modeling_lightpost.py

# coding=utf-8
# Copyright 2024 Lightpost ApS. All rights reserved.
#
# This code is based on Qwen2, which itself is based on EleutherAI's GPT-NeoX library and the GPT-NeoX and
# OPT implementations in the Transformers library. The code has been modified to support Lightpost's
# architectural differences, including memory attention and other adaptations that distinguish it from the
# original GPT-NeoX and OPT architectures.
#
# 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 Lightpost model."""

import math
from typing import List, Optional, Tuple, Union

import torch
import torch.utils.checkpoint
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss

from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache, SlidingWindowCache, StaticCache
from transformers.generation import GenerationMixin
from transformers.modeling_attn_mask_utils import AttentionMaskConverter
from transformers.modeling_outputs import (
    BaseModelOutputWithPast,
    CausalLMOutputWithPast,
    QuestionAnsweringModelOutput,
    SequenceClassifierOutputWithPast,
    TokenClassifierOutput,
)
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
from transformers.modeling_utils import PreTrainedModel
from transformers.utils import (
    add_code_sample_docstrings,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    is_flash_attn_2_available,
    is_flash_attn_greater_or_equal_2_10,
    logging,
    replace_return_docstrings,
)
from .configuration_lightpost import LightpostConfig


if is_flash_attn_2_available():
    from transformers.modeling_flash_attention_utils import _flash_attention_forward


logger = logging.get_logger(__name__)


_CHECKPOINT_FOR_DOC = "Lightpost/Lightpost2-7B-beta"
_CONFIG_FOR_DOC = "LightpostConfig"

class MemoryAttention(nn.Module):
    def __init__(
        self,
        config: LightpostConfig,
    ):
        super(MemoryAttention, self).__init__()
        self.embed_dim = config.hidden_size
        self.memory_size = config.mem_size
        self.dropout = config.attention_dropout
        self.scaling = self.embed_dim ** -0.5

        self.num_heads = config.num_attention_heads

        self.attn_dropout = nn.Dropout(self.dropout)

        # Define a learnable memory for the value vectors
        self.q_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False)
        self.keys = nn.Parameter(0.01 * torch.randn(self.memory_size, self.embed_dim))
        self.learnable_memory = nn.Parameter(0.01 * torch.randn(self.memory_size, self.embed_dim))

    @staticmethod
    def from_state_dict(state_dict, config):
        """
        Instantiate a MemoryAttention object from a state dictionary.

        Args:
            state_dict (dict): The state dictionary containing the model parameters.
            config (object): Configuration object with attributes like hidden_size and num_attention_heads.

        Returns:
            MemoryAttention: An instance of the MemoryAttention class.
        """
        learnable_memory_size = state_dict["learnable_memory"].shape[0]
        config.mem_size = learnable_memory_size
        mem_attn = MemoryAttention(
            config=config,
        )
        mem_attn.load_state_dict(state_dict)
        return mem_attn

    def forward(
        self,
        inputs,
    ):
        # Assume queries are in (batch, seq, embed) format
        queries = self.q_proj(inputs)

        # Calculate attention to each key in memory
        attn_weights = torch.matmul(queries, self.keys.transpose(0,1)) * self.scaling # (batch, seq, memory_size)

        # Apply softmax
        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(queries.dtype) # (batch, seq, memory_size)
        attn_weights = self.attn_dropout(attn_weights) # (batch, seq, memory_size)

        # Compute attention output
        attn_output = torch.matmul(attn_weights, self.learnable_memory) # (batch, seq, embed_dim)

        return attn_output




    def forward_mh(self, queries):
        """
        Args:
            queries: Tensor of shape (batch_size, seq_length, embed_dim)
        
        Returns:
            attn_output: Tensor of shape (batch_size, seq_length, embed_dim)
        """
        bsz, q_len, _ = queries.shape

        # Reshape queries for multi-head attention
        # From (bsz, q_len, embed_dim) to (bsz, num_heads, q_len, head_dim)
        queries = queries.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)  # (bsz, num_heads, q_len, head_dim)

        # Project keys
        # From (memory_size, embed_dim) to (num_heads, memory_size, head_dim)
        keys = self.k_proj(self.learnable_memory)  # (memory_size, embed_dim)
        keys = keys.view(self.memory_size, self.num_heads, self.head_dim).transpose(0, 1)  # (num_heads, memory_size, head_dim)

        # Compute attention weights
        # queries: (bsz, num_heads, q_len, head_dim)
        # keys: (num_heads, memory_size, head_dim)
        # We need to perform matrix multiplication for each head separately
        # Resulting attn_weights shape: (bsz, num_heads, q_len, memory_size)
        
        # Expand keys to (1, num_heads, head_dim, memory_size) for broadcasting
        keys = keys.unsqueeze(0).transpose(-2, -1)  # (1, num_heads, head_dim, memory_size)
        
        # Perform batched matrix multiplication
        attn_weights = torch.matmul(queries, keys)  # (bsz, num_heads, q_len, memory_size)

        # Apply scaling factor
        attn_weights = attn_weights * self.scaling

        # Apply softmax
        attn_weights = nn.functional.softmax(attn_weights, dim=-1)  # (bsz, num_heads, q_len, memory_size)
        attn_weights = self.attn_dropout(attn_weights)

        # Compute attention output
        # learnable_memory: (memory_size, embed_dim) -> (num_heads, memory_size, head_dim)
        memory = self.learnable_memory.view(self.memory_size, self.num_heads, self.head_dim).transpose(0, 1)  # (num_heads, memory_size, head_dim)
        
        # Expand memory for batched matrix multiplication
        # memory: (num_heads, memory_size, head_dim) -> (1, num_heads, memory_size, head_dim)
        memory = memory.unsqueeze(0)  # (1, num_heads, memory_size, head_dim)
        
        # Compute attention output
        # attn_weights: (bsz, num_heads, q_len, memory_size)
        # memory: (1, num_heads, memory_size, head_dim)
        # Resulting attn_output: (bsz, num_heads, q_len, head_dim)
        attn_output = torch.matmul(attn_weights, memory)  # (bsz, num_heads, q_len, head_dim)

        # Concatenate heads and reshape to (bsz, q_len, embed_dim)
        attn_output = attn_output.transpose(1, 2).contiguous().view(bsz, q_len, self.embed_dim)  # (bsz, q_len, embed_dim)

        return attn_output



# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Lightpost
class LightpostRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        LightpostRMSNorm is equivalent to T5LayerNorm
        """
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states):
        input_dtype = hidden_states.dtype
        hidden_states = hidden_states.to(torch.float32)
        variance = hidden_states.pow(2).mean(-1, keepdim=True)
        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
        return self.weight * hidden_states.to(input_dtype)

    def extra_repr(self):
        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"


# Copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding with Llama->Lightpost
class LightpostRotaryEmbedding(nn.Module):
    def __init__(
        self,
        config: LightpostConfig,
        device=None,
    ):
        super().__init__()

        # Use the config object directly
        if config.rope_scaling is not None:
            self.rope_type = config.rope_scaling.get("rope_type", "default")
        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

    def _dynamic_frequency_update(self, position_ids, device):
        """
        dynamic RoPE layers should recompute `inv_freq` in the following situations:
        1 - growing beyond the cached sequence length (allow scaling)
        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
        """
        seq_len = torch.max(position_ids) + 1
        if seq_len > self.max_seq_len_cached:  # growth
            inv_freq, self.attention_scaling = self.rope_init_fn(
                self.config, device, seq_len=seq_len
            )
            self.register_buffer("inv_freq", inv_freq, persistent=False)  # TODO joao: may break with compilation
            self.max_seq_len_cached = seq_len

        if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len:  # reset
            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
            self.max_seq_len_cached = self.original_max_seq_len

    @torch.no_grad()
    def forward(self, x, position_ids):
        if "dynamic" in self.rope_type:
            self._dynamic_frequency_update(position_ids, device=x.device)

        # Core RoPE block
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
        position_ids_expanded = position_ids[:, None, :].float()
        # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
        device_type = x.device.type
        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
        with torch.autocast(device_type=device_type, enabled=False):
            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos()
            sin = emb.sin()

        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
        cos = cos * self.attention_scaling
        sin = sin * self.attention_scaling

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


# 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 dimension along which to unsqueeze the rotary position embeddings (cos and sin) for proper broadcasting.
            If q and k have shape [batch_size, heads, seq_len, head_dim], use unsqueeze_dim=1 to insert a dimension
            after batch_size. If q and k have shape [batch_size, seq_len, heads, head_dim], use unsqueeze_dim=2 to
            insert a dimension after seq_len. This ensures the rotary embeddings can be properly broadcast to match
            the query and key tensor shapes.
    Returns:
        `tuple(torch.Tensor)` with 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.mistral.modeling_mistral.MistralMLP with Mistral->Lightpost
class LightpostMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, hidden_state):
        return self.down_proj(self.act_fn(self.gate_proj(hidden_state)) * self.up_proj(hidden_state))


# Copied from transformers.models.llama.modeling_llama.repeat_kv
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    """
    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
    """
    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
    if n_rep == 1:
        return hidden_states
    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)


class LightpostAttention(nn.Module):
    """
    Multi-headed attention from 'Attention Is All You Need' paper. For long sequences, this implementation uses 
    sliding window attention similar to Longformer and Sparse Transformers, where each token attends to a local window of 
    surrounding tokens rather than the full sequence. This allows for efficient processing of very long sequences while 
    maintaining the key benefits of self-attention within each window.
    """

    def __init__(self, config: LightpostConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx

        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.hidden_size // self.num_heads
        self.num_key_value_heads = config.num_key_value_heads
        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
        self.max_position_embeddings = config.max_position_embeddings
        self.rope_theta = config.rope_theta
        self.is_causal = True
        self.attention_dropout = config.attention_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}"
                f" and `num_heads`: {self.num_heads})."
            )
        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=True)
        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)


        if self.config.mem_layers is not None and self.layer_idx in self.config.mem_layers:
            self.mem_attn = MemoryAttention(config=self.config)
        else:
            self.mem_attn = None

        self.rotary_emb = LightpostRotaryEmbedding(config=self.config)

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]],
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Cache] = None,
        output_attentions: bool = False,
        use_cache: bool = False,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        bsz, q_len, _ = hidden_states.size()

        query_states = self.q_proj(hidden_states)
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)

        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)

        cos, sin = position_embeddings
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

        if past_key_value is not None:
            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)

        # repeat k/v heads if n_kv_heads < n_heads
        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)

        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
        if attention_mask is not None:  # no matter the length, we just slice it
            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
            attn_weights = attn_weights + causal_mask

        # upcast attention to fp32
        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
        attn_output = torch.matmul(attn_weights, value_states)

        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
            raise ValueError(
                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
                f" {attn_output.size()}"
            )

        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)

        attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value


class LightpostFlashAttention2(LightpostAttention):
    """
    Lightpost flash attention module that inherits from `LightpostAttention`. The weights remain identical to the base class,
    with modifications only to the forward pass to properly integrate with flash attention's API and handle padding tokens.
    For sliding window attention (SWA), it is applied only to the bottom config.max_window_layers layers of the model.
    """

    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
    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 alignement, 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 = not is_flash_attn_greater_or_equal_2_10()

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]],
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Cache] = None,
        output_attentions: bool = False,
        use_cache: bool = False,
        cache_position: Optional[torch.LongTensor] = None,
    ):
        bsz, q_len, _ = hidden_states.size()

        query_states = self.q_proj(hidden_states)
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)

        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)

        cos, sin = position_embeddings
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

        if past_key_value is not None:
            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)

        # repeat k/v heads if n_kv_heads < n_heads
        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)
        dropout_rate = 0.0 if not self.training else self.attention_dropout

        # 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_states.dtype
        if input_dtype == torch.float32:
            if torch.is_autocast_enabled():
                target_dtype = 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.q_proj.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_states = query_states.to(target_dtype)
            key_states = key_states.to(target_dtype)
            value_states = value_states.to(target_dtype)

        # Reashape to the expected shape for Flash Attention
        query_states = query_states.transpose(1, 2)
        key_states = key_states.transpose(1, 2)
        value_states = value_states.transpose(1, 2)

        if (
            self.config.use_sliding_window
            and getattr(self.config, "sliding_window", None) is not None
            and self.layer_idx >= self.config.max_window_layers
        ):
            sliding_window = self.config.sliding_window
        else:
            sliding_window = None

        attn_output = _flash_attention_forward(
            query_states,
            key_states,
            value_states,
            attention_mask,
            q_len,
            position_ids=position_ids,
            dropout=dropout_rate,
            sliding_window=sliding_window,
            is_causal=self.is_causal,
            use_top_left_mask=self._flash_attn_uses_top_left_mask,
        )

        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
        attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value


class LightpostSdpaAttention(LightpostAttention):
    """
    This module implements Lightpost attention using PyTorch's scaled dot-product attention (SDPA) functionality. It extends
    the base `LightpostAttention` class, preserving all weights and parameters. The only modification is in the forward
    pass implementation to leverage the optimized SDPA interface.
    """

    # Adapted from LightpostAttention.forward
    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]],
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Cache] = None,
        output_attentions: bool = False,
        use_cache: bool = False,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        if output_attentions:
            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
            logger.warning_once(
                "LightpostModel is using LightpostSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
                'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
            )
            return super().forward(
                hidden_states=hidden_states,
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_value=past_key_value,
                output_attentions=output_attentions,
                use_cache=use_cache,
            )

        bsz, q_len, _ = hidden_states.size()

        query_states = self.q_proj(hidden_states)
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)

        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)

        cos, sin = position_embeddings
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

        if past_key_value is not None:
            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)

        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)

        causal_mask = attention_mask
        if attention_mask is not None:  # no matter the length, we just slice it
            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]

        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
        # Reference: https://github.com/pytorch/pytorch/issues/112577.
        if query_states.device.type == "cuda" and attention_mask is not None:
            query_states = query_states.contiguous()
            key_states = key_states.contiguous()
            value_states = value_states.contiguous()

        # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
        # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
        # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
        is_causal = True if causal_mask is None and q_len > 1 else False

        attn_output = torch.nn.functional.scaled_dot_product_attention(
            query_states,
            key_states,
            value_states,
            attn_mask=causal_mask,
            dropout_p=self.attention_dropout if self.training else 0.0,
            is_causal=is_causal,
        )

        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.view(bsz, q_len, self.hidden_size)

        if self.mem_attn:
            attn_output = attn_output +self.mem_attn(hidden_states)

        attn_output = self.o_proj(attn_output)


        return attn_output, None, past_key_value


LIGHTPOST_ATTENTION_CLASSES = {
    "eager": LightpostAttention,
    "flash_attention_2": LightpostFlashAttention2,
    "sdpa": LightpostSdpaAttention,
}

# Adapted from QWEN2DecoderLayer
class LightpostDecoderLayer(nn.Module):
    def __init__(self, config: LightpostConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size

        if config.sliding_window and config._attn_implementation != "flash_attention_2":
            logger.warning_once(
                f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
                "unexpected results may be encountered."
            )
        self.self_attn = LIGHTPOST_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)


        self.mlp = LightpostMLP(config)
        self.input_layernorm = LightpostRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = LightpostRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]],
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = False,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs,
    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            position_embeddings (`Tuple[torch.FloatTensor, torch.FloatTensor]`):
                Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
                with `head_dim` being the embedding dimension of each attention head.
            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
                `(batch, sequence_length)` where padding elements are indicated by 0.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            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`).
            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
                Indices depicting the position of the input sequence tokens in the sequence.
            kwargs (`dict`, *optional*):
                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
                into the model
        """

        residual = hidden_states

        hidden_states = self.input_layernorm(hidden_states)

        # Self Attention
        hidden_states, self_attn_weights, present_key_value = self.self_attn(
            hidden_states=hidden_states,
            position_embeddings=position_embeddings,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_value=past_key_value,
            output_attentions=output_attentions,
            use_cache=use_cache,
            cache_position=cache_position,
        )

        hidden_states = residual + hidden_states
        
        # Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights,)

        if use_cache:
            outputs += (present_key_value,)

        return outputs


LIGHTPOST_START_DOCSTRING = r"""
    This model extends [`PreTrainedModel`] and provides access to common functionality like model downloading, saving,
    input embedding resizing, and head pruning. See the parent class documentation for details on these methods.

    As a standard PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module), this model can be
    used like any other PyTorch module. Refer to PyTorch's documentation for general usage patterns and behaviors.

    Parameters:
        config ([`LightpostConfig`]):
            Configuration object containing model parameters. Note that initializing with a config only sets up the model
            architecture - to load pretrained weights, use [`~PreTrainedModel.from_pretrained`].
"""


@add_start_docstrings(
    """
    The base Lightpost Model that outputs raw hidden states from the transformer layers, 
    without any task-specific head (like language modeling or classification) on top. 
    This provides the core transformer functionality that task-specific models can build upon.
    """,
    LIGHTPOST_START_DOCSTRING,
)
class LightpostPreTrainedModel(PreTrainedModel):
    config_class = LightpostConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["LightpostDecoderLayer"]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn_2 = True
    _supports_sdpa = True
    _supports_cache_class = True
    _supports_quantized_cache = True
    _supports_static_cache = True

    def _init_weights(self, module):
        std = self.config.initializer_range
        if isinstance(module, nn.Linear):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()


LIGHTPOST_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
            Input token IDs. These are indices into the model's vocabulary. Padding tokens will be ignored.
            Can be obtained using a tokenizer from the `AutoTokenizer` class.

        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
            Attention mask to avoid attending to padding tokens:
            - 1 for tokens to attend to
            - 0 for tokens to ignore
            See the model's `_prepare_decoder_attention_mask` method for implementation details.

        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Position indices for input tokens, ranging from 0 to config.n_positions - 1.
            Used for positional embeddings.

        past_key_values (`Cache`, *optional*):
            Cached key/value states from previous forward passes to speed up sequential decoding.
            Must be a `Cache` instance (see [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache)).
            When using cached states, only the new tokens need to be provided in input_ids.

        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
            Pre-computed input embeddings. Alternative to passing input_ids.
            Useful for more control over token embedding process.

        use_cache (`bool`, *optional*):
            Whether to return key/value states for use in subsequent forward passes.

        output_attentions (`bool`, *optional*):
            Whether to return attention weights from all layers.

        output_hidden_states (`bool`, *optional*):
            Whether to return hidden states from all layers.

        return_dict (`bool`, *optional*):
            Whether to return a ModelOutput object instead of a tuple.

        cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
            Indices showing true sequence position of input tokens, ignoring padding.
            Used for cache position tracking and inference.
"""


@add_start_docstrings(
    "The bare Lightpost Model outputting raw hidden-states without any specific head on top.",
    LIGHTPOST_START_DOCSTRING,
)
class LightpostModel(LightpostPreTrainedModel):
    """
    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`LightpostDecoderLayer`]

    Args:
        config: LightpostConfig
    """

    def __init__(self, config: LightpostConfig):
        super().__init__(config)
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
        self.layers = nn.ModuleList(
            [LightpostDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self._attn_implementation = config._attn_implementation
        self.norm = LightpostRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = LightpostRotaryEmbedding(config=config)

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


    def set_mem(self, memory_size: int, mem_layers: None | int | list[int] = None):
        if mem_layers is None:
            mem_layers = list(range(len(self.layers)))
        elif isinstance(mem_layers, int):
            mem_layers = [mem_layers]
        
        mem_layers = list(mem_layers)

        print(f"Setting memory size to {memory_size} for layers {mem_layers}")
        self.config.mem_size = memory_size
        self.config.mem_layers = mem_layers

        for ix, layer in enumerate(self.layers):
            if ix in mem_layers:                
                if memory_size == 0 or memory_size is None:
                    layer.self_attn.mem_attn = None
                elif hasattr(layer.self_attn, 'mem_attn'):
                    device = next(layer.parameters()).device
                    dtype = next(layer.parameters()).dtype
                    layer.self_attn.mem_attn = MemoryAttention(config=self.config).to(device, dtype=dtype)
            else:
                if hasattr(layer.self_attn, 'mem_attn'):
                    delattr(layer.self_attn, 'mem_attn')

    def save_mem(self, path: str):
        data = {"version": 1, "layers": {}}
        for ix, layer in enumerate(self.layers):
            if hasattr(layer.self_attn, 'mem_attn') and layer.self_attn.mem_attn is not None:
                data["layers"][ix] = layer.self_attn.mem_attn.state_dict()

        torch.save(data, path)
    
    def load_mem(self, path: str):
        data = torch.load(path, weights_only=True)

        if data['version'] != 1:
            raise ValueError(f"Unsupported version: {data['version']}")

        for ix, state_dict in data["layers"].items():

            if not hasattr(self.layers[ix], 'self_attn'):
                raise ValueError(f"MemoryAttention module not found in layer {ix}")

            device = next(self.layers[ix].parameters()).device
            self.layers[ix].self_attn.mem_attn = MemoryAttention.from_state_dict(state_dict, self.config).to(device)

        # Ensure that the config is updated with the correct memory size
        self.config.mem_layers = list(data["layers"].keys())
        self.config.mem_size = self.layers[self.config.mem_layers[0]].self_attn.mem_attn.memory_size


    def get_input_embeddings(self):
        return self.embed_tokens

    def set_input_embeddings(self, value):
        self.embed_tokens = value

    @add_start_docstrings_to_model_forward(LIGHTPOST_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = 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, BaseModelOutputWithPast]:
        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

        # Ensure `past_key_values` is a `Cache` instance
        if use_cache and past_key_values is None:
            past_key_values = DynamicCache()

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

        if cache_position is None:
            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
            cache_position = torch.arange(
                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], 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
        )

        hidden_states = inputs_embeds

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

        # decoder layers
        all_hidden_states = () if output_hidden_states else None
        all_self_attns = () if output_attentions else None
        next_decoder_cache = None

        for decoder_layer in self.layers:
            if output_hidden_states:
                all_hidden_states += (hidden_states,)

            if self.gradient_checkpointing and self.training:
                layer_outputs = self._gradient_checkpointing_func(
                    decoder_layer.__call__,
                    hidden_states,
                    causal_mask,
                    position_ids,
                    past_key_values,
                    output_attentions,
                    use_cache,
                    cache_position,
                    position_embeddings,
                )
            else:
                layer_outputs = decoder_layer(
                    hidden_states,
                    position_embeddings=position_embeddings,
                    attention_mask=causal_mask,
                    position_ids=position_ids,
                    past_key_value=past_key_values,
                    output_attentions=output_attentions,
                    use_cache=use_cache,
                    cache_position=cache_position
                )

            hidden_states = layer_outputs[0]

            if use_cache:
                next_decoder_cache = layer_outputs[2 if output_attentions else 1]

            if output_attentions:
                all_self_attns += (layer_outputs[1],)

        hidden_states = self.norm(hidden_states)

        # add hidden states from the last decoder layer
        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        next_cache = next_decoder_cache if use_cache else None

        if not return_dict:
            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=next_cache,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
        )

    # Copied from transformers.models.phi3.modeling_phi3.Phi3Model._update_causal_mask
    def _update_causal_mask(
        self,
        attention_mask: torch.Tensor,
        input_tensor: torch.Tensor,
        cache_position: torch.Tensor,
        past_key_values: Cache,
        output_attentions: bool,
    ):
        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)
        using_sliding_window_cache = isinstance(past_key_values, SlidingWindowCache)

        # 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 or using_sliding_window_cache)
            and not output_attentions
        ):
            if AttentionMaskConverter._ignore_causal_mask_sdpa(
                attention_mask,
                inputs_embeds=input_tensor,
                past_key_values_length=past_seen_tokens,
                sliding_window=self.config.sliding_window,
                is_training=self.training,
            ):
                return None

        dtype, device = input_tensor.dtype, input_tensor.device
        min_dtype = torch.finfo(dtype).min
        sequence_length = input_tensor.shape[1]
        # SlidingWindowCache or StaticCache
        if using_sliding_window_cache or using_static_cache:
            target_length = past_key_values.get_max_cache_shape()
        # DynamicCache or no cache
        else:
            target_length = (
                attention_mask.shape[-1]
                if isinstance(attention_mask, torch.Tensor)
                else past_seen_tokens + sequence_length + 1
            )

        # 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],
            config=self.config,
            past_key_values=past_key_values,
        )

        if (
            self.config._attn_implementation == "sdpa"
            and attention_mask is not None
            and attention_mask.device.type == "cuda"
            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.mistral.modeling_mistral.MistralModel._prepare_4d_causal_attention_mask_with_cache_position with Mistral->Lightpost
    def _prepare_4d_causal_attention_mask_with_cache_position(
        attention_mask: torch.Tensor,
        sequence_length: int,
        target_length: int,
        dtype: torch.dtype,
        device: torch.device,
        cache_position: torch.Tensor,
        batch_size: int,
        config: LightpostConfig,
        past_key_values: Cache,
    ):
        """
        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.
            device (`torch.device`):
                The device to plcae the 4D attention mask on.
            cache_position (`torch.Tensor`):
                Indices depicting the position of the input sequence tokens in the sequence.
            batch_size (`torch.Tensor`):
                Batch size.
            config (`LightpostConfig`):
                The model's configuration class
            past_key_values (`Cache`):
                The cache class that is being used currently to generate
        """
        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=device
            )
            diagonal_attend_mask = torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
            if config.sliding_window is not None:
                # if we have sliding window, we should not attend to tokens beyond sliding window length, so we mask them out also
                # the check is needed to verify is current checkpoint was trained with sliding window or not
                if not isinstance(past_key_values, SlidingWindowCache) or sequence_length > target_length:
                    sliding_attend_mask = torch.arange(target_length, device=device) <= (
                        cache_position.reshape(-1, 1) - config.sliding_window
                    )
                    diagonal_attend_mask.bitwise_or_(sliding_attend_mask)
            causal_mask *= diagonal_attend_mask
            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
                if attention_mask.shape[-1] > target_length:
                    attention_mask = attention_mask[:, :target_length]
                mask_length = attention_mask.shape[-1]
                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
                padding_mask = padding_mask == 0
                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
                    padding_mask, min_dtype
                )
        return causal_mask


class LightpostForCausalLM(LightpostPreTrainedModel, GenerationMixin):
    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.model = LightpostModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

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

    def get_input_embeddings(self):
        return self.model.embed_tokens

    def set_input_embeddings(self, value):
        self.model.embed_tokens = value

    def get_output_embeddings(self):
        return self.lm_head

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

    def set_decoder(self, decoder):
        self.model = decoder

    def get_decoder(self):
        return self.model

    @add_start_docstrings_to_model_forward(LIGHTPOST_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: 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,
        num_logits_to_keep: int = 0,
        **loss_kwargs,
    ) -> Union[Tuple, CausalLMOutputWithPast]:
        r"""
        Args:
            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.

            num_logits_to_keep (`int`, *optional*):
                Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate logits for all
                `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
                token can save memory, which becomes pretty significant for long sequences or large vocabulary size.

        Returns:

        Example:

        ```python
        >>> from transformers import AutoTokenizer, LightpostForCausalLM

        >>> model = LightpostForCausalLM.from_pretrained(PATH_TO_WEIGHTS)
        >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_TOKENIZER)

        >>> prompt = "Hey, are you conscious? Can you talk to me?"
        >>> inputs = tokenizer(prompt, return_tensors="pt")

        >>> # Generate
        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
        ```"""

        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            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 = outputs[0]
        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])

        loss = None
        if labels is not None:
            loss = self.loss_function(logits, labels, self.vocab_size, **loss_kwargs)
            self._input_ids = input_ids
            self._logits = logits
            self._labels = labels
            self._attention_mask = attention_mask
            self._loss_kwargs = loss_kwargs
            self._num_logits_to_keep = num_logits_to_keep
            

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

        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


@add_start_docstrings(
    """
    The Lightpost Model transformer with a sequence classification head on top (linear layer).

    [`LightpostForSequenceClassification`] uses the last token in order to do the classification, as other causal models
    (e.g. GPT-2) 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).
    """,
    LIGHTPOST_START_DOCSTRING,
)
class LightpostForSequenceClassification(LightpostPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels
        self.model = LightpostModel(config)
        self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)

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

    def get_input_embeddings(self):
        return self.model.embed_tokens

    def set_input_embeddings(self, value):
        self.model.embed_tokens = value

    @add_start_docstrings_to_model_forward(LIGHTPOST_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: 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,
    ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        transformer_outputs = self.model(
            input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            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:
            sequence_lengths = -1
        else:
            if input_ids is not None:
                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
                sequence_lengths = sequence_lengths % input_ids.shape[-1]
                sequence_lengths = sequence_lengths.to(logits.device)
            else:
                sequence_lengths = -1

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

        loss = None
        if labels is not None:
            labels = labels.to(logits.device)
            if self.config.problem_type is None:
                if self.num_labels == 1:
                    self.config.problem_type = "regression"
                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
                    self.config.problem_type = "single_label_classification"
                else:
                    self.config.problem_type = "multi_label_classification"

            if self.config.problem_type == "regression":
                loss_fct = MSELoss()
                if self.num_labels == 1:
                    loss = loss_fct(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.view(-1, self.num_labels), labels.view(-1))
            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,
        )


@add_start_docstrings(
    """
    The Lightpost Model transformer with a token classification head on top (a linear layer on top of the hidden-states
    output) e.g. for Named-Entity-Recognition (NER) tasks.
    """,
    LIGHTPOST_START_DOCSTRING,
)
# Copied from transformers.models.llama.modeling_llama.LlamaForTokenClassification with Llama->Lightpost, LLAMA->QWEN2
class LightpostForTokenClassification(LightpostPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels
        self.model = LightpostModel(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.score = nn.Linear(config.hidden_size, config.num_labels)

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

    def get_input_embeddings(self):
        return self.model.embed_tokens

    def set_input_embeddings(self, value):
        self.model.embed_tokens = value

    @add_start_docstrings_to_model_forward(LIGHTPOST_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(
        checkpoint=_CHECKPOINT_FOR_DOC,
        output_type=TokenClassifierOutput,
        config_class=_CONFIG_FOR_DOC,
    )
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: 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,
    ) -> Union[Tuple, TokenClassifierOutput]:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        outputs = self.model(
            input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        sequence_output = outputs[0]
        sequence_output = self.dropout(sequence_output)
        logits = self.score(sequence_output)

        loss = None
        if labels is not None:
            loss = self.loss_function(logits, labels, self.config)

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

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


@add_start_docstrings(
    """
The Lightpost Model transformer with a span classification head on top for extractive question-answering tasks like
SQuAD (a linear layer on top of the hidden-states output to compute `span start logits` and `span end logits`).
    """,
    LIGHTPOST_START_DOCSTRING,
)
# Copied from transformers.models.mistral.modeling_mistral.MistralForQuestionAnswering with Mistral->Lightpost
class LightpostForQuestionAnswering(LightpostPreTrainedModel):
    base_model_prefix = "model"

    def __init__(self, config):
        super().__init__(config)
        self.model = LightpostModel(config)
        self.qa_outputs = nn.Linear(config.hidden_size, 2)

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

    def get_input_embeddings(self):
        return self.model.embed_tokens

    def set_input_embeddings(self, value):
        self.model.embed_tokens = value

    @add_start_docstrings_to_model_forward(LIGHTPOST_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Union[Cache, List[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,
        **kwargs,
    ) -> Union[Tuple, QuestionAnsweringModelOutput]:
        r"""
        start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for position (index) of the start of the labelled span for computing the token classification loss.
            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
            are not taken into account for computing the loss.
        end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for position (index) of the end of the labelled span for computing the token classification loss.
            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
            are not taken into account for computing the loss.
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        outputs = self.model(
            input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            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()

        loss = None
        if start_positions is not None and end_positions is not None:
            loss = self.loss_function(start_logits, end_logits, start_positions, end_positions, **kwargs)

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

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