AutoAndroidController/py/venv/Lib/site-packages/transformers/models/falcon/modeling_falcon.py

# 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, Union

import torch
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, LayerNorm, MSELoss
from torch.nn import functional as F

from ...activations import get_activation
from ...cache_utils import Cache, DynamicCache, StaticCache
from ...generation import GenerationMixin
from ...modeling_attn_mask_utils import (
    AttentionMaskConverter,
)
from ...modeling_flash_attention_utils import flash_attn_supports_top_left_mask, is_flash_attn_available
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import (
    BaseModelOutputWithPastAndCrossAttentions,
    CausalLMOutputWithCrossAttentions,
    QuestionAnsweringModelOutput,
    SequenceClassifierOutputWithPast,
    TokenClassifierOutput,
)
from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
from ...modeling_utils import PreTrainedModel
from ...utils import (
    auto_docstring,
    logging,
)
from .configuration_falcon import FalconConfig


if is_flash_attn_available():
    from ...modeling_flash_attention_utils import _flash_attention_forward

logger = logging.get_logger(__name__)


# 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


# Copied from transformers.models.llama.modeling_llama.rotate_half
def rotate_half(x):
    """Rotates half the hidden dims of the input."""
    x1 = x[..., : x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2 :]
    return torch.cat((-x2, x1), dim=-1)


# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
    """Applies Rotary Position Embedding to the query and key tensors.

    Args:
        q (`torch.Tensor`): The query tensor.
        k (`torch.Tensor`): The key tensor.
        cos (`torch.Tensor`): The cosine part of the rotary embedding.
        sin (`torch.Tensor`): The sine part of the rotary embedding.
        position_ids (`torch.Tensor`, *optional*):
            Deprecated and unused.
        unsqueeze_dim (`int`, *optional*, defaults to 1):
            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
    Returns:
        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
    """
    cos = cos.unsqueeze(unsqueeze_dim)
    sin = sin.unsqueeze(unsqueeze_dim)
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed


# Copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding with Llama->Falcon
class FalconRotaryEmbedding(nn.Module):
    inv_freq: torch.Tensor  # fix linting for `register_buffer`

    def __init__(self, config: FalconConfig, device=None):
        super().__init__()
        # BC: "rope_type" was originally "type"
        if hasattr(config, "rope_scaling") and isinstance(config.rope_scaling, dict):
            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
        else:
            self.rope_type = "default"
        self.max_seq_len_cached = config.max_position_embeddings
        self.original_max_seq_len = config.max_position_embeddings

        self.config = config
        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]

        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self.original_inv_freq = self.inv_freq

    @torch.no_grad()
    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
    def forward(self, x, position_ids):
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
        position_ids_expanded = position_ids[:, None, :].float()

        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos() * self.attention_scaling
            sin = emb.sin() * self.attention_scaling

        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)


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`):
            input tensor
        residual (`torch.tensor`):
            residual tensor
        prob (`float`):
            dropout probability
        training (`bool`):
            training mode
    """
    out = F.dropout(x, p=prob, training=training)
    out = residual + out
    return out


class FalconAttention(nn.Module):
    def __init__(self, config: FalconConfig, layer_idx=None):
        super().__init__()

        self.config = config
        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
        self.max_position_embeddings = config.max_position_embeddings
        self.rope_theta = config.rope_theta
        self.is_causal = True
        self.layer_idx = layer_idx
        if layer_idx is None:
            logger.warning_once(
                f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
                "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
                "when creating this class."
            )

        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})."
            )

        # 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`): [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 dimension

        Args:
            x (`torch.tensor`): [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,
        position_ids: Optional[torch.LongTensor] = None,
        layer_past: Optional[Cache] = None,
        head_mask: Optional[torch.Tensor] = None,
        use_cache: bool = False,
        output_attentions: bool = False,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
    ):
        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)

        if alibi is None:
            cos, sin = position_embeddings
            query_layer, key_layer = apply_rotary_pos_emb(query_layer, key_layer, cos, sin)

        if layer_past is not None:
            cache_kwargs = {"cache_position": cache_position}
            if alibi is None:
                cache_kwargs.update({"sin": sin, "cos": cos})
            key_layer, value_layer = layer_past.update(key_layer, value_layer, self.layer_idx, cache_kwargs)

        kv_length = key_layer.shape[-2]
        if (
            self.config._attn_implementation == "sdpa"
            and query_layer.device.type == "cuda"
            and attention_mask is not None
        ):
            # For torch<=2.1.2, SDPA with memory-efficient backend is bugged with non-contiguous inputs with custom attn_mask,
            # Reference: https://github.com/pytorch/pytorch/issues/112577.
            query_layer = query_layer.contiguous()
            key_layer = key_layer.contiguous()
            value_layer = value_layer.contiguous()

        if attention_mask is not None:
            attention_mask = attention_mask[:, :, :, : key_layer.shape[-2]]

        if alibi is None:
            if self.config._attn_implementation == "sdpa" and not output_attentions:
                # We dispatch to SDPA's Flash Attention or Efficient kernels via this if statement instead of an
                # inline conditional assignment to support both torch.compile's `dynamic=True` and `fullgraph=True`
                # The query_length > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not
                # create a causal mask in case query_length == 1.
                is_causal = self.is_causal and attention_mask is None and query_length > 1
                attn_output = torch.nn.functional.scaled_dot_product_attention(
                    query_layer,
                    key_layer,
                    value_layer,
                    attn_mask=attention_mask,
                    dropout_p=0.0,
                    is_causal=is_causal,
                )
                attention_scores = None
            else:
                attention_scores = query_layer @ key_layer.transpose(-1, -2)
                attention_scores /= math.sqrt(self.head_dim)

                attention_scores = F.softmax(attention_scores + attention_mask, dim=-1, dtype=hidden_states.dtype)
                # It is unclear why neither dropout nor head_mask is applied here (while it is with alibi).
                attn_output = attention_scores @ value_layer

            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)

            attn_output = self.dense(attn_output)

            return attn_output, attention_scores

        else:
            if self.config._attn_implementation == "sdpa" and not output_attentions and head_mask is None:
                # We dispatch to SDPA's Flash Attention or Efficient kernels via this if statement instead of an
                # inline conditional assignment to support both torch.compile's `dynamic=True` and `fullgraph=True`
                is_causal = self.is_causal and attention_mask is None and query_length > 1
                attn_output = torch.nn.functional.scaled_dot_product_attention(
                    query_layer,
                    key_layer,
                    value_layer,
                    attn_mask=attention_mask,
                    dropout_p=self.attention_dropout.p if self.training else 0.0,
                    is_causal=is_causal,
                )
                attention_probs = None
                attn_output = attn_output.transpose(1, 2)
                attn_output = attn_output.reshape(batch_size, query_length, self.num_heads * self.head_dim)

                attn_output = self.dense(attn_output)
            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)

                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, 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]
                attn_output = (attention_probs_reshaped @ value_layer).flatten(0, 1)

                # change view [batch_size, q_length, num_heads * head_dim]
                attn_output = self._merge_heads(attn_output)

                attn_output = self.dense(attn_output)

            return attn_output, attention_probs


class FalconFlashAttention2(FalconAttention):
    """
    Falcon flash attention module. This module inherits from `FalconAttention` as the weights of the module stays
    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
    flash attention and deal with padding tokens in case the input contains any of them.
    """

    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignment, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
        self._flash_attn_uses_top_left_mask = flash_attn_supports_top_left_mask()

    def forward(
        self,
        hidden_states: torch.Tensor,
        alibi: Optional[torch.Tensor],
        attention_mask: torch.Tensor,
        position_ids: Optional[torch.LongTensor] = None,
        layer_past: Optional[Cache] = None,
        head_mask: Optional[torch.Tensor] = None,
        use_cache: bool = False,
        output_attentions: bool = False,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
    ):
        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)

        if alibi is None:
            cos, sin = position_embeddings
            query_layer, key_layer = apply_rotary_pos_emb(query_layer, key_layer, cos, sin)

        if layer_past is not None:
            cache_kwargs = {"cache_position": cache_position}
            if alibi is None:
                cache_kwargs.update({"sin": sin, "cos": cos})
            key_layer, value_layer = layer_past.update(key_layer, value_layer, self.layer_idx, cache_kwargs)

        # TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache
        # to be able to avoid many of these transpose/reshape/view.
        query_layer = query_layer.transpose(1, 2)
        key_layer = key_layer.transpose(1, 2)
        value_layer = value_layer.transpose(1, 2)

        if alibi is not None:
            raise ValueError("`alibi` is not supported when `use_flash_attn` is True")

        attn_dropout = self.config.attention_dropout if self.training else 0.0

        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
        # therefore the input hidden states gets silently casted in float32. Hence, we need
        # cast them back in float16 just to be sure everything works as expected.
        input_dtype = query_layer.dtype
        device_type = query_layer.device.type if query_layer.device.type != "mps" else "cpu"
        if input_dtype == torch.float32:
            if torch.is_autocast_enabled():
                target_dtype = (
                    torch.get_autocast_dtype(device_type)
                    if hasattr(torch, "get_autocast_dtype")
                    else torch.get_autocast_gpu_dtype()
                )
            # Handle the case where the model is quantized
            elif hasattr(self.config, "_pre_quantization_dtype"):
                target_dtype = self.config._pre_quantization_dtype
            else:
                target_dtype = self.query_key_value.weight.dtype

            logger.warning_once(
                f"The input hidden states seems to be silently casted in float32, this might be related to"
                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
                f" {target_dtype}."
            )

            query_layer = query_layer.to(target_dtype)
            key_layer = key_layer.to(target_dtype)
            value_layer = value_layer.to(target_dtype)

        attn_output = _flash_attention_forward(
            query_layer,
            key_layer,
            value_layer,
            attention_mask,
            query_length,
            position_ids=position_ids,
            dropout=attn_dropout,
            is_causal=self.is_causal,
            use_top_left_mask=self._flash_attn_uses_top_left_mask,
        )

        attn_weights = attn_output.reshape(batch_size, query_length, self.num_heads * self.head_dim)
        attn_output = self.dense(attn_weights)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights


class FalconMLP(nn.Module):
    def __init__(self, config: FalconConfig):
        super().__init__()
        hidden_size = config.hidden_size

        self.dense_h_to_4h = FalconLinear(hidden_size, config.ffn_hidden_size, bias=config.bias)
        self.act = get_activation(config.activation)
        self.dense_4h_to_h = FalconLinear(config.ffn_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


FALCON_ATTENTION_CLASSES = {
    "eager": FalconAttention,
    "sdpa": FalconAttention,  # FalconAttention originally implemented both a forward with & without SDPA
    "flash_attention_2": FalconFlashAttention2,
}


class FalconDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: FalconConfig, layer_idx=None):
        super().__init__()
        hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads

        self.self_attention = FALCON_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)
        self.mlp = FalconMLP(config)
        self.hidden_dropout = config.hidden_dropout
        self.config = config

        if config.num_ln_in_parallel_attn is None and config.new_decoder_architecture:
            config.num_ln_in_parallel_attn = 2

        if not config.parallel_attn:
            self.post_attention_layernorm = LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
            self.input_layernorm = LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
        else:
            if config.num_ln_in_parallel_attn == 2:
                # 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)

    def forward(
        self,
        hidden_states: torch.Tensor,
        alibi: Optional[torch.Tensor],
        attention_mask: torch.Tensor,
        position_ids: Optional[torch.LongTensor] = None,
        layer_past: Optional[Union[Cache, tuple[torch.Tensor, torch.Tensor]]] = None,
        head_mask: Optional[torch.Tensor] = None,
        use_cache: bool = False,
        output_attentions: bool = False,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
        **kwargs,
    ):
        residual = hidden_states

        if self.config.new_decoder_architecture and self.config.num_ln_in_parallel_attn == 2:
            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.
        attention_output, attn_weights = self.self_attention(
            attention_layernorm_out,
            layer_past=layer_past,
            attention_mask=attention_mask,
            position_ids=position_ids,
            alibi=alibi,
            head_mask=head_mask,
            use_cache=use_cache,
            output_attentions=output_attentions,
            cache_position=cache_position,
            position_embeddings=position_embeddings,
        )

        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)

        if (
            self.config.new_decoder_architecture
            and self.config.parallel_attn
            and self.config.num_ln_in_parallel_attn == 1
        ):
            mlp_layernorm_out = attention_layernorm_out

        # 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)

        return output, attn_weights


@auto_docstring
class FalconPreTrainedModel(PreTrainedModel):
    config: FalconConfig
    base_model_prefix = "transformer"
    supports_gradient_checkpointing = True
    _no_split_modules = ["FalconDecoderLayer"]
    _supports_flash_attn = True
    _supports_sdpa = True

    _can_compile_fullgraph = True

    def __init__(self, *inputs, **kwargs):
        super().__init__(*inputs, **kwargs)

    def _init_weights(self, module: nn.Module):
        """Initialize the weights."""
        if isinstance(module, (nn.Linear, 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)

    # Adapted from transformers.modeling_utils.PreTrainedModel._check_and_enable_sdpa
    @classmethod
    def _check_and_enable_sdpa(cls, config, hard_check_only: bool = False):
        _is_bettertransformer = getattr(cls, "use_bettertransformer", False)
        if _is_bettertransformer:
            return config

        if not hard_check_only:
            config._attn_implementation = "sdpa"
        return config


@auto_docstring
class FalconModel(FalconPreTrainedModel):
    def __init__(self, config: FalconConfig):
        super().__init__(config)

        self.embed_dim = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.use_alibi = config.alibi

        # Embedding + LN Embedding
        self.word_embeddings = nn.Embedding(config.vocab_size, self.embed_dim)

        # Transformer blocks
        self.h = nn.ModuleList([FalconDecoderLayer(config, layer_idx=i) for i in range(config.num_hidden_layers)])

        # Final Layer Norm
        self.ln_f = LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)

        self.rotary_emb = FalconRotaryEmbedding(config=config)

        self.gradient_checkpointing = False

        # Initialize weights and apply final processing
        self.post_init()

    def get_input_embeddings(self):
        return self.word_embeddings

    def set_input_embeddings(self, new_embeddings: torch.Tensor):
        self.word_embeddings = new_embeddings

    @auto_docstring
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Union[Cache, tuple[tuple[torch.Tensor, torch.Tensor], ...]]] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.LongTensor] = None,
        inputs_embeds: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Union[tuple[torch.Tensor, ...], BaseModelOutputWithPastAndCrossAttentions]:
        r"""
        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.get_seq_length()`
            (`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)
        """
        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 (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        if self.gradient_checkpointing and self.training:
            if use_cache:
                logger.warning_once(
                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                )
                use_cache = False

        if inputs_embeds is None:
            inputs_embeds = self.word_embeddings(input_ids)

        if use_cache and past_key_values is None:
            past_key_values = DynamicCache(config=self.config)

        # Compute alibi tensor: check build_alibi_tensor documentation
        alibi = None
        past_key_values_length = past_key_values.get_seq_length() if past_key_values is not None else 0
        batch_size, seq_length, _ = inputs_embeds.shape
        if self.use_alibi:
            mask = (
                torch.ones(
                    (batch_size, seq_length + past_key_values_length), device=inputs_embeds.device, dtype=torch.long
                )
                if attention_mask is None
                else attention_mask
            )
            alibi = build_alibi_tensor(mask, self.num_heads, dtype=inputs_embeds.dtype)

        if cache_position is None:
            cache_position = torch.arange(
                past_key_values_length, past_key_values_length + seq_length, device=inputs_embeds.device
            )

        if position_ids is None:
            position_ids = cache_position.unsqueeze(0)

        causal_mask = self._update_causal_mask(
            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions, head_mask, alibi
        )

        # Prepare head mask if needed
        # 1.0 in head_mask indicate we keep the head
        # attention_probs has shape batch_size x num_heads x N x N
        # head_mask has shape n_layer x batch x num_heads x N x N
        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
        hidden_states = inputs_embeds

        # create position embeddings to be shared across the decoder layers
        position_embeddings = self.rotary_emb(hidden_states, position_ids)

        all_self_attentions = () if output_attentions else None
        all_hidden_states = () if output_hidden_states else None

        for i, block in enumerate(self.h):
            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

            outputs = block(
                hidden_states,
                layer_past=past_key_values,
                attention_mask=causal_mask,
                position_ids=position_ids,
                head_mask=head_mask[i],
                use_cache=use_cache,
                output_attentions=output_attentions,
                alibi=alibi,
                cache_position=cache_position,
                position_embeddings=position_embeddings,
            )

            hidden_states = outputs[0]
            if output_attentions:
                all_self_attentions = all_self_attentions + (outputs[1],)

        # Add last hidden state
        hidden_states = self.ln_f(hidden_states)

        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        if not return_dict:
            return tuple(
                v for v in [hidden_states, past_key_values, all_hidden_states, all_self_attentions] if v is not None
            )

        return BaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
            hidden_states=all_hidden_states,
            attentions=all_self_attentions,
        )

    def _update_causal_mask(
        self,
        attention_mask: torch.Tensor,
        input_tensor: torch.Tensor,
        cache_position: torch.Tensor,
        past_key_values: Cache,
        output_attentions: bool,
        head_mask: torch.Tensor,
        alibi: torch.Tensor,
    ):
        # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static
        # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes.
        # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using
        # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114

        if self.config._attn_implementation == "flash_attention_2":
            if attention_mask is not None and 0.0 in attention_mask:
                return attention_mask
            return None

        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
        # to infer the attention mask.
        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
        using_static_cache = isinstance(past_key_values, StaticCache)

        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
        if (
            self.config._attn_implementation == "sdpa"
            and not using_static_cache
            and not output_attentions
            and head_mask is None
            and alibi is None
        ):
            if AttentionMaskConverter._ignore_causal_mask_sdpa(
                attention_mask,
                inputs_embeds=input_tensor,
                past_key_values_length=past_seen_tokens,
                is_training=self.training,
            ):
                return None

        dtype, device = input_tensor.dtype, input_tensor.device
        min_dtype = torch.finfo(dtype).min
        batch_size, sequence_length, _ = input_tensor.shape
        if using_static_cache:
            target_length = past_key_values.get_max_cache_shape()
        else:
            target_length = (
                attention_mask.shape[-1]
                if isinstance(attention_mask, torch.Tensor)
                else past_seen_tokens + sequence_length
            )

        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
            attention_mask,
            sequence_length=sequence_length,
            target_length=target_length,
            dtype=dtype,
            device=device,
            cache_position=cache_position,
            batch_size=input_tensor.shape[0],
        )

        # We take care to integrate alibi bias in the causal_mask here
        if head_mask is None and alibi is not None:
            alibi = alibi.reshape(batch_size, -1, *alibi.shape[1:])
            causal_mask = torch.masked_fill(
                alibi / math.sqrt(self.config.hidden_size // self.num_heads),
                causal_mask < -1,
                min_dtype,
            )

        if (
            self.config._attn_implementation == "sdpa"
            and attention_mask is not None
            and attention_mask.device.type in ["cuda", "xpu", "npu"]
            and not output_attentions
        ):
            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
            # Details: https://github.com/pytorch/pytorch/issues/110213
            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)

        return causal_mask

    @staticmethod
    # Copied from transformers.models.gptj.modeling_gptj.GPTJModel._prepare_4d_causal_attention_mask_with_cache_position
    def _prepare_4d_causal_attention_mask_with_cache_position(
        attention_mask: torch.Tensor,
        sequence_length: int,
        target_length: int,
        dtype: torch.dtype,
        cache_position: torch.Tensor,
        batch_size: int,
        **kwargs,
    ):
        """
        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.

        Args:
            attention_mask (`torch.Tensor`):
                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape
                `(batch_size, 1, query_length, key_value_length)`.
            sequence_length (`int`):
                The sequence length being processed.
            target_length (`int`):
                The target length: when generating with static cache, the mask should be as long as the static cache,
                to account for the 0 padding, the part of the cache that is not filled yet.
            dtype (`torch.dtype`):
                The dtype to use for the 4D attention mask.
            cache_position (`torch.Tensor`):
                Indices depicting the position of the input sequence tokens in the sequence.
            batch_size (`torch.Tensor`):
                Batch size.
        """
        if attention_mask is not None and attention_mask.dim() == 4:
            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
            causal_mask = attention_mask
        else:
            min_dtype = torch.finfo(dtype).min
            causal_mask = torch.full(
                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=cache_position.device
            )
            if sequence_length != 1:
                causal_mask = torch.triu(causal_mask, diagonal=1)
            causal_mask *= torch.arange(target_length, device=cache_position.device) > cache_position.reshape(-1, 1)
            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
            if attention_mask is not None:
                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
                mask_length = attention_mask.shape[-1]
                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(
                    causal_mask.device
                )
                padding_mask = padding_mask == 0
                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
                    padding_mask, min_dtype
                )

        return causal_mask


@auto_docstring(
    custom_intro="""
    The Falcon Model transformer with a language modeling head on top (linear layer with weights tied to the input embeddings).
    """
)
class FalconForCausalLM(FalconPreTrainedModel, GenerationMixin):
    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config: FalconConfig):
        super().__init__(config)
        self.transformer = FalconModel(config)
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

        # Initialize weights and apply final processing
        self.post_init()

    def set_output_embeddings(self, new_embeddings: torch.Tensor):
        self.lm_head = new_embeddings

    @auto_docstring
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Union[Cache, tuple[tuple[torch.Tensor, torch.Tensor], ...]]] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        labels: Optional[torch.Tensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        logits_to_keep: Union[int, torch.Tensor] = 0,
        **kwargs,
    ) -> Union[tuple[torch.Tensor], CausalLMOutputWithCrossAttentions]:
        r"""
        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.get_seq_length()`
            (`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)
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
            `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
            are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
        """

        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        transformer_outputs = self.transformer(
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            cache_position=cache_position,
        )
        hidden_states = transformer_outputs[0]

        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
        lm_logits = self.lm_head(hidden_states[:, slice_indices, :])

        loss = None
        if labels is not None:
            loss = self.loss_function(
                lm_logits,
                labels,
                vocab_size=self.config.vocab_size,
                **kwargs,
            )

        if not return_dict:
            output = (lm_logits,) + transformer_outputs[1:]
            return ((loss,) + output) if loss is not None else output

        return CausalLMOutputWithCrossAttentions(
            loss=loss,
            logits=lm_logits,
            past_key_values=transformer_outputs.past_key_values,
            hidden_states=transformer_outputs.hidden_states,
            attentions=transformer_outputs.attentions,
        )


@auto_docstring(
    custom_intro="""
    The Falcon Model transformer with a sequence classification head on top (linear layer).

    [`FalconForSequenceClassification`] uses the last token in order to do the classification, as other causal models
    (e.g. GPT-1) do.

    Since it does classification on the last token, it requires to know the position of the last token. If a
    `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
    no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
    padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
    each row of the batch).
    """
)
class FalconForSequenceClassification(FalconPreTrainedModel):
    def __init__(self, config: FalconConfig):
        super().__init__(config)
        self.num_labels = config.num_labels
        self.transformer = FalconModel(config)
        self.score = nn.Linear(config.hidden_size, config.num_labels, bias=False)

        # Initialize weights and apply final processing
        self.post_init()

    @auto_docstring
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        attention_mask: Optional[torch.Tensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        labels: Optional[torch.Tensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[tuple[torch.Tensor], SequenceClassifierOutputWithPast]:
        r"""
        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.get_seq_length()`
            (`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)
        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

        transformer_outputs = self.transformer(
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        hidden_states = transformer_outputs[0]
        logits = self.score(hidden_states)

        if input_ids is not None:
            batch_size = input_ids.shape[0]
        else:
            batch_size = inputs_embeds.shape[0]

        if self.config.pad_token_id is None and batch_size != 1:
            raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
        if self.config.pad_token_id is None:
            last_non_pad_token = -1
        elif input_ids is not None:
            # To handle both left- and right- padding, we take the rightmost token that is not equal to pad_token_id
            non_pad_mask = (input_ids != self.config.pad_token_id).to(logits.device, torch.int32)
            token_indices = torch.arange(input_ids.shape[-1], device=logits.device, dtype=torch.int32)
            last_non_pad_token = (token_indices * non_pad_mask).argmax(-1)
        else:
            last_non_pad_token = -1
            logger.warning_once(
                f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
                "unexpected if using padding tokens in conjunction with `inputs_embeds.`"
            )

        pooled_logits = logits[torch.arange(batch_size, device=logits.device), last_non_pad_token]

        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(pooled_logits.squeeze(), labels.squeeze())
                else:
                    loss = loss_fct(pooled_logits, labels)
            elif self.config.problem_type == "single_label_classification":
                loss_fct = CrossEntropyLoss()
                loss = loss_fct(pooled_logits, labels)
            elif self.config.problem_type == "multi_label_classification":
                loss_fct = BCEWithLogitsLoss()
                loss = loss_fct(pooled_logits, labels)
        if not return_dict:
            output = (pooled_logits,) + transformer_outputs[1:]
            return ((loss,) + output) if loss is not None else output

        return SequenceClassifierOutputWithPast(
            loss=loss,
            logits=pooled_logits,
            past_key_values=transformer_outputs.past_key_values,
            hidden_states=transformer_outputs.hidden_states,
            attentions=transformer_outputs.attentions,
        )


@auto_docstring
class FalconForTokenClassification(FalconPreTrainedModel):
    def __init__(self, config: FalconConfig):
        super().__init__(config)
        self.num_labels = config.num_labels

        self.transformer = FalconModel(config)
        if getattr(config, "classifier_dropout", None) is not None:
            classifier_dropout = config.classifier_dropout
        elif getattr(config, "hidden_dropout", None) is not None:
            classifier_dropout = config.hidden_dropout
        else:
            classifier_dropout = 0.1
        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()

    @auto_docstring
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        attention_mask: Optional[torch.Tensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        labels: Optional[torch.Tensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[tuple[torch.Tensor], TokenClassifierOutput]:
        r"""
        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.get_seq_length()`
            (`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)
        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

        transformer_outputs = self.transformer(
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        hidden_states = transformer_outputs[0]
        hidden_states = self.dropout(hidden_states)
        logits = self.classifier(hidden_states)

        loss = None
        if labels is not None:
            batch_size, seq_length = labels.shape
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(
                logits.view(batch_size * seq_length, self.num_labels), labels.view(batch_size * seq_length)
            )

        if not return_dict:
            output = (logits,) + transformer_outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return TokenClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=transformer_outputs.hidden_states,
            attentions=transformer_outputs.attentions,
        )


@auto_docstring
class FalconForQuestionAnswering(FalconPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.transformer = FalconModel(config)
        self.qa_outputs = nn.Linear(config.hidden_size, 2)

        # Initialize weights and apply final processing
        self.post_init()

    @auto_docstring
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        start_positions: Optional[torch.LongTensor] = None,
        end_positions: Optional[torch.LongTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[tuple, QuestionAnsweringModelOutput]:
        r"""
        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.get_seq_length()`
            (`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)
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        outputs = self.transformer(
            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,
        )

        sequence_output = outputs[0]

        logits = self.qa_outputs(sequence_output)
        start_logits, end_logits = logits.split(1, dim=-1)
        start_logits = start_logits.squeeze(-1).contiguous()
        end_logits = end_logits.squeeze(-1).contiguous()

        total_loss = None
        if start_positions is not None and end_positions is not None:
            # If we are on multi-GPU, split add a dimension
            if len(start_positions.size()) > 1:
                start_positions = start_positions.squeeze(-1)
            if len(end_positions.size()) > 1:
                end_positions = end_positions.squeeze(-1)
            # sometimes the start/end positions are outside our model inputs, we ignore these terms
            ignored_index = start_logits.size(1)
            start_positions = start_positions.clamp(0, ignored_index)
            end_positions = end_positions.clamp(0, ignored_index)

            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
            start_loss = loss_fct(start_logits, start_positions)
            end_loss = loss_fct(end_logits, end_positions)
            total_loss = (start_loss + end_loss) / 2

        if not return_dict:
            output = (start_logits, end_logits) + outputs[2:]
            return ((total_loss,) + output) if total_loss is not None else output

        return QuestionAnsweringModelOutput(
            loss=total_loss,
            start_logits=start_logits,
            end_logits=end_logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


__all__ = [
    "FalconForCausalLM",
    "FalconModel",
    "FalconPreTrainedModel",
    "FalconForSequenceClassification",
    "FalconForTokenClassification",
    "FalconForQuestionAnswering",
]