *args, **kwargs):
+        return self.decoder(*args, **kwargs)
+
+
+class LSGBartForCausalLM(BartForCausalLM, LSGBartPretrainedModel):
+
+    def __init__(self, config: LSGBartConfig):
+
+        config = copy.deepcopy(config)
+        config.is_decoder = True
+        config.is_encoder_decoder = False
+        LSGBartPretrainedModel.__init__(self, config)
+        self.model = LSGBartDecoderWrapper(config)
+
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+
+def str_to_class(classname):
+    return getattr(sys.modules[__name__], classname)
+
+# Register model in Auto API
+try:
+    LSGBartConfig.register_for_auto_class()
+    for key, value in AUTO_MAP.items():
+        str_to_class(value.split(".")[-1]).register_for_auto_class(key)
+except:
+    warn("AutoRegister isn't available, you'll have to manually copy modeling.py after .save_pretrained(...).")
+    warn("Update to transformers >= 4.17.0 to fix.")

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:414b98e7d1f31950bf2737cfa78f72e63340ef5ac0be251272d4d9152e9e7f60
+size 653914167

special_tokens_map.json

@@ -0,0 +1 @@
+{"bos_token": "<s>", "eos_token": "</s>", "unk_token": "<unk>", "sep_token": "</s>", "pad_token": "<pad>", "cls_token": "<s>", "mask_token": {"content": "<mask>", "single_word": false, "lstrip": true, "rstrip": false, "normalized": false}}

tokenizer_config.json

@@ -0,0 +1 @@
+{"errors": "replace", "bos_token": "<s>", "eos_token": "</s>", "sep_token": "</s>", "cls_token": "<s>", "unk_token": "<unk>", "pad_token": "<pad>", "mask_token": "<mask>", "add_prefix_space": false, "trim_offsets": true, "model_max_length": 16384, "special_tokens_map_file": null, "name_or_path": "/data/ccondevaux/lsg/text-summarization/tmp_final/mediasum/lsg_local_16384", "tokenizer_class": "BartTokenizer"}

train_results.json

@@ -0,0 +1,8 @@
+{
+    "epoch": 1.0,
+    "train_loss": 1.5347596183515975,
+    "train_runtime": 69651.7637,
+    "train_samples": 443596,
+    "train_samples_per_second": 6.369,
+    "train_steps_per_second": 0.199
+}

trainer_state.json

@@ -0,0 +1,187 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.9999729483584162,
+  "global_step": 13862,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.04,
+      "learning_rate": 2.8839221341023793e-05,
+      "loss": 1.4339,
+      "step": 500
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 5.767844268204759e-05,
+      "loss": 1.4587,
+      "step": 1000
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 7.927535070140282e-05,
+      "loss": 1.5324,
+      "step": 1500
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 7.606893787575151e-05,
+      "loss": 1.5695,
+      "step": 2000
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 7.286252505010021e-05,
+      "loss": 1.5931,
+      "step": 2500
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 6.96561122244489e-05,
+      "loss": 1.5994,
+      "step": 3000
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 6.64496993987976e-05,
+      "loss": 1.6028,
+      "step": 3500
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 6.32432865731463e-05,
+      "loss": 1.5965,
+      "step": 4000
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 6.0036873747494996e-05,
+      "loss": 1.6002,
+      "step": 4500
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 5.683046092184369e-05,
+      "loss": 1.6017,
+      "step": 5000
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 5.362404809619239e-05,
+      "loss": 1.5735,
+      "step": 5500
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 5.041763527054109e-05,
+      "loss": 1.5765,
+      "step": 6000
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 4.7211222444889784e-05,
+      "loss": 1.5713,
+      "step": 6500
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 4.400480961923849e-05,
+      "loss": 1.5619,
+      "step": 7000
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 4.0798396793587175e-05,
+      "loss": 1.5509,
+      "step": 7500
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 3.759198396793588e-05,
+      "loss": 1.5421,
+      "step": 8000
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 3.438557114228457e-05,
+      "loss": 1.5299,
+      "step": 8500
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 3.117915831663327e-05,
+      "loss": 1.5285,
+      "step": 9000
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 2.7972745490981967e-05,
+      "loss": 1.5326,
+      "step": 9500
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 2.4766332665330663e-05,
+      "loss": 1.5119,
+      "step": 10000
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 2.1559919839679358e-05,
+      "loss": 1.5147,
+      "step": 10500
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 1.8353507014028057e-05,
+      "loss": 1.4978,
+      "step": 11000
+    },
+    {
+      "epoch": 0.83,
+      "learning_rate": 1.5147094188376754e-05,
+      "loss": 1.4914,
+      "step": 11500
+    },
+    {
+      "epoch": 0.87,
+      "learning_rate": 1.1940681362725453e-05,
+      "loss": 1.4889,
+      "step": 12000
+    },
+    {
+      "epoch": 0.9,
+      "learning_rate": 8.734268537074148e-06,
+      "loss": 1.4894,
+      "step": 12500
+    },
+    {
+      "epoch": 0.94,
+      "learning_rate": 5.527855711422846e-06,
+      "loss": 1.4743,
+      "step": 13000
+    },
+    {
+      "epoch": 0.97,
+      "learning_rate": 2.321442885771543e-06,
+      "loss": 1.4627,
+      "step": 13500
+    },
+    {
+      "epoch": 1.0,
+      "step": 13862,
+      "total_flos": 7.817060132659814e+17,
+      "train_loss": 1.5347596183515975,
+      "train_runtime": 69651.7637,
+      "train_samples_per_second": 6.369,
+      "train_steps_per_second": 0.199
+    }
+  ],
+  "max_steps": 13862,
+  "num_train_epochs": 1,
+  "total_flos": 7.817060132659814e+17,
+  "trial_name": null,
+  "trial_params": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:23b05dd0e9ec7f982e19087b8e7878149c1ef65c9ba808ba9b25866700cfc6ed
+size 3439