Create modeling_falcon.py

modeling_falcon.py

@@ -0,0 +1,624 @@
+# coding=utf-8
+# Copyright 2023 the Falcon authors and HuggingFace Inc. team.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch Falcon model."""
+
+import math
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, LayerNorm, MSELoss
+from torch.nn import functional as F
+
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutputWithPast,
+    TokenClassifierOutput,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward, logging
+from .configuration_falcon import FalconConfig
+
+
+logger = logging.get_logger(__name__)
+
+FALCON_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "tiiuae/falcon-40b",
+    "tiiuae/falcon-40b-instruct",
+    "tiiuae/falcon-7b",
+    "tiiuae/falcon-7b-instruct",
+    "tiiuae/falcon-rw-7b",
+    "tiiuae/falcon-rw-1b",
+]
+_CHECKPOINT_FOR_DOC = "Rocketknight1/falcon-rw-1b"
+_CONFIG_FOR_DOC = "FalconConfig"
+
+
+# NOTE(Hesslow): Unfortunately we did not fuse matmul and bias during training, this means that there's one additional quantization to bfloat16 between the operations.
+# In order not to degrade the quality of our HF-port, we keep these characteristics in the final model.
+class FalconLinear(nn.Linear):
+    def forward(self, input: torch.Tensor) -> torch.Tensor:
+        hidden_states = input @ self.weight.T
+        if self.bias is None:
+            return hidden_states
+        return hidden_states + self.bias
+
+
+# rotary pos emb helpers (torch.jit.script does not seem to support staticmethod...)
+def rotate_half(x):
+    x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+class FalconRotaryEmbedding(nn.Module):
+    """Implementation of RotaryEmbedding from GPT-NeoX.
+    This implementation is designed to operate on queries and keys that are compatible with `[batch_size,
+    n_heads_per_partition, seq_len, head_dim]` (e.g. MinGPTAttention format).
+    """
+
+    def __init__(self, head_dim: int, base=10000):
+        super().__init__()
+        inv_freq = 1.0 / (base ** (torch.arange(0, head_dim, 2).float() / head_dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.head_dim = head_dim
+        self.seq_len_cached = -1
+        self.cos_cached: torch.Tensor | None = None
+        self.sin_cached: torch.Tensor | None = None
+
+    def cos_sin(self, seq_len: int, past_key_values_length: int, device="cpu", dtype=torch.bfloat16) -> torch.Tensor:
+        total_length = seq_len + past_key_values_length
+        if total_length > self.seq_len_cached:
+            self.seq_len_cached = total_length
+            t = torch.arange(total_length, device=device, dtype=self.inv_freq.dtype)
+            freqs = torch.einsum("i,j->ij", t, self.inv_freq)
+            emb = torch.cat((freqs, freqs), dim=-1).to(device)
+
+            if dtype in [torch.float16, torch.bfloat16]:
+                emb = emb.float()
+
+            self.cos_cached = emb.cos()[None, :, :]
+            self.sin_cached = emb.sin()[None, :, :]
+
+            self.cos_cached = self.cos_cached.type(dtype)
+            self.sin_cached = self.sin_cached.type(dtype)
+
+        return (
+            self.cos_cached[:, past_key_values_length : seq_len + past_key_values_length],
+            self.sin_cached[:, past_key_values_length : seq_len + past_key_values_length],
+        )
+
+    def forward(self, query, key, past_key_values_length=0):
+        batch, seq_len, head_dim = query.shape
+        cos, sin = self.cos_sin(seq_len, past_key_values_length, query.device, query.dtype)
+        return (query * cos) + (rotate_half(query) * sin), (key * cos) + (rotate_half(key) * sin)
+
+
+def _make_causal_mask(
+    input_ids_shape: torch.Size, device: torch.device, past_key_values_length: int
+) -> torch.BoolTensor:
+    """
+    Make causal mask used for self-attention. This mask does not take the existing attention mask into account - it
+    just blocks tokens from attending forwards in the sequence. The output shape will be `[batch_size, 1,
+    target_length, target_length+past_key_values_length]`.
+    """
+    batch_size, target_length = input_ids_shape
+
+    mask = torch.triu(torch.ones((target_length, target_length), dtype=torch.bool, device=device), diagonal=1)
+    # If past_key_values_length is 0 this is an empty tensor and the concatenation is a no-op.
+    # This code style is an unfortunate consequence of getting your TF engineer to port models; doing it this
+    # way avoids a data-dependent conditional, which will help me when I have to port this to XLA later.
+    past_mask = torch.zeros((target_length, past_key_values_length), dtype=torch.bool, device=device)
+    mask = torch.cat([past_mask, mask], dim=-1)
+    expanded_mask = mask[None, None, :, :].expand(batch_size, 1, target_length, target_length + past_key_values_length)
+    return expanded_mask
+
+
+def _expand_mask(mask: torch.Tensor, past_key_values_length: int) -> torch.BoolTensor:
+    """
+    Expands attention_mask from `[batch_size, seq_length]` to `[batch_size, 1, seq_length, seq_length + past_length]`.
+    """
+    batch_size, total_length = mask.shape
+    seq_length = total_length - past_key_values_length if past_key_values_length is not None else total_length
+
+    expanded_mask = ~(mask[:, None, None, :].to(torch.bool))
+    return expanded_mask.expand(batch_size, 1, seq_length, total_length)
+
+
+def build_alibi_tensor(attention_mask: torch.Tensor, num_heads: int, dtype: torch.dtype) -> torch.Tensor:
+    batch_size, seq_length = attention_mask.shape
+    closest_power_of_2 = 2 ** math.floor(math.log2(num_heads))
+    base = torch.tensor(
+        2 ** (-(2 ** -(math.log2(closest_power_of_2) - 3))), device=attention_mask.device, dtype=torch.float32
+    )
+    powers = torch.arange(1, 1 + closest_power_of_2, device=attention_mask.device, dtype=torch.int32)
+    slopes = torch.pow(base, powers)
+
+    if closest_power_of_2 != num_heads:
+        extra_base = torch.tensor(
+            2 ** (-(2 ** -(math.log2(2 * closest_power_of_2) - 3))), device=attention_mask.device, dtype=torch.float32
+        )
+        num_remaining_heads = min(closest_power_of_2, num_heads - closest_power_of_2)
+        extra_powers = torch.arange(1, 1 + 2 * num_remaining_heads, 2, device=attention_mask.device, dtype=torch.int32)
+        slopes = torch.cat([slopes, torch.pow(extra_base, extra_powers)], dim=0)
+
+    # Note: alibi will added to the attention bias that will be applied to the query, key product of attention
+    # => therefore alibi will have to be of shape (batch_size, num_heads, query_length, key_length)
+    # => here we set (batch_size=1, num_heads=num_heads, query_length=1, key_length=max_length)
+    # => the query_length dimension will then be broadcasted correctly
+    # This is more or less identical to T5's relative position bias:
+    # https://github.com/huggingface/transformers/blob/f681437203baa7671de3174b0fa583c349d9d5e1/src/transformers/models/t5/modeling_t5.py#L527
+    arange_tensor = ((attention_mask.cumsum(dim=-1) - 1) * attention_mask)[:, None, :]
+    alibi = slopes[..., None].bfloat16() * arange_tensor
+    return alibi.reshape(batch_size * num_heads, 1, seq_length).to(dtype)
+
+
+# Copied from transformers.models.bloom.modeling_bloom.dropout_add
+def dropout_add(x: torch.Tensor, residual: torch.Tensor, prob: float, training: bool) -> torch.Tensor:
+    """
+    Dropout add function
+    Args:
+        x (`torch.tensor`, *required*):
+            input tensor
+        residual (`torch.tensor`, *required*):
+            residual tensor
+        prob (`float`, *required*):
+            dropout probability
+        training (`bool`, *required*):
+            training mode
+    """
+    out = F.dropout(x, p=prob, training=training)
+    out = residual + out
+    return out
+
+
+class FalconAttention(nn.Module):
+    def __init__(self, config: FalconConfig):
+        super().__init__()
+
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.split_size = self.hidden_size
+        self.hidden_dropout = config.hidden_dropout
+
+        if self.head_dim * self.num_heads != self.hidden_size:
+            raise ValueError(
+                f"`hidden_size` must be divisible by num_heads (got `hidden_size`: {self.hidden_size} and `num_heads`:"
+                f" {self.num_heads})."
+            )
+
+        self.maybe_rotary = FalconRotaryEmbedding(config.head_dim) if config.rotary else lambda q, k, t: (q, k)
+
+        # Layer-wise attention scaling
+        self.inv_norm_factor = 1.0 / math.sqrt(self.head_dim)
+        self.beta = self.inv_norm_factor
+        if config.new_decoder_architecture:
+            qkv_out_dim = (config.num_kv_heads * 2 + config.num_attention_heads) * self.head_dim
+        elif config.multi_query:
+            qkv_out_dim = self.hidden_size + 2 * self.head_dim
+        else:
+            qkv_out_dim = 3 * self.hidden_size
+        self.query_key_value = FalconLinear(self.hidden_size, qkv_out_dim, bias=config.bias)
+        self.new_decoder_architecture = config.new_decoder_architecture
+        self.multi_query = config.multi_query
+        self.dense = FalconLinear(self.hidden_size, self.hidden_size, bias=config.bias)
+        self.attention_dropout = nn.Dropout(config.attention_dropout)
+        self.num_kv_heads = config.num_kv_heads if (self.new_decoder_architecture or not self.multi_query) else 1
+
+    def _split_heads(self, fused_qkv: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Split the last dimension into (num_heads, head_dim), results share same memory storage as `fused_qkv`
+        Args:
+            fused_qkv (`torch.tensor`, *required*): [batch_size, seq_length, num_heads * 3 * head_dim]
+        Returns:
+            query: [batch_size, seq_length, num_heads, head_dim] key: [batch_size, seq_length, num_heads, head_dim]
+            value: [batch_size, seq_length, num_heads, head_dim]
+        """
+        if self.new_decoder_architecture:
+            batch, seq_len, _ = fused_qkv.shape
+            qkv = fused_qkv.view(batch, seq_len, -1, self.num_heads // self.num_kv_heads + 2, self.head_dim)
+            query = qkv[:, :, :, :-2]
+            key = qkv[:, :, :, [-2]]
+            value = qkv[:, :, :, [-1]]
+            key = torch.broadcast_to(key, query.shape)
+            value = torch.broadcast_to(value, query.shape)
+
+            query, key, value = [x.flatten(2, 3) for x in (query, key, value)]
+            return query, key, value
+        elif not self.multi_query:
+            batch_size, seq_length, three_times_hidden_size = fused_qkv.shape
+            fused_qkv = fused_qkv.view(batch_size, seq_length, self.num_heads, 3, self.head_dim)
+            return fused_qkv[..., 0, :], fused_qkv[..., 1, :], fused_qkv[..., 2, :]
+        else:
+            batch_size, seq_length, three_times_hidden_size = fused_qkv.shape
+            fused_qkv = fused_qkv.view(batch_size, seq_length, self.num_heads + 2, self.head_dim)
+            return fused_qkv[..., :-2, :], fused_qkv[..., [-2], :], fused_qkv[..., [-1], :]
+
+    # Copied from transformers.models.bloom.modeling_bloom.BloomAttention._merge_heads
+    def _merge_heads(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Merge heads together over the last dimenstion
+        Args:
+            x (`torch.tensor`, *required*): [batch_size * num_heads, seq_length, head_dim]
+        Returns:
+            torch.tensor: [batch_size, seq_length, num_heads * head_dim]
+        """
+        # What we want to achieve is:
+        # batch_size * num_heads, seq_length, head_dim -> batch_size, seq_length, num_heads * head_dim
+        batch_size_and_num_heads, seq_length, _ = x.shape
+        batch_size = batch_size_and_num_heads // self.num_heads
+
+        # First view to decompose the batch size
+        # batch_size * num_heads, seq_length, head_dim -> batch_size, num_heads, seq_length, head_dim
+        x = x.view(batch_size, self.num_heads, seq_length, self.head_dim)
+
+        # batch_size, num_heads, seq_length, head_dim -> batch_size, seq_length, num_heads, head_dim
+        x = x.permute(0, 2, 1, 3)
+
+        # batch_size, seq_length, num_heads, head_dim -> batch_size, seq_length, num_heads * head_dim
+        return x.reshape(batch_size, seq_length, self.num_heads * self.head_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        alibi: Optional[torch.Tensor],
+        attention_mask: torch.Tensor,
+        layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        use_cache: bool = False,
+        output_attentions: bool = False,
+    ):
+        fused_qkv = self.query_key_value(hidden_states)  # [batch_size, seq_length, 3 x hidden_size]
+        num_kv_heads = self.num_heads if self.new_decoder_architecture else self.num_kv_heads
+        # 3 x [batch_size, seq_length, num_heads, head_dim]
+        (query_layer, key_layer, value_layer) = self._split_heads(fused_qkv)
+
+        batch_size, query_length, _, _ = query_layer.shape
+
+        query_layer = query_layer.transpose(1, 2).reshape(batch_size * self.num_heads, query_length, self.head_dim)
+        key_layer = key_layer.transpose(1, 2).reshape(
+            batch_size * num_kv_heads,
+            query_length,
+            self.head_dim,
+        )
+        value_layer = value_layer.transpose(1, 2).reshape(batch_size * num_kv_heads, query_length, self.head_dim)
+
+        past_kv_length = 0 if layer_past is None else layer_past[0].shape[1]
+        query_layer, key_layer = self.maybe_rotary(query_layer, key_layer, past_kv_length)
+
+        if layer_past is not None:
+            past_key, past_value = layer_past
+            # concatenate along seq_length dimension:
+            #  - key: [batch_size * self.num_heads, kv_length, head_dim]
+            #  - value: [batch_size * self.num_heads, kv_length, head_dim]
+            key_layer = torch.cat((past_key, key_layer), dim=1)
+            value_layer = torch.cat((past_value, value_layer), dim=1)
+
+        _, kv_length, _ = key_layer.shape
+        if use_cache:
+            present = (key_layer, value_layer)
+        else:
+            present = None
+
+        attention_mask_float = (attention_mask * 1.0).masked_fill(attention_mask, float("-1e9")).to(query_layer.dtype)
+
+        query_layer_ = query_layer.reshape(batch_size, self.num_heads, -1, self.head_dim)
+        key_layer_ = key_layer.reshape(batch_size, num_kv_heads, -1, self.head_dim)
+        value_layer_ = value_layer.reshape(batch_size, num_kv_heads, -1, self.head_dim)
+
+        if alibi is None:
+            if output_attentions:
+                # F.scaled_dot_product_attention doesn't return the attention weights, so we have
+                # to do it by hand if we want them
+                attention_scores = query_layer_ @ key_layer_.transpose(-1, -2)
+                attention_scores /= math.sqrt(self.head_dim)
+
+                attention_scores = F.softmax(
+                    attention_scores + attention_mask_float, dim=-1, dtype=hidden_states.dtype
+                )
+                attn_output = attention_scores @ value_layer_
+            else:
+                attn_output = F.scaled_dot_product_attention(
+                    query_layer_, key_layer_, value_layer_, attention_mask_float, 0.0, is_causal=False
+                )
+                attention_scores = None
+
+            attn_output = attn_output.view(batch_size, self.num_heads, query_length, self.head_dim)
+            attn_output = attn_output.permute(0, 2, 1, 3)
+            attn_output = attn_output.reshape(batch_size, query_length, self.num_heads * self.head_dim)
+
+            output_tensor = self.dense(attn_output)
+
+            if output_attentions:
+                return output_tensor, present, attention_scores
+            else:
+                return output_tensor, present
+
+        else:
+            matmul_result = query_layer_ @ key_layer_.transpose(-1, -2)
+
+            # change view to [batch_size, num_heads, q_length, kv_length]
+            attention_scores = matmul_result.view(batch_size, self.num_heads, query_length, kv_length)
+
+            # cast attention scores to fp32, compute scaled softmax and cast back to initial dtype - [batch_size, num_heads, q_length, kv_length]
+            input_dtype = attention_scores.dtype
+            # `float16` has a minimum value of -65504.0, whereas `bfloat16` and `float32` have a minimum value of `-3.4e+38`
+            if input_dtype == torch.float16 or input_dtype == torch.bfloat16:
+                attention_scores = attention_scores.to(torch.float32)
+            # Matt (HF) note: We could possibly use F.scaled_dot_product_attention here too, by
+            # adding (alibi * self.inv_norm_factor) to attention_mask_float. I think this would be mathematically
+            # equivalent and more performant, but there might be a numerical difference. If you're reading this
+            # and you'd like to experiment and maybe file a PR, feel free!
+            attention_logits = attention_scores + alibi.view(batch_size, self.num_heads, 1, -1)
+            attention_logits *= self.inv_norm_factor
+            attention_probs = F.softmax(attention_logits + attention_mask_float, dim=-1, dtype=hidden_states.dtype)
+            # [batch_size, num_heads, q_length, kv_length]
+            attention_probs = self.attention_dropout(attention_probs)
+
+            if head_mask is not None:
+                attention_probs = attention_probs * head_mask
+
+            # change view [batch_size, num_heads, q_length, kv_length]
+            attention_probs_reshaped = attention_probs.view(batch_size, self.num_heads, query_length, kv_length)
+
+            # matmul: [batch_size * num_heads, q_length, head_dim]
+            context_layer = (attention_probs_reshaped @ value_layer_).flatten(0, 1)
+
+            # change view [batch_size, num_heads, q_length, head_dim]
+            context_layer = self._merge_heads(context_layer)
+
+            output_tensor = self.dense(context_layer)
+
+            if output_attentions:
+                return output_tensor, present, attention_probs
+            else:
+                return output_tensor, present
+
+
+class FalconMLP(nn.Module):
+    def __init__(self, config: FalconConfig):
+        super().__init__()
+        hidden_size = config.hidden_size
+
+        self.dense_h_to_4h = FalconLinear(hidden_size, 4 * hidden_size, bias=config.bias)
+        self.act = nn.GELU()
+        self.dense_4h_to_h = FalconLinear(4 * hidden_size, hidden_size, bias=config.bias)
+        self.hidden_dropout = config.hidden_dropout
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.act(self.dense_h_to_4h(x))
+        x = self.dense_4h_to_h(x)
+        return x
+
+
+class FalconDecoderLayer(nn.Module):
+    def __init__(self, config: FalconConfig):
+        super().__init__()
+        hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.self_attention = FalconAttention(config)
+        self.mlp = FalconMLP(config)
+        self.hidden_dropout = config.hidden_dropout
+        self.config = config
+
+        if config.new_decoder_architecture:
+            # The layer norm before self-attention
+            self.ln_attn = LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+            # The layer norm before the MLP
+            self.ln_mlp = LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+        else:
+            self.input_layernorm = LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+            if not config.parallel_attn:
+                self.post_attention_layernorm = LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        alibi: Optional[torch.Tensor],
+        attention_mask: torch.Tensor,
+        layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        use_cache: bool = False,
+        output_attentions: bool = False,
+    ):
+        residual = hidden_states
+
+        if self.config.new_decoder_architecture:
+            attention_layernorm_out = self.ln_attn(hidden_states)
+            mlp_layernorm_out = self.ln_mlp(hidden_states)
+        else:
+            attention_layernorm_out = self.input_layernorm(hidden_states)
+
+        # Self attention.
+        attn_outputs = self.self_attention(
+            attention_layernorm_out,
+            layer_past=layer_past,
+            attention_mask=attention_mask,
+            alibi=alibi,
+            head_mask=head_mask,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+        )
+
+        attention_output = attn_outputs[0]
+
+        if not self.config.new_decoder_architecture:
+            if self.config.parallel_attn:
+                mlp_layernorm_out = attention_layernorm_out
+            else:
+                residual = dropout_add(
+                    attention_output, residual, self.config.attention_dropout, training=self.training
+                )
+                mlp_layernorm_out = self.post_attention_layernorm(residual)
+
+        outputs = attn_outputs[1:]
+
+        # MLP.
+        mlp_output = self.mlp(mlp_layernorm_out)
+
+        if self.config.new_decoder_architecture or self.config.parallel_attn:
+            mlp_output += attention_output
+
+        output = dropout_add(mlp_output, residual, self.config.hidden_dropout, training=self.training)
+
+        if use_cache:
+            outputs = (output,) + outputs
+        else:
+            outputs = (output,) + outputs[1:]
+
+        return outputs  # hidden_states, present, attentions
+
+
+FALCON_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings etc.)
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+    Parameters:
+        config ([`FalconConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+FALCON_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
+            `input_ids_length` = `sequence_length` if `past_key_values` is `None` else `past_key_values[0][0].shape[2]`
+            (`sequence_length` of input past key value states). Indices of input sequence tokens in the vocabulary.
+            If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as
+            `input_ids`.
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+            [What are input IDs?](../glossary#input-ids)
+        past_key_values (`Tuple[Tuple[torch.Tensor]]` of length `config.num_hidden_layers`):
+            Contains precomputed hidden-states (key and values in the attention blocks) as computed by the model (see
+            `past_key_values` output below). Can be used to speed up sequential decoding. The `input_ids` which have
+            their past given to this model should not be passed as `input_ids` as they have already been computed.
+            Each element of `past_key_values` is a tuple (past_key, past_value):
+            - past_key: [batch_size * num_heads, head_dim, kv_length]
+            - past_value: [batch_size * num_heads, kv_length, head_dim]
+        attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+            [What are attention masks?](../glossary#attention-mask)
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+            If `past_key_values` is used, optionally only the last `inputs_embeds` have to be input (see
+            `past_key_values`).
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~file_utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+class FalconPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = FalconConfig
+    base_model_prefix = "transformer"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["FalconDecoderLayer"]
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+    def _init_weights(self, module: nn.Module):
+        """Initialize the weights."""
+        if isinstance(module, nn.Linear) or isinstance(module, FalconLinear):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+    # Copied from transformers.models.bloom.modeling_bloom.BloomPreTrainedModel._set_gradient_checkpointing with BloomModel->FalconModel
+    def _set_gradient_checkpointing(self, module: nn.Module, value: bool = False):
+        if isinstance(module, FalconModel):
+            module.gradient_checkpointing = value
+
+    @staticmethod
+    def _convert_cache_to_standard_format(
+        past_key_value: Tuple[Tuple[torch.Tensor, torch.Tensor]], batch_size: int
+    ) -> Tuple[Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Standardizes the format of the cache so as to match most implementations, i.e. to tuple(tuple([batch_size,
+        num_heads, ...]))
+        """
+        batch_size_times_num_heads, kv_length, head_dim = past_key_value[0][0].shape
+        # [batch_size * self.num_heads, kv_length, head_dim] -> [batch_size, num_heads, kv_length, head_dim]
+        # Note that don't want to use self.num_attention_heads because the number of heads may vary depending
+        # on whether we use multi_query attention.
+        num_heads = batch_size_times_num_heads // batch_size
+        return tuple(
+            (
+                layer_past[0].view(batch_size, num_heads, kv_length, head_dim),
+                layer_past[1].view(batch_size, num_heads, kv_length, head_dim),
+            )
+            for layer_past in past_key_value
+        )
+
+    @staticmethod
+    def _convert_to_rw_cache(
+        past_key_value: Tuple[Tuple[torch.Tensor, torch.Tensor]]
+    ) -> Tuple[Tuple[torch.Tensor, torch.Tensor]]:
+        batch_size, num_heads, kv_length, head_dim = past_key_value[0][0].shape
+        batch_size_times_num_heads = batch_size * num_heads
+        # [batch_size, num_heads, kv_length, head_dim] -> [batch_size * num_heads, kv_length, head_dim]
+        return tuple(
+            (
+                layer_past[0].view(batch_size_times_num_heads, kv_length, head_dim),
+                layer_past[1].view(batch_size_times_num_heads, kv_length, head_dim),
+            )
+            for layer_past in past_key_value
+        )
+
+
+@add_start_docstrings(
+    "The bare Falcon Model transformer outputting raw hidden-states without any specific head on top.",
+    FALCON_START_DOCSTRING,
+)
+class FalconModel(FalconPreTrainedModel):
+    def __init__(self, config: FalconConfig):
+        super().__init__(config)
+