modeling_imp.py

# Copyright (c) MILVLG team.
# Licensed under the Apache 2.0 license.
#
# Some code here is copied from the project Phi-2 (https://huggingface.co/microsoft/phi-2),
# SigLIP@transformers==4.37.0.dev0 (https://huggingface.co/google/siglip-so400m-patch14-384),
# and Llava (https://github.com/haotian-liu/LLaVA), and modified by 
# Zhenwei Shao ([email protected]) @ MILVLG. We thank them for their great works.
# And their original licenses and copyright should be inherited (see the statements
# in `configuration_imp.py` for more details).


# Be careful: The way how `past_key_values.seqlen_offset` is updated is modified from
# the implementation of original Phi-2. See the comments below for details.

from __future__ import annotations
import os
import math
import re
from dataclasses import dataclass, field
from typing import Any, Dict, Optional, Tuple, Union, List
from abc import ABC, abstractmethod

import torch
import torch.nn as nn
from einops import rearrange, repeat
from transformers.cache_utils import Cache, DynamicCache
from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
from transformers import (
    PretrainedConfig, 
    PreTrainedModel,
    AutoConfig,
    AutoModelForCausalLM
)    
from transformers.modeling_utils import PreTrainedModel
from transformers.activations import ACT2FN
from transformers.modeling_outputs import (
    BaseModelOutputWithPast,
    CausalLMOutputWithPast,
    SequenceClassifierOutputWithPast,
    TokenClassifierOutput,
)
import sys
from .configuration_imp import Phi3Config, ImpPhi3Config
from .vision_encoder import VisionTower
# from .vision_encoder import CLIPVisionTower

try:
    from flash_attn.bert_padding import pad_input, unpad_input
    from flash_attn.layers.rotary import RotaryEmbedding as FlashRotaryEmbedding
    from flash_attn.modules.mha import FlashCrossAttention, FlashSelfAttention
    from flash_attn.ops.fused_dense import FusedDense
except:
    pad_input, unpad_input = None, None
    FlashRotaryEmbedding = None
    FlashSelfAttention, FlashCrossAttention = None, None
    FusedDense = None


@dataclass
class InferenceParams:
    """Inference parameters passed to model to efficiently calculate
    and store context during inference.

    Reference:
        https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/utils/generation.py.

    Args:
        max_seqlen: Maximum sequence length.
        max_batch_size: Maximum batch size.
        seqlen_offset: Sequence length offset.
        batch_size_offset: Batch size offset.
        key_value_memory_dict: Key value memory dictionary.
        lengths_per_sample: Lengths per sample.

    """

    max_seqlen: int = field(metadata={"help": "Maximum sequence length."})

    max_batch_size: int = field(metadata={"help": "Maximum batch size."})

    seqlen_offset: int = field(default=0, metadata={"help": "Sequence length offset."})

    batch_size_offset: int = field(default=0, metadata={"help": "Batch size offset."})

    key_value_memory_dict: Dict[str, Any] = field(
        default_factory=dict, metadata={"help": "Key value memory dictionary."}
    )

    lengths_per_sample: torch.Tensor = field(default=None, metadata={"help": "Lengths per sample."})




# Copied from transformers.models.gemma.modeling_gemma.GemmaRotaryEmbedding with gemma->phi3, Gemma->Phi3
class Phi3RotaryEmbedding(nn.Module):
    def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
        super().__init__()

        self.dim = dim
        self.max_position_embeddings = max_position_embeddings
        self.base = base
        self.register_buffer("inv_freq", None, persistent=False)

    @torch.no_grad()
    def forward(self, x, position_ids, seq_len=None):
        # x: [bs, num_attention_heads, seq_len, head_size]
        if self.inv_freq is None:
            self.inv_freq = 1.0 / (
                self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim)
            )
        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 since bfloat16 loses precision on long contexts
        # 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()
        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)


class Phi3SuScaledRotaryEmbedding(Phi3RotaryEmbedding):
    def __init__(self, dim, config, device=None):
        super().__init__(dim, config.max_position_embeddings, config.rope_theta, device)

        self.short_factor = config.rope_scaling["short_factor"]
        self.long_factor = config.rope_scaling["long_factor"]
        self.original_max_position_embeddings = config.original_max_position_embeddings

    @torch.no_grad()
    def forward(self, x, position_ids, seq_len=None):
        seq_len = torch.max(position_ids) + 1
        if seq_len > self.original_max_position_embeddings:
            ext_factors = torch.tensor(self.long_factor, dtype=torch.float32, device=x.device)
        else:
            ext_factors = torch.tensor(self.short_factor, dtype=torch.float32, device=x.device)

        inv_freq_shape = torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim
        self.inv_freq = 1.0 / (ext_factors * self.base**inv_freq_shape)

        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 since bfloat16 loses precision on long contexts
        # 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)

            scale = self.max_position_embeddings / self.original_max_position_embeddings
            if scale <= 1.0:
                scaling_factor = 1.0
            else:
                scaling_factor = math.sqrt(1 + math.log(scale) / math.log(self.original_max_position_embeddings))

            cos = emb.cos() * scaling_factor
            sin = emb.sin() * scaling_factor
        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)


class Phi3YarnScaledRotaryEmbedding(Phi3RotaryEmbedding):
    def __init__(self, dim, config, device=None):
        super().__init__(dim, config.max_position_embeddings, config.rope_theta, device)

        self.short_factor = config.rope_scaling["short_factor"]
        self.long_factor = config.rope_scaling["long_factor"]
        self.original_max_position_embeddings = config.original_max_position_embeddings

    @torch.no_grad()
    def forward(self, x, position_ids, seq_len=None):
        seq_len = torch.max(position_ids) + 1
        if seq_len > self.original_max_position_embeddings:
            ext_factors = torch.tensor(self.long_factor, dtype=torch.float32, device=x.device)
        else:
            ext_factors = torch.tensor(self.short_factor, dtype=torch.float32, device=x.device)

        inv_freq_shape = torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim
        self.inv_freq = 1.0 / (ext_factors * self.base**inv_freq_shape)

        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 since bfloat16 loses precision on long contexts
        # 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)

            scale = self.max_position_embeddings / self.original_max_position_embeddings
            if scale <= 1.0:
                scaling_factor = 1.0
            else:
                scaling_factor = 0.1 * math.log(scale) + 1.0

            cos = emb.cos() * scaling_factor
            sin = emb.sin() * scaling_factor
        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 '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



class Phi3MLP(nn.Module):
    def __init__(self, config):
        super().__init__()

        self.config = config
        self.gate_up_proj = nn.Linear(config.hidden_size, 2 * config.intermediate_size, bias=False)
        self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)

        self.activation_fn = ACT2FN[config.hidden_act]

    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
        up_states = self.gate_up_proj(hidden_states)

        gate, up_states = up_states.chunk(2, dim=-1)
        up_states = up_states * self.activation_fn(gate)

        return self.down_proj(up_states)

class Phi3RMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        Phi3RMSNorm 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)



# Copied from transformers.models.llama.modeling_llama.repeat_kv with llama->phi
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 Phi3Attention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(self, config: Phi3Config, layer_idx: Optional[int] = None):
        super().__init__()
        self.config = config
        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."
            # )
            pass

        self.attention_dropout = config.attention_dropout
        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.original_max_position_embeddings = config.original_max_position_embeddings
        self.rope_theta = config.rope_theta
        self.rope_scaling = config.rope_scaling
        self.is_causal = True

        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})."
            )

        op_size = self.num_heads * self.head_dim + 2 * (self.num_key_value_heads * self.head_dim)
        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
        self.qkv_proj = nn.Linear(self.hidden_size, op_size, bias=False)
        self._init_rope()

    def _init_rope(self):
        if self.rope_scaling is None:
            self.rotary_emb = Phi3RotaryEmbedding(
                self.head_dim,
                max_position_embeddings=self.max_position_embeddings,
                base=self.rope_theta,
            )
        else:
            scaling_type = self.config.rope_scaling["type"]
            if scaling_type == "su":
                self.rotary_emb = Phi3SuScaledRotaryEmbedding(self.head_dim, self.config)
            elif scaling_type == "yarn":
                self.rotary_emb = Phi3YarnScaledRotaryEmbedding(self.head_dim, self.config)
            else:
                raise ValueError(f"Unknown RoPE scaling type {scaling_type}")

    def forward(
        self,
        hidden_states: 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,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        # logger.warning_once("You are not running the flash-attention implementation, expect numerical differences.")

        bsz, q_len, _ = hidden_states.size()

        qkv = self.qkv_proj(hidden_states)
        query_pos = self.num_heads * self.head_dim
        query_states = qkv[..., :query_pos]
        key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
        value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]

        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)

        kv_seq_len = key_states.shape[-2]
        if past_key_value is not None:
            if self.layer_idx is None:
                raise ValueError(
                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
                    "with a layer index."
                )
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
        cos, sin = self.rotary_emb(value_states, position_ids, seq_len=kv_seq_len)

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

        if past_key_value is not None:
            cache_kwargs = {"sin": sin, "cos": cos}  # 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 attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
            raise ValueError(
                f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
                f" {attn_weights.size()}"
            )

        if attention_mask is not None:
            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
                raise ValueError(
                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
                )
            attn_weights = attn_weights + attention_mask

        # upcast attention to fp32
        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(value_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 Phi3FlashAttention2(Phi3Attention):
    """
    Phi-3 flash attention module. This module inherits from `Phi3Attention` 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.
    """

    # 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,
        attention_mask: Optional[torch.LongTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Cache] = None,
        output_attentions: bool = False,
        use_cache: bool = False,
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        # Phi3FlashAttention2 attention does not support output_attentions

        if not _flash_supports_window_size:
            # logger.warning_once(
            #     "The current flash attention version does not support sliding window attention. Please use `attn_implementation='eager'` or upgrade flash-attn library."
            # )
            raise ValueError("The current flash attention version does not support sliding window attention.")

        output_attentions = False

        if "padding_mask" in kwargs:
            warnings.warn(
                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
            )

            # overwrite attention_mask with padding_mask
            attention_mask = kwargs.pop("padding_mask")

        bsz, q_len, _ = hidden_states.size()

        qkv = self.qkv_proj(hidden_states)
        query_pos = self.num_heads * self.head_dim
        query_states = qkv[..., :query_pos]
        key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
        value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]

        # Flash attention requires the input to have the shape
        # batch_size x seq_length x head_dim x hidden_dim
        # therefore we just need to keep the original shape
        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)

        kv_seq_len = key_states.shape[-2]
        if past_key_value is not None:
            if self.layer_idx is None:
                raise ValueError(
                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
                    "with a layer index."
                )
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)

        # Because the input can be padded, the absolute sequence length depends on the max position id.
        rotary_seq_len = max(kv_seq_len, position_ids[:, -1].max().item()) + 1
        cos, sin = self.rotary_emb(value_states, position_ids, seq_len=rotary_seq_len)

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

        use_sliding_windows = (
            _flash_supports_window_size
            and getattr(self.config, "sliding_window", None) is not None
            and kv_seq_len > self.config.sliding_window
        )

        if past_key_value is not None:
            # Activate slicing cache only if the config has a value `sliding_windows` attribute
            cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0
            if (
                getattr(self.config, "sliding_window", None) is not None
                and kv_seq_len > self.config.sliding_window
                and cache_has_contents
            ):
                slicing_tokens = 1 - self.config.sliding_window

                past_key = past_key_value[self.layer_idx][0]
                past_value = past_key_value[self.layer_idx][1]

                past_key = past_key[:, :, slicing_tokens:, :].contiguous()
                past_value = past_value[:, :, slicing_tokens:, :].contiguous()

                if past_key.shape[-2] != self.config.sliding_window - 1:
                    raise ValueError(
                        f"past key must have a shape of (`batch_size, num_heads, self.config.sliding_window-1, head_dim`), got"
                        f" {past_key.shape}"
                    )

                if attention_mask is not None:
                    attention_mask = attention_mask[:, slicing_tokens:]
                    attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1)

            cache_kwargs = {"sin": sin, "cos": cos}  # 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_dropout = self.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 the correct dtype just to be sure everything works as expected.
        # This might slowdown training & inference so it is recommended to not cast the LayerNorms
        # in fp32.

        if query_states.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.qkv_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)

        attn_output = self._flash_attention_forward(
            query_states,
            key_states,
            value_states,
            attention_mask,
            q_len,
            dropout=attn_dropout,
            use_sliding_windows=use_sliding_windows,
        )

        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

    # Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2._flash_attention_forward
    def _flash_attention_forward(
        self,
        query_states,
        key_states,
        value_states,
        attention_mask,
        query_length,
        dropout=0.0,
        softmax_scale=None,
        use_sliding_windows=False,
    ):
        """
        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
        first unpad the input, then computes the attention scores and pad the final attention scores.

        Args:
            query_states (`torch.Tensor`):
                Input query states to be passed to Flash Attention API
            key_states (`torch.Tensor`):
                Input key states to be passed to Flash Attention API
            value_states (`torch.Tensor`):
                Input value states to be passed to Flash Attention API
            attention_mask (`torch.Tensor`):
                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
                position of padding tokens and 1 for the position of non-padding tokens.
            dropout (`float`):
                Attention dropout
            softmax_scale (`float`, *optional*):
                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
            use_sliding_windows (`bool`, *optional*):
                Whether to activate sliding window attention.
        """
        if not self._flash_attn_uses_top_left_mask:
            causal = self.is_causal
        else:
            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
            causal = self.is_causal and query_length != 1

        # Contains at least one padding token in the sequence
        if attention_mask is not None:
            batch_size = query_states.shape[0]
            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
                query_states, key_states, value_states, attention_mask, query_length
            )

            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens

            if not use_sliding_windows:
                attn_output_unpad = flash_attn_varlen_func(
                    query_states,
                    key_states,
                    value_states,
                    cu_seqlens_q=cu_seqlens_q,
                    cu_seqlens_k=cu_seqlens_k,
                    max_seqlen_q=max_seqlen_in_batch_q,
                    max_seqlen_k=max_seqlen_in_batch_k,
                    dropout_p=dropout,
                    softmax_scale=softmax_scale,
                    causal=causal,
                )
            else:
                attn_output_unpad = flash_attn_varlen_func(
                    query_states,
                    key_states,
                    value_states,
                    cu_seqlens_q=cu_seqlens_q,
                    cu_seqlens_k=cu_seqlens_k,
                    max_seqlen_q=max_seqlen_in_batch_q,
                    max_seqlen_k=max_seqlen_in_batch_k,
                    dropout_p=dropout,
                    softmax_scale=softmax_scale,
                    causal=causal,
                    window_size=(self.config.sliding_window, self.config.sliding_window),
                )

            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
        else:
            if not use_sliding_windows:
                attn_output = flash_attn_func(
                    query_states,
                    key_states,
                    value_states,
                    dropout,
                    softmax_scale=softmax_scale,
                    causal=causal,
                )
            else:
                attn_output = flash_attn_func(
                    query_states,
                    key_states,
                    value_states,
                    dropout,
                    softmax_scale=softmax_scale,
                    causal=causal,
                    window_size=(self.config.sliding_window, self.config.sliding_window),
                )

        return attn_output

    # Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2._upad_input
    def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
        batch_size, kv_seq_len, num_heads, head_dim = key_layer.shape

        # On the first iteration we need to properly re-create the padding mask
        # by slicing it on the proper place
        if kv_seq_len != attention_mask.shape[-1]:
            attention_mask_num_tokens = attention_mask.shape[-1]
            attention_mask = attention_mask[:, attention_mask_num_tokens - kv_seq_len :]

        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)

        key_layer = index_first_axis(key_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
        value_layer = index_first_axis(value_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)

        if query_length == kv_seq_len:
            query_layer = index_first_axis(
                query_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k
            )
            cu_seqlens_q = cu_seqlens_k
            max_seqlen_in_batch_q = max_seqlen_in_batch_k
            indices_q = indices_k
        elif query_length == 1:
            max_seqlen_in_batch_q = 1
            cu_seqlens_q = torch.arange(
                batch_size + 1, dtype=torch.int32, device=query_layer.device
            )  # There is a memcpy here, that is very bad.
            indices_q = cu_seqlens_q[:-1]
            query_layer = query_layer.squeeze(1)
        else:
            # The -q_len: slice assumes left padding.
            attention_mask = attention_mask[:, -query_length:]
            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)

        return (
            query_layer,
            key_layer,
            value_layer,
            indices_q,
            (cu_seqlens_q, cu_seqlens_k),
            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
        )


# copied from transformers.models.llama.modeling_llama.LlamaSdpaAttention with Llama->Phi3
# TODO @Arthur no longer copied from LLama after static cache
class Phi3SdpaAttention(Phi3Attention):
    """
    Phi3 attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
    `Phi3Attention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
    SDPA API.
    """

    # Adapted from Phi3Attention.forward
    def forward(
        self,
        hidden_states: 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,
    ) -> 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(
            #     "Phi3Model is using Phi3SdpaAttention, 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()

        qkv = self.qkv_proj(hidden_states)
        query_pos = self.num_heads * self.head_dim
        query_states = qkv[..., :query_pos]
        key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
        value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]

        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)

        kv_seq_len = key_states.shape[-2]
        if past_key_value is not None:
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
        cos, sin = self.rotary_emb(value_states, position_ids, seq_len=kv_seq_len)

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

        if past_key_value is not None:
            cache_kwargs = {"sin": sin, "cos": cos}  # 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)

        if attention_mask is not None:
            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
                raise ValueError(
                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
                )

        # 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()

        attn_output = torch.nn.functional.scaled_dot_product_attention(
            query_states,
            key_states,
            value_states,
            attn_mask=attention_mask,
            dropout_p=self.attention_dropout if self.training else 0.0,
            # 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=self.is_causal and attention_mask is None and q_len > 1,
        )

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

        attn_output = self.o_proj(attn_output)

        return attn_output, None, past_key_value




PHI3_ATTENTION_CLASSES = {
    "eager": Phi3Attention,
    "flash_attention_2": Phi3FlashAttention2,
    "sdpa": Phi3SdpaAttention,
}

class Phi3DecoderLayer(nn.Module):
    def __init__(self, config: Phi3Config, layer_idx: int):
        super().__init__()

        self.config = config
        self.self_attn = PHI3_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx=layer_idx)

        self.mlp = Phi3MLP(config)
        self.input_layernorm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

        self.resid_attn_dropout = nn.Dropout(config.resid_pdrop)
        self.resid_mlp_dropout = nn.Dropout(config.resid_pdrop)
        self.post_attention_layernorm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(
        self,
        hidden_states: 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,
        **kwargs,
    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
        if "padding_mask" in kwargs:
            warnings.warn(
                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
            )
        """
        Args:
            hidden_states (`torch.FloatTensor`):
                input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
            position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
                Indices of positions of each input sequence tokens in the position embeddings. Selected in the range
                `[0, config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids)
            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
        """

        residual = hidden_states

        hidden_states = self.input_layernorm(hidden_states)

        # Self Attention
        attn_outputs, self_attn_weights, present_key_value = self.self_attn(
            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,
        )

        hidden_states = residual + self.resid_attn_dropout(attn_outputs)

        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + self.resid_mlp_dropout(hidden_states)

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights,)

        if use_cache:
            outputs += (present_key_value,)

        return outputs



class Phi3PreTrainedModel(PreTrainedModel):
    config_class = Phi3Config
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["Phi3DecoderLayer"]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn_2 = True
    _supports_sdpa = False
    _supports_cache_class = True

    _version = "0.0.5"

    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_()
    
    def prepare_inputs_for_generation(
        self,
        input_ids: torch.LongTensor,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[Union[torch.LongTensor, torch.BoolTensor]] = None,
        **kwargs,
    ) -> Dict[str, Any]:
        if past_key_values is not None:
            if isinstance(past_key_values, Cache):
                cache_length = past_key_values.get_seq_length()
                past_length = past_key_values.seen_tokens
                max_cache_length = past_key_values.get_max_length()
            else:
                cache_length = past_length = past_key_values[0][0].shape[2]
                max_cache_length = None

            # Keep only the unprocessed tokens:
            # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
            # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
            # input)
            if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
                input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
            # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
            # input_ids based on the past_length.
            elif past_length < input_ids.shape[1]:
                input_ids = input_ids[:, past_length:]
            # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.

            # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
            if (
                max_cache_length is not None
                and attention_mask is not None
                and cache_length + input_ids.shape[1] > max_cache_length
            ):
                attention_mask = attention_mask[:, -max_cache_length:]

        position_ids = kwargs.get("position_ids", None)
        if attention_mask is not None and position_ids is None:
            # create position_ids on the fly for batch generation
            position_ids = attention_mask.long().cumsum(-1) - 1
            position_ids.masked_fill_(attention_mask == 0, 1)
            if past_key_values:
                position_ids = position_ids[:, -input_ids.shape[1] :]

        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
        if inputs_embeds is not None and past_key_values is None:
            model_inputs = {"inputs_embeds": inputs_embeds}
        else:
            model_inputs = {"input_ids": input_ids}

        model_inputs.update(
            {
                "position_ids": position_ids,
                "past_key_values": past_key_values,
                "use_cache": kwargs.get("use_cache"),
                "attention_mask": attention_mask,
            }
        )
        return model_inputs




class LlavaMetaModel(ABC):
    """
    Define the APIs for building components that are related to image perceiving.
    This implementation is based on the implementation from the Llave project.
    """

    def get_vision_tower(self):
        vision_tower = getattr(self, 'vision_tower', None)
        if type(vision_tower) is list:
            vision_tower = vision_tower[0]
        return vision_tower
    
    def build_vision_tower(self, config):
        self.vision_tower = VisionTower(config.vision_tower_cfg)
        # self.vision_tower = CLIPVisionTower(config.vision_tower_cfg)

    def build_vision_projector(self, config):
        projector_type = getattr(config, 'mm_projector_type', 'linear')

        if projector_type == 'linear':
            self.mm_projector = nn.Linear(config.mm_hidden_size, config.hidden_size)
            return

        mlp_gelu_match = re.match(r'^mlp(\d+)x_gelu$', projector_type)
        if mlp_gelu_match:
            mlp_depth = int(mlp_gelu_match.group(1))
            modules = [nn.Linear(config.mm_hidden_size, config.hidden_size)]
            for _ in range(1, mlp_depth):
                modules.append(nn.GELU())
                modules.append(nn.Linear(config.hidden_size, config.hidden_size))
            self.mm_projector = nn.Sequential(*modules)
            return

        if projector_type == 'identity':
            self.mm_projector = nn.Identity()
            return

        raise ValueError(f'Unknown projector type: {projector_type}')


class ImpPhi3Model(Phi3PreTrainedModel, LlavaMetaModel):
    """Imp model. This implementation is modified from the implementation of Phi-2"""

    config_class = ImpPhi3Config

    def __init__(self, config: ImpPhi3Config) -> None:
        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.embed_dropout = nn.Dropout(config.embd_pdrop)
        self.layers = nn.ModuleList(
            [Phi3DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self._attn_implementation = config._attn_implementation
        self.norm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

        self.gradient_checkpointing = False

        if hasattr(config, "mm_vision_tower"):
            self.build_vision_tower(config)
            self.build_vision_projector(config)
        # Initialize weights and apply final processing
        self.post_init()

    
    def get_input_embeddings(self) -> nn.Embedding:
        return self.embed_tokens
    
    def set_input_embeddings(self, new_embeddings: nn.Embedding) -> None:
        self.embed_tokens = value

    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,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = 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

        # retrieve input_ids and inputs_embeds
        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
        elif input_ids is not None:
            batch_size, seq_length = input_ids.shape[:2]
        elif inputs_embeds is not None:
            batch_size, seq_length = inputs_embeds.shape[:2]
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        past_key_values_length = 0

        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 use_cache:
            use_legacy_cache = not isinstance(past_key_values, Cache)
            if use_legacy_cache:
                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
            past_key_values_length = past_key_values.get_usable_length(seq_length)

        if position_ids is None:
            device = input_ids.device if input_ids is not None else inputs_embeds.device
            position_ids = torch.arange(
                past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
            )
            position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
        else:
            position_ids = position_ids.view(-1, seq_length).long()

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

        if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache:
            is_padding_right = attention_mask[:, -1].sum().item() != batch_size
            if is_padding_right:
                raise ValueError(
                    "You are attempting to perform batched generation with padding_side='right'"
                    " this may lead to unexpected behaviour for Flash Attention version of Phi3. Make sure to "
                    " call `tokenizer.padding_side  = 'left'` before tokenizing the input. "
                )

        if self._attn_implementation == "flash_attention_2":
            # 2d mask is passed through the layers
            attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
        else:
            # 4d mask is passed through the layers
            attention_mask = _prepare_4d_causal_attention_mask(
                attention_mask,
                (batch_size, seq_length),
                inputs_embeds,
                past_key_values_length,
                sliding_window=self.config.sliding_window,
            )

        hidden_states = inputs_embeds

        # 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,
                    attention_mask,
                    position_ids,
                    past_key_values,
                    output_attentions,
                    use_cache,
                )
            else:
                layer_outputs = decoder_layer(
                    hidden_states,
                    attention_mask=attention_mask,
                    position_ids=position_ids,
                    past_key_value=past_key_values,
                    output_attentions=output_attentions,
                    use_cache=use_cache,
                )

            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 = None
        if use_cache:
            next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache
        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,
        )



class LlavaMetaForCausalLM(ABC):
    """This implementation is based on the implementation from the Llave project."""

    def init_constants(self, config):
        self.IGNORE_INDEX = getattr(config, 'ignore_index', -100)
        self.IMAGE_TOKEN_INDEX = getattr(config, 'image_token_index', 50296)
        self.DEFAULT_IMAGE_TOKEN = getattr(config, 'image_token', "<image>")

    @abstractmethod
    def get_model(self):
        pass

    def get_vision_tower(self):
        return self.get_model().get_vision_tower()

    def encode_images(self, images):
        image_features = self.get_model().get_vision_tower()(images)
        image_features = self.get_model().mm_projector(image_features)
        return image_features

    def prepare_inputs_labels_for_multimodal(
        self, input_ids, position_ids, attention_mask, past_key_values, labels, images
    ):
        vision_tower = self.get_vision_tower()
        # if vision_tower is None or images is None or past_key_values.seqlen_offset != 0:
        if past_key_values is not None:
            target_shape = past_key_values[0][0].shape[2] + 1
            attention_mask = torch.ones(
                (attention_mask.shape[0], target_shape),
                dtype=attention_mask.dtype,
                device=attention_mask.device
            )
            position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1
            # print('position_ids', position_ids.shape)
            # print(input_ids[:, -1:].item())
            return input_ids[:, -1:], position_ids, attention_mask, past_key_values, None, labels

        if type(images) is list or images.ndim == 5:
            concat_images = torch.cat([image for image in images], dim=0)
            # concat_images.requires_grad_(True)
            image_features = self.encode_images(concat_images)
            split_sizes = [image.shape[0] for image in images]
            image_features = torch.split(image_features, split_sizes, dim=0)
            image_features = [x.flatten(0, 1).to(self.device) for x in image_features]
        else:
            # images.requires_grad_(True)
            image_features = self.encode_images(images).to(self.device)

        # TODO: image start / end is not implemented here to support pretraining.
        if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False):
            raise NotImplementedError

        # Let's just add dummy tensors if they do not exist,
        # it is a headache to deal with None all the time.
        # But it is not ideal, and if you have a better idea,
        # please open an issue / submit a PR, thanks.
        _labels = labels
        _position_ids = position_ids
        _attention_mask = attention_mask
        if attention_mask is None:
            attention_mask = torch.ones_like(input_ids, dtype=torch.bool)
        else:
            attention_mask = attention_mask.bool()
        if position_ids is None:
            position_ids = torch.arange(0, input_ids.shape[1], dtype=torch.long, device=input_ids.device)
        if labels is None:
            labels = torch.full_like(input_ids, self.IGNORE_INDEX)

        # remove the padding using attention_mask -- TODO: double check
        input_ids = [cur_input_ids[cur_attention_mask] for cur_input_ids, cur_attention_mask in zip(input_ids, attention_mask)]
        labels = [cur_labels[cur_attention_mask] for cur_labels, cur_attention_mask in zip(labels, attention_mask)]

        new_input_embeds = []
        new_labels = []
        cur_image_idx = 0
        for batch_idx, cur_input_ids in enumerate(input_ids):
            num_images = (cur_input_ids == self.IMAGE_TOKEN_INDEX).sum()
            if num_images == 0:
                cur_image_features = image_features[cur_image_idx]
                cur_input_embeds_1 = self.get_model().embed_tokens(cur_input_ids)
                cur_input_embeds = torch.cat([cur_input_embeds_1, cur_image_features[0:0]], dim=0)
                new_input_embeds.append(cur_input_embeds)
                new_labels.append(labels[batch_idx])
                cur_image_idx += 1
                continue

            image_token_indices = [-1] + torch.where(cur_input_ids == self.IMAGE_TOKEN_INDEX)[0].tolist() + [cur_input_ids.shape[0]]
            cur_input_ids_noim = []
            cur_labels = labels[batch_idx]
            cur_labels_noim = []
            for i in range(len(image_token_indices) - 1):
                cur_input_ids_noim.append(cur_input_ids[image_token_indices[i]+1:image_token_indices[i+1]])
                cur_labels_noim.append(cur_labels[image_token_indices[i]+1:image_token_indices[i+1]])
            split_sizes = [x.shape[0] for x in cur_labels_noim]
            cur_input_embeds = self.get_model().embed_tokens(torch.cat(cur_input_ids_noim))
            # print(cur_input_embeds.shape)
            cur_input_embeds_no_im = torch.split(cur_input_embeds, split_sizes, dim=0)
            cur_new_input_embeds = []
            cur_new_labels = []

            for i in range(num_images + 1):
                cur_new_input_embeds.append(cur_input_embeds_no_im[i])
                cur_new_labels.append(cur_labels_noim[i])
                if i < num_images:
                    cur_image_features = image_features[cur_image_idx]
                    cur_image_idx += 1
                    cur_new_input_embeds.append(cur_image_features)
                    cur_new_labels.append(torch.full((cur_image_features.shape[0],), self.IGNORE_INDEX, device=cur_labels.device, dtype=cur_labels.dtype))

            cur_new_input_embeds = torch.cat(cur_new_input_embeds)
            cur_new_labels = torch.cat(cur_new_labels)

            new_input_embeds.append(cur_new_input_embeds)
            new_labels.append(cur_new_labels)

        # Truncate sequences to max length as image embeddings can make the sequence longer
        tokenizer_model_max_length = getattr(self.config, 'tokenizer_model_max_length', None)
        if tokenizer_model_max_length is not None:
            new_input_embeds = [x[:tokenizer_model_max_length] for x in new_input_embeds]
            new_labels = [x[:tokenizer_model_max_length] for x in new_labels]

        # Combine them
        max_len = max(x.shape[0] for x in new_input_embeds)
        batch_size = len(new_input_embeds)

        new_input_embeds_padded = []
        new_labels_padded = torch.full((batch_size, max_len), self.IGNORE_INDEX, dtype=new_labels[0].dtype, device=new_labels[0].device)
        attention_mask = torch.zeros((batch_size, max_len), dtype=attention_mask.dtype, device=attention_mask.device)
        position_ids = torch.zeros((batch_size, max_len), dtype=position_ids.dtype, device=position_ids.device)

        for i, (cur_new_embed, cur_new_labels) in enumerate(zip(new_input_embeds, new_labels)):
            cur_len = cur_new_embed.shape[0]
            if getattr(self.config, 'tokenizer_padding_side', 'right') == "left":
                new_input_embeds_padded.append(torch.cat((
                    torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device),
                    cur_new_embed
                ), dim=0))
                if cur_len > 0:
                    new_labels_padded[i, -cur_len:] = cur_new_labels
                    attention_mask[i, -cur_len:] = True
                    position_ids[i, -cur_len:] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
            else:
                new_input_embeds_padded.append(torch.cat((
                    cur_new_embed,
                    torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device)
                ), dim=0))
                if cur_len > 0:
                    new_labels_padded[i, :cur_len] = cur_new_labels
                    attention_mask[i, :cur_len] = True
                    position_ids[i, :cur_len] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)

        new_input_embeds = torch.stack(new_input_embeds_padded, dim=0)

        if new_input_embeds.shape[-2] > 2000:
            self.need_clear_cache = True

        if _labels is None:
            new_labels = None
        else:
            new_labels = new_labels_padded

        if _attention_mask is None:
            attention_mask = None
        else:
            attention_mask = attention_mask.to(dtype=_attention_mask.dtype)

        if _position_ids is None:
            position_ids = None

        return None, position_ids, attention_mask, past_key_values, new_input_embeds, new_labels


class ImpPhi3ForCausalLM(Phi3PreTrainedModel, LlavaMetaForCausalLM):
    """Impphi3 for Causal Language Modeling."""

    config_class = ImpPhi3Config

    def __init__(self, config: ImpPhi3Config) -> None:
        super().__init__(config)

        self.model = ImpPhi3Model(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        self.need_clear_cache = False
        
        self.post_init()
        self.init_constants(config)

    # def get_output_embeddings(self) -> nn.Linear:
    #     return self.lm_head

    # def set_output_embeddings(self, new_embeddings: nn.Linear) -> None:
    #     self.lm_head = new_embeddings

    def get_model(self):
        return self.model

    
    def image_preprocess(self, images):
        return self.get_vision_tower().image_processor(images)['pixel_values']
    

    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,
        images: Optional[torch.FloatTensor] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, CausalLMOutputWithPast]:

        if inputs_embeds is None:
            (
                input_ids,
                position_ids,
                attention_mask,
                past_key_values,
                inputs_embeds,
                labels
            ) = self.prepare_inputs_labels_for_multimodal(
                input_ids,
                position_ids,
                attention_mask,
                past_key_values,
                labels,
                images
            )
        
        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

        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,
        )

        hidden_states = outputs[0]
        logits = self.lm_head(hidden_states)
        # logits = logits.float()

        loss = None
        if labels is not None:
            # Shift so that tokens < n predict n
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = CrossEntropyLoss()
            shift_logits = shift_logits.view(-1, self.config.vocab_size)
            shift_labels = shift_labels.view(-1)
            # Enable model parallelism
            shift_labels = shift_labels.to(shift_logits.device)
            loss = loss_fct(shift_logits, shift_labels)

        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,
        )


        # # inputs_embeds.requires_grad_(True)
        # return super().forward(
        #     input_ids=input_ids,
        #     attention_mask=attention_mask,
        #     position_ids=position_ids,
        #     past_key_values=past_key_values,
        #     inputs_embeds=inputs_embeds,
        #     labels=labels,
        #     use_cache=use_cache,
        #     output_attentions=output_attentions,
        #     output_hidden_states=output_hidden_states,
        #     return_dict=return_dict
        # )

    def prepare_inputs_for_generation(self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs):
        images = kwargs.pop("images", None)
        _inputs = super().prepare_inputs_for_generation(
            input_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, **kwargs
        )
        if images is not None:
            _inputs['images'] = images
        return _inputs

AutoConfig.register("imp_phi3", ImpPhi3Config)
AutoModelForCausalLM.register(ImpPhi3Config, ImpPhi3ForCausalLM)