final_code.py

import torch
from torch.utils.data import Dataset, DataLoader
from transformers import MT5ForConditionalGeneration, MT5Tokenizer, AdamW
from transformers import AutoModel, AutoTokenizer
from sklearn.metrics.pairwise import cosine_similarity
import pandas as pd
import matplotlib.pyplot as plt
import numpy as np
from huggingface_hub import HfApi, HfFolder, Repository, notebook_login, create_repo, upload_folder
import os
import shutil

# ========== CONFIG ==========
HF_USERNAME = "aarath97"
HF_REPO = "mt5-dogri-translation"
MODEL_NAME = "google/mt5-large"
BATCH_SIZE = 2
LR = 1e-5
DPO_STEPS = 100
HGRL_STEPS = 100
COMBINED_STEPS = 50
GAMMA = 3.5
ALPHA = 0.5
BETA = 0.5

# ========== LOAD DATA ==========
df = pd.read_excel("dogri_train.xlsx")
train_data = list(zip(df['Dogri'], df['English'], df['Unpreffered']))

# ========== TOKENIZERS ==========
tokenizer = MT5Tokenizer.from_pretrained(MODEL_NAME)
sbert = AutoModel.from_pretrained("sentence-transformers/all-MiniLM-L6-v2")
sbert_tokenizer = AutoTokenizer.from_pretrained("sentence-transformers/all-MiniLM-L6-v2")

# ========== UTILITIES ==========
def compute_similarity(sent1, sent2):
    emb1 = sbert(**sbert_tokenizer(sent1, return_tensors='pt')).last_hidden_state.mean(1)
    emb2 = sbert(**sbert_tokenizer(sent2, return_tensors='pt')).last_hidden_state.mean(1)
    return cosine_similarity(emb1.detach().numpy(), emb2.detach().numpy())[0][0]

def hyper_gamma_reward(rho):
    return rho * np.exp(-GAMMA * (1 - rho))

# ========== DATASET ==========
class DogriDataset(Dataset):
    def __init__(self, data):
        self.data = data

    def __len__(self):
        return len(self.data)

    def __getitem__(self, idx):
        return self.data[idx]

dataloader = DataLoader(DogriDataset(train_data), batch_size=BATCH_SIZE, shuffle=True)

# ========== TRAINING ==========
model = MT5ForConditionalGeneration.from_pretrained(MODEL_NAME).to("cuda")
optimizer = AdamW(model.parameters(), lr=LR)

dpo_losses, hgrl_losses, final_losses = [], [], []

# ---- DPO Training ----
for step in range(DPO_STEPS):
    batch = next(iter(dataloader))
    loss_batch = []
    for src, ref, unpref in zip(*batch):
        input_ids = tokenizer(src, return_tensors='pt', truncation=True, padding=True).input_ids.to("cuda")
        ref_ids = tokenizer(ref, return_tensors='pt', truncation=True, padding=True).input_ids.to("cuda")
        unpref_ids = tokenizer(unpref, return_tensors='pt', truncation=True, padding=True).input_ids.to("cuda")

        ref_logprob = model(input_ids=input_ids, labels=ref_ids).loss
        unpref_logprob = model(input_ids=input_ids, labels=unpref_ids).loss

        logit_diff = -ref_logprob.item() + unpref_logprob.item()
        beta = 1.0
        loss = -torch.log(torch.sigmoid(torch.tensor(beta * logit_diff)))
        loss_batch.append(loss)

    loss_val = torch.stack(loss_batch).mean()
    loss_val.backward()
    optimizer.step()
    optimizer.zero_grad()
    dpo_losses.append(loss_val.item())

# ---- HGRL Training ----
for step in range(HGRL_STEPS):
    batch = next(iter(dataloader))
    loss_batch = []
    for src, ref, _ in zip(*batch):
        input_ids = tokenizer(src, return_tensors='pt').input_ids.to("cuda")
        gen_ids = model.generate(input_ids)
        gen_text = tokenizer.decode(gen_ids[0], skip_special_tokens=True)

        rho = compute_similarity(gen_text, ref)
        reward = hyper_gamma_reward(rho)

        labels = tokenizer(gen_text, return_tensors='pt').input_ids.to("cuda")
        logprob = model(input_ids=input_ids, labels=labels).loss

        loss = -reward * logprob
        loss_batch.append(loss)

    loss_val = torch.stack(loss_batch).mean()
    loss_val.backward()
    optimizer.step()
    optimizer.zero_grad()
    hgrl_losses.append(loss_val.item())

# ---- Combined Training ----
for step in range(COMBINED_STEPS):
    batch = next(iter(dataloader))
    loss_dpo_batch, loss_hgrl_batch = [], []
    for src, ref, unpref in zip(*batch):
        input_ids = tokenizer(src, return_tensors='pt').input_ids.to("cuda")
        ref_ids = tokenizer(ref, return_tensors='pt').input_ids.to("cuda")
        unpref_ids = tokenizer(unpref, return_tensors='pt').input_ids.to("cuda")

        logprob_ref = model(input_ids=input_ids, labels=ref_ids).loss
        logprob_unpref = model(input_ids=input_ids, labels=unpref_ids).loss
        dpo_loss = -torch.log(torch.sigmoid(torch.tensor(logprob_unpref.item() - logprob_ref.item())))
        loss_dpo_batch.append(dpo_loss)

        gen_ids = model.generate(input_ids)
        gen_text = tokenizer.decode(gen_ids[0], skip_special_tokens=True)
        rho = compute_similarity(gen_text, ref)
        reward = hyper_gamma_reward(rho)

        labels = tokenizer(gen_text, return_tensors='pt').input_ids.to("cuda")
        logprob = model(input_ids=input_ids, labels=labels).loss
        hgrl_loss = -reward * logprob
        loss_hgrl_batch.append(hgrl_loss)

    loss_dpo_mean = torch.stack(loss_dpo_batch).mean()
    loss_hgrl_mean = torch.stack(loss_hgrl_batch).mean()
    combined_loss = ALPHA * loss_dpo_mean + BETA * loss_hgrl_mean
    combined_loss.backward()
    optimizer.step()
    optimizer.zero_grad()
    final_losses.append(combined_loss.item())

# ========== SAVE OUTPUTS ==========
plt.plot(dpo_losses, label="DPO")
plt.plot(hgrl_losses, label="HGRL")
plt.plot(final_losses, label="Combined")
plt.xlabel("Steps")
plt.ylabel("Loss")
plt.legend()
plt.savefig("loss_curve.png")

with open("loss_report.txt", "w") as f:
    f.write("DPO Final Loss: {:.4f}\n".format(dpo_losses[-1]))
    f.write("HGRL Final Loss: {:.4f}\n".format(hgrl_losses[-1]))
    f.write("Combined Final Loss: {:.4f}\n".format(final_losses[-1]))

# ========== TEST AND SAVE TRANSLATIONS ==========
test_df = pd.read_excel("in22conv.xlsx")
test_outputs = []
for line in test_df.iloc[:, 0].tolist():
    input_ids = tokenizer(line, return_tensors='pt').input_ids.to("cuda")
    outputs = model.generate(input_ids)
    translation = tokenizer.decode(outputs[0], skip_special_tokens=True)
    test_outputs.append(translation)

output_df = pd.DataFrame({"Dogri": test_df.iloc[:, 0], "English": test_outputs})
output_df.to_excel("translated_output.xlsx", index=False)


from collections import OrderedDict
from typing import Any, Mapping, Optional

from transformers import PreTrainedTokenizer
from transformers.configuration_utils import PretrainedConfig
from transformers.onnx import OnnxConfig, OnnxSeq2SeqConfigWithPast
from transformers.onnx.utils import compute_effective_axis_dimension
from transformers.utils import TensorType, is_torch_available


# Copied from transformers.models.m2m_100.configuration_m2m_100.M2M100Config->IndicTrans
class IndicTransConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`IT2Model`]. It is used to instantiate an
    IT2 model according to the specified arguments, defining the model architecture. Instantiating a configuration
    with the defaults will yield a similar configuration to that of the IT2

    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
    documentation from [`PretrainedConfig`] for more information.


    Args:
        vocab_size (`int`, *optional*, defaults to 50265):
            Vocabulary size of the IT2 model. Defines the number of different tokens that can be represented by the
            `inputs_ids` passed when calling [`IT2Model`] or
        d_model (`int`, *optional*, defaults to 1024):
            Dimensionality of the layers and the pooler layer.
        encoder_layers (`int`, *optional*, defaults to 12):
            Number of encoder layers.
        decoder_layers (`int`, *optional*, defaults to 12):
            Number of decoder layers.
        encoder_attention_heads (`int`, *optional*, defaults to 16):
            Number of attention heads for each attention layer in the Transformer encoder.
        decoder_attention_heads (`int`, *optional*, defaults to 16):
            Number of attention heads for each attention layer in the Transformer decoder.
        decoder_ffn_dim (`int`, *optional*, defaults to 4096):
            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
        encoder_ffn_dim (`int`, *optional*, defaults to 4096):
            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
        activation_function (`str` or `function`, *optional*, defaults to `"gelu"`):
            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
            `"relu"`, `"silu"` and `"gelu_new"` are supported.
        dropout (`float`, *optional*, defaults to 0.1):
            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
        attention_dropout (`float`, *optional*, defaults to 0.0):
            The dropout ratio for the attention probabilities.
        activation_dropout (`float`, *optional*, defaults to 0.0):
            The dropout ratio for activations inside the fully connected layer.
        classifier_dropout (`float`, *optional*, defaults to 0.0):
            The dropout ratio for classifier.
        max_position_embeddings (`int`, *optional*, defaults to 1024):
            The maximum sequence length that this model might ever be used with. Typically set this to something large
            just in case (e.g., 512 or 1024 or 2048).
        init_std (`float`, *optional*, defaults to 0.02):
            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
        encoder_layerdrop (`float`, *optional*, defaults to 0.0):
            The LayerDrop probability for the encoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
            for more details.
        decoder_layerdrop (`float`, *optional*, defaults to 0.0):
            The LayerDrop probability for the decoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
            for more details.
        use_cache (`bool`, *optional*, defaults to `True`):
            Whether or not the model should return the last key/values attentions (not used by all models).
    ```"""
    model_type = "IndicTrans"
    keys_to_ignore_at_inference = ["past_key_values"]
    attribute_map = {
        "num_attention_heads": "encoder_attention_heads",
        "hidden_size": "d_model",
    }

    def __init__(
        self,
        encoder_vocab_size=None,
        decoder_vocab_size=None,
        encoder_embed_dim=512,
        decoder_embed_dim=512,
        max_source_positions=210,
        max_target_positions=210,
        encoder_layers=6,
        encoder_ffn_dim=2048,
        encoder_attention_heads=8,
        decoder_layers=6,
        decoder_ffn_dim=2048,
        decoder_attention_heads=8,
        encoder_layerdrop=0.00,
        decoder_layerdrop=0.00,
        use_cache=True,
        is_encoder_decoder=True,
        activation_function="relu",
        encoder_normalize_before=False,
        decoder_normalize_before=False,
        layernorm_embedding=False,
        share_decoder_input_output_embed=False,
        dropout=0.1,
        attention_dropout=0.0,
        activation_dropout=0.0,
        init_std=0.02,
        scale_embedding=True,
        decoder_start_token_id=2,
        pad_token_id=1,
        bos_token_id=0,
        eos_token_id=2,
        attn_implementation="eager",
        **kwargs,
    ):
        self.encoder_vocab_size = encoder_vocab_size
        self.decoder_vocab_size = decoder_vocab_size
        self.encoder_normalize_before = encoder_normalize_before
        self.decoder_normalize_before = decoder_normalize_before
        self.layernorm_embedding = layernorm_embedding
        self.max_source_positions = max_source_positions
        self.max_target_positions = max_target_positions
        self.encoder_embed_dim = encoder_embed_dim
        self.decoder_embed_dim = decoder_embed_dim
        self.encoder_ffn_dim = encoder_ffn_dim
        self.encoder_layers = encoder_layers
        self.encoder_attention_heads = encoder_attention_heads
        self.decoder_ffn_dim = decoder_ffn_dim
        self.decoder_layers = decoder_layers
        self.decoder_attention_heads = decoder_attention_heads
        self.dropout = dropout
        self.attention_dropout = attention_dropout
        self.activation_dropout = activation_dropout
        self.activation_function = activation_function
        self.init_std = init_std
        self.encoder_layerdrop = encoder_layerdrop
        self.decoder_layerdrop = decoder_layerdrop
        self.use_cache = use_cache
        self.num_hidden_layers = encoder_layers
        self.scale_embedding = scale_embedding
        self.share_decoder_input_output_embed = share_decoder_input_output_embed
        self.attn_implementation = attn_implementation
        
        super().__init__(
            pad_token_id=pad_token_id,
            bos_token_id=bos_token_id,
            eos_token_id=eos_token_id,
            is_encoder_decoder=is_encoder_decoder,
            decoder_start_token_id=decoder_start_token_id,
            **kwargs,
        )


class IndicTransOnnxConfig(OnnxSeq2SeqConfigWithPast):
    @property
    def inputs(self) -> Mapping[str, Mapping[int, str]]:
        common_inputs = OrderedDict(
            [
                ("input_ids", {0: "batch", 1: "encoder_sequence"}),
                ("attention_mask", {0: "batch", 1: "encoder_sequence"}),
            ]
        )

        if self.use_past:
            common_inputs["decoder_input_ids"] = {0: "batch"}
            common_inputs["decoder_attention_mask"] = {
                0: "batch",
                1: "past_decoder_sequence + sequence",
            }
        else:
            common_inputs["decoder_input_ids"] = {0: "batch", 1: "decoder_sequence"}
            common_inputs["decoder_attention_mask"] = {
                0: "batch",
                1: "decoder_sequence",
            }

        if self.use_past:
            self.fill_with_past_key_values_(common_inputs, direction="inputs")
        return common_inputs

    # Copied from BartOnnxConfig._generate_dummy_inputs_for_sequence_classification_and_question_answering
    # A better name would be _generate_dummy_inputs_for_encoder_and_decoder because sequence classification and question
    # answering are not supported for IT2, but this name is preserved to be able to check that the copy matches what
    # was done for BART so that it can be updated if need be.
    def _generate_dummy_inputs_for_sequence_classification_and_question_answering(
        self,
        tokenizer: PreTrainedTokenizer,
        batch_size: int = -1,
        seq_length: int = -1,
        is_pair: bool = False,
        framework: Optional[TensorType] = None,
    ) -> Mapping[str, Any]:
        # Copied from OnnxConfig.generate_dummy_inputs
        # Did not use super(OnnxConfigWithPast, self).generate_dummy_inputs for code clarity.
        # If dynamic axis (-1) we forward with a fixed dimension of 2 samples to avoid optimizations made by ONNX
        batch_size = compute_effective_axis_dimension(
            batch_size,
            fixed_dimension=OnnxConfig.default_fixed_batch,
            num_token_to_add=0,
        )

        # If dynamic axis (-1) we forward with a fixed dimension of 8 tokens to avoid optimizations made by ONNX
        token_to_add = tokenizer.num_special_tokens_to_add(is_pair)
        seq_length = compute_effective_axis_dimension(
            seq_length,
            fixed_dimension=OnnxConfig.default_fixed_sequence,
            num_token_to_add=token_to_add,
        )

        # Generate dummy inputs according to compute batch and sequence
        dummy_input = [" ".join([tokenizer.unk_token]) * seq_length] * batch_size
        common_inputs = dict(tokenizer(dummy_input, return_tensors=framework))
        return common_inputs

    # Copied from transformers.models.bart.configuration_bart.BartOnnxConfig._generate_dummy_inputs_for_default_and_seq2seq_lm
    def _generate_dummy_inputs_for_default_and_seq2seq_lm(
        self,
        tokenizer: PreTrainedTokenizer,
        batch_size: int = -1,
        seq_length: int = -1,
        is_pair: bool = False,
        framework: Optional[TensorType] = None,
    ) -> Mapping[str, Any]:
        encoder_inputs = self._generate_dummy_inputs_for_sequence_classification_and_question_answering(
            tokenizer, batch_size, seq_length, is_pair, framework
        )

        # Generate decoder inputs
        decoder_seq_length = seq_length if not self.use_past else 1
        decoder_inputs = self._generate_dummy_inputs_for_sequence_classification_and_question_answering(
            tokenizer, batch_size, decoder_seq_length, is_pair, framework
        )
        decoder_inputs = {
            f"decoder_{name}": tensor for name, tensor in decoder_inputs.items()
        }
        common_inputs = dict(**encoder_inputs, **decoder_inputs)

        if self.use_past:
            if not is_torch_available():
                raise ValueError(
                    "Cannot generate dummy past_keys inputs without PyTorch installed."
                )
            else:
                import torch
            batch, encoder_seq_length = common_inputs["input_ids"].shape
            decoder_seq_length = common_inputs["decoder_input_ids"].shape[1]
            (
                num_encoder_attention_heads,
                num_decoder_attention_heads,
            ) = self.num_attention_heads
            encoder_shape = (
                batch,
                num_encoder_attention_heads,
                encoder_seq_length,
                self._config.hidden_size // num_encoder_attention_heads,
            )
            decoder_past_length = decoder_seq_length + 3
            decoder_shape = (
                batch,
                num_decoder_attention_heads,
                decoder_past_length,
                self._config.hidden_size // num_decoder_attention_heads,
            )

            common_inputs["decoder_attention_mask"] = torch.cat(
                [
                    common_inputs["decoder_attention_mask"],
                    torch.ones(batch, decoder_past_length),
                ],
                dim=1,
            )

            common_inputs["past_key_values"] = []
            # If the number of encoder and decoder layers are present in the model configuration, both are considered
            num_encoder_layers, num_decoder_layers = self.num_layers
            min_num_layers = min(num_encoder_layers, num_decoder_layers)
            max_num_layers = (
                max(num_encoder_layers, num_decoder_layers) - min_num_layers
            )
            remaining_side_name = (
                "encoder" if num_encoder_layers > num_decoder_layers else "decoder"
            )

            for _ in range(min_num_layers):
                common_inputs["past_key_values"].append(
                    (
                        torch.zeros(decoder_shape),
                        torch.zeros(decoder_shape),
                        torch.zeros(encoder_shape),
                        torch.zeros(encoder_shape),
                    )
                )
            # TODO: test this.
            shape = encoder_shape if remaining_side_name == "encoder" else decoder_shape
            for _ in range(min_num_layers, max_num_layers):
                common_inputs["past_key_values"].append(
                    (torch.zeros(shape), torch.zeros(shape))
                )
        return common_inputs

    generate_dummy_inputs = _generate_dummy_inputs_for_default_and_seq2seq_lm

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

import torch
import torch.nn as nn
from torch.nn import functional as F

from transformers.activations import ACT2FN

from transformers.modeling_attn_mask_utils import (
    _prepare_4d_attention_mask,
    _prepare_4d_attention_mask_for_sdpa,
    _prepare_4d_causal_attention_mask,
    _prepare_4d_causal_attention_mask_for_sdpa,
)

from transformers.integrations.deepspeed import is_deepspeed_zero3_enabled
from transformers.modeling_outputs import (
    BaseModelOutput,
    BaseModelOutputWithPastAndCrossAttentions,
    Seq2SeqLMOutput,
    Seq2SeqModelOutput
)

from transformers.utils import (
    logging,
    is_flash_attn_2_available,
    is_flash_attn_greater_or_equal_2_10,
)
 
from transformers.modeling_utils import PreTrainedModel
from transformers.generation.utils import GenerationMixin

from .configuration_indictrans import IndicTransConfig


logger = logging.get_logger(__name__)

INDICTRANS_PRETRAINED_MODEL_ARCHIVE_LIST = [""]

try:
    if is_flash_attn_2_available():
        from flash_attn import flash_attn_func, flash_attn_varlen_func
        from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa    
except:
    pass


# Copied from transformers.models.llama.modeling_llama._get_unpad_data
def _get_unpad_data(attention_mask):
    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
    max_seqlen_in_batch = seqlens_in_batch.max().item()
    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
    return (
        indices,
        cu_seqlens,
        max_seqlen_in_batch,
    )


# Copied from transformers.models.bart.modeling_bart.shift_tokens_right
def shift_tokens_right(
    input_ids: torch.Tensor, pad_token_id: int, decoder_start_token_id: int
):
    """
    Shift input ids one token to the right.
    """
    shifted_input_ids = input_ids.new_zeros(input_ids.shape)
    shifted_input_ids[:, 1:] = input_ids[:, :-1].clone()
    shifted_input_ids[:, 0] = decoder_start_token_id

    if pad_token_id is None:
        raise ValueError("self.model.config.pad_token_id has to be defined.")
    # replace possible -100 values in labels by `pad_token_id`
    shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)

    return shifted_input_ids


def create_position_ids_from_input_ids(
    input_ids, padding_idx, past_key_values_length=0
):
    """
    Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding symbols
    are ignored. This is modified from fairseq's `utils.make_positions`.
    """
    # The series of casts and type-conversions here are carefully balanced to both work with ONNX export and XLA.
    mask = input_ids.ne(padding_idx).int()
    incremental_indices = (
        torch.cumsum(mask, dim=1).type_as(mask) + past_key_values_length
    ) * mask
    return incremental_indices.long() + padding_idx


# Copied from transformers.models.m2m_100.modeling_m2m_100.M2M100SinusoidalPositionalEmbedding->IndicTrans
class IndicTransSinusoidalPositionalEmbedding(nn.Module):
    """This module produces sinusoidal positional embeddings of any length."""

    def __init__(
        self, num_positions: int, embedding_dim: int, padding_idx: Optional[int] = None
    ):
        super().__init__()
        self.offset = 2
        self.embedding_dim = embedding_dim
        self.padding_idx = padding_idx
        self.make_weights(num_positions + self.offset, embedding_dim, padding_idx)

    def make_weights(
        self, num_embeddings: int, embedding_dim: int, padding_idx: Optional[int] = None
    ):
        emb_weights = self.get_embedding(num_embeddings, embedding_dim, padding_idx)
        if hasattr(self, "weights"):
            # in forward put the weights on the correct dtype and device of the param
            emb_weights = emb_weights.to(
                dtype=self.weights.dtype, device=self.weights.device
            )

        self.register_buffer("weights", emb_weights, persistent=False)

    @staticmethod
    def get_embedding(
        num_embeddings: int, embedding_dim: int, padding_idx: Optional[int] = None
    ):
        """
        Build sinusoidal embeddings.

        This matches the implementation in tensor2tensor, but differs slightly from the description in Section 3.5 of
        "Attention Is All You Need".
        """
        half_dim = embedding_dim // 2
        emb = math.log(10000) / (half_dim - 1)
        emb = torch.exp(torch.arange(half_dim, dtype=torch.float) * -emb)
        emb = torch.arange(num_embeddings, dtype=torch.float).unsqueeze(
            1
        ) * emb.unsqueeze(0)
        emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1).view(
            num_embeddings, -1
        )
        if embedding_dim % 2 == 1:
            # zero pad
            emb = torch.cat([emb, torch.zeros(num_embeddings, 1)], dim=1)
        if padding_idx is not None:
            emb[padding_idx, :] = 0

        return emb.to(torch.get_default_dtype())

    @torch.no_grad()
    def forward(
        self,
        input_ids: torch.Tensor = None,
        inputs_embeds: torch.Tensor = None,
        past_key_values_length: int = 0,
    ):
        if input_ids is not None:
            bsz, seq_len = input_ids.size()
            # Create the position ids from the input token ids. Any padded tokens remain padded.
            position_ids = create_position_ids_from_input_ids(
                input_ids, self.padding_idx, past_key_values_length
            ).to(input_ids.device)
        else:
            bsz, seq_len = inputs_embeds.size()[:-1]
            position_ids = self.create_position_ids_from_inputs_embeds(
                inputs_embeds, past_key_values_length
            )

        # expand embeddings if needed
        max_pos = self.padding_idx + 1 + seq_len + past_key_values_length
        if max_pos > self.weights.size(0):
            self.make_weights(
                max_pos + self.offset, self.embedding_dim, self.padding_idx
            )

        return (
            self.weights.index_select(0, position_ids.view(-1))
            .view(bsz, seq_len, self.weights.shape[-1])
            .detach()
        )

    def create_position_ids_from_inputs_embeds(
        self, inputs_embeds, past_key_values_length
    ):
        """
        We are provided embeddings directly. We cannot infer which are padded so just generate sequential position ids.

        Args:
            inputs_embeds: torch.Tensor

        Returns: torch.Tensor
        """
        input_shape = inputs_embeds.size()[:-1]
        sequence_length = input_shape[1]

        position_ids = torch.arange(
            self.padding_idx + 1,
            sequence_length + self.padding_idx + 1,
            dtype=torch.long,
            device=inputs_embeds.device,
        )
        return (
            position_ids.unsqueeze(0).expand(input_shape).contiguous()
            + past_key_values_length
        )


# Copied from transformers.models.bart.modeling_bart.BartAttention with Bart->IndicTrans
class IndicTransAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(
        self,
        embed_dim: int,
        num_heads: int,
        dropout: float = 0.0,
        is_decoder: bool = False,
        bias: bool = True,
        is_causal: bool = False,
        config: Optional[IndicTransConfig] = None,
    ):
        super().__init__()
        self.embed_dim = embed_dim
        self.num_heads = num_heads
        self.dropout = dropout
        self.head_dim = embed_dim // num_heads
        self.config = config

        if (self.head_dim * num_heads) != self.embed_dim:
            raise ValueError(
                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
                f" and `num_heads`: {num_heads})."
            )
        self.scaling = self.head_dim**-0.5
        self.is_decoder = is_decoder
        self.is_causal = is_causal

        self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)

    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
        return (
            tensor.view(bsz, seq_len, self.num_heads, self.head_dim)
            .transpose(1, 2)
            .contiguous()
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        key_value_states: Optional[torch.Tensor] = None,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
        attention_mask: Optional[torch.Tensor] = None,
        layer_head_mask: Optional[torch.Tensor] = None,
        output_attentions: bool = False,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        """Input shape: Batch x Time x Channel"""

        # if key_value_states are provided this layer is used as a cross-attention layer
        # for the decoder
        is_cross_attention = key_value_states is not None

        bsz, tgt_len, _ = hidden_states.size()

        # get query proj
        query_states = self.q_proj(hidden_states) * self.scaling
        # get key, value proj
        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
        # is checking that the `sequence_length` of the `past_key_value` is the same as
        # the provided `key_value_states` to support prefix tuning
        if (
            is_cross_attention
            and past_key_value is not None
            and past_key_value[0].shape[2] == key_value_states.shape[1]
        ):
            # reuse k,v, cross_attentions
            key_states = past_key_value[0]
            value_states = past_key_value[1]
        elif is_cross_attention:
            # cross_attentions
            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
        elif past_key_value is not None:
            # reuse k, v, self_attention
            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
            key_states = torch.cat([past_key_value[0], key_states], dim=2)
            value_states = torch.cat([past_key_value[1], value_states], dim=2)
        else:
            # self_attention
            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)

        if self.is_decoder:
            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
            # Further calls to cross_attention layer can then reuse all cross-attention
            # key/value_states (first "if" case)
            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
            # all previous decoder key/value_states. Further calls to uni-directional self-attention
            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
            # if encoder bi-directional self-attention `past_key_value` is always `None`
            past_key_value = (key_states, value_states)

        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
        query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
        key_states = key_states.reshape(*proj_shape)
        value_states = value_states.reshape(*proj_shape)

        src_len = key_states.size(1)
        attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))

        if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
            raise ValueError(
                f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
                f" {attn_weights.size()}"
            )

        if attention_mask is not None:
            if attention_mask.size() != (bsz, 1, tgt_len, src_len):
                raise ValueError(
                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
                )
            attn_weights = (
                attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
                + attention_mask
            )
            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)

        attn_weights = F.softmax(attn_weights, dim=-1)

        if layer_head_mask is not None:
            if layer_head_mask.size() != (self.num_heads,):
                raise ValueError(
                    f"Head mask for a single layer should be of size {(self.num_heads,)}, but is"
                    f" {layer_head_mask.size()}"
                )
            attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(
                bsz, self.num_heads, tgt_len, src_len
            )
            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)

        if output_attentions:
            # this operation is a bit awkward, but it's required to
            # make sure that attn_weights keeps its gradient.
            # In order to do so, attn_weights have to be reshaped
            # twice and have to be reused in the following
            attn_weights_reshaped = attn_weights.view(
                bsz, self.num_heads, tgt_len, src_len
            )
            attn_weights = attn_weights_reshaped.view(
                bsz * self.num_heads, tgt_len, src_len
            )
        else:
            attn_weights_reshaped = None

        attn_probs = F.dropout(attn_weights, p=self.dropout, training=self.training)

        attn_output = torch.bmm(attn_probs, value_states)

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

        attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
        attn_output = attn_output.transpose(1, 2)

        # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
        # partitioned across GPUs when using tensor-parallelism.
        attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)

        attn_output = self.out_proj(attn_output)

        return attn_output, attn_weights_reshaped, past_key_value
    

class IndicTransFlashAttention2(IndicTransAttention):
    """
    IndicTrans flash attention module. This module inherits from `IndicTransAttention` 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 _reshape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim)

    def forward(
        self,
        hidden_states: torch.Tensor,
        key_value_states: Optional[torch.Tensor] = None,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
        attention_mask: Optional[torch.Tensor] = None,
        layer_head_mask: Optional[torch.Tensor] = None,
        output_attentions: bool = False,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        # IndicTransFlashAttention2 attention does not support output_attentions
        if output_attentions:
            raise ValueError("IndicTransFlashAttention2 attention does not support output_attentions")

        # if key_value_states are provided this layer is used as a cross-attention layer
        # for the decoder
        is_cross_attention = key_value_states is not None

        bsz, q_len, _ = hidden_states.size()

        # get query proj
        query_states = self._reshape(self.q_proj(hidden_states), -1, bsz)
        # get key, value proj
        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
        # is checking that the `sequence_length` of the `past_key_value` is the same as
        # the provided `key_value_states` to support prefix tuning
        if (
            is_cross_attention
            and past_key_value is not None
            and past_key_value[0].shape[2] == key_value_states.shape[1]
        ):
            # reuse k,v, cross_attentions
            key_states = past_key_value[0].transpose(1, 2)
            value_states = past_key_value[1].transpose(1, 2)
        elif is_cross_attention:
            # cross_attentions
            key_states = self._reshape(self.k_proj(key_value_states), -1, bsz)
            value_states = self._reshape(self.v_proj(key_value_states), -1, bsz)
        elif past_key_value is not None:
            # reuse k, v, self_attention
            key_states = self._reshape(self.k_proj(hidden_states), -1, bsz)
            value_states = self._reshape(self.v_proj(hidden_states), -1, bsz)
            key_states = torch.cat([past_key_value[0].transpose(1, 2), key_states], dim=1)
            value_states = torch.cat([past_key_value[1].transpose(1, 2), value_states], dim=1)
        else:
            # self_attention
            key_states = self._reshape(self.k_proj(hidden_states), -1, bsz)
            value_states = self._reshape(self.v_proj(hidden_states), -1, bsz)

        if self.is_decoder:
            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
            # Further calls to cross_attention layer can then reuse all cross-attention
            # key/value_states (first "if" case)
            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
            # all previous decoder key/value_states. Further calls to uni-directional self-attention
            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
            # if encoder bi-directional self-attention `past_key_value` is always `None`
            past_key_value = (key_states.transpose(1, 2), value_states.transpose(1, 2))

        kv_seq_len = key_states.shape[-2]
        if past_key_value is not None:
            kv_seq_len += past_key_value[0].shape[-2]

        # 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. (LlamaRMSNorm handles it correctly)

        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)

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

        attn_output = attn_output.reshape(bsz, q_len, -1)
        attn_output = self.out_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value

    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2._flash_attention_forward
    def _flash_attention_forward(
        self, query_states, key_states, value_states, attention_mask, query_length, dropout=0.0, softmax_scale=None
    ):
        """
        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)
        """
        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

            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,
            )

            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
        else:
            attn_output = flash_attn_func(
                query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal
            )

        return attn_output

    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2._upad_input
    def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
        batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape

        key_layer = index_first_axis(
            key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
        )
        value_layer = index_first_axis(
            value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
        )
        if query_length == kv_seq_len:
            query_layer = index_first_axis(
                query_layer.reshape(batch_size * kv_seq_len, self.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),
        )


class IndicTransSdpaAttention(IndicTransAttention):
    def forward(
        self,
        hidden_states: torch.Tensor,
        key_value_states: Optional[torch.Tensor] = None,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
        attention_mask: Optional[torch.Tensor] = None,
        layer_head_mask: Optional[torch.Tensor] = None,
        output_attentions: bool = False,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        """Input shape: Batch x Time x Channel"""
        if output_attentions or layer_head_mask is not None:
            # TODO: Improve this warning with e.g. `model.config._attn_implementation = "manual"` once this is implemented.
            logger.warning_once(
                "IndicTransModel is using IndicTransSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True` or `layer_head_mask` not None. 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,
                key_value_states=key_value_states,
                past_key_value=past_key_value,
                attention_mask=attention_mask,
                layer_head_mask=layer_head_mask,
                output_attentions=output_attentions,
            )

        # if key_value_states are provided this layer is used as a cross-attention layer
        # for the decoder
        is_cross_attention = key_value_states is not None

        bsz, tgt_len, _ = hidden_states.size()

        # get query proj
        query_states = self.q_proj(hidden_states)
        # get key, value proj
        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
        # is checking that the `sequence_length` of the `past_key_value` is the same as
        # the provided `key_value_states` to support prefix tuning
        if (
            is_cross_attention
            and past_key_value is not None
            and past_key_value[0].shape[2] == key_value_states.shape[1]
        ):
            # reuse k,v, cross_attentions
            key_states = past_key_value[0]
            value_states = past_key_value[1]
        elif is_cross_attention:
            # cross_attentions
            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
        elif past_key_value is not None:
            # reuse k, v, self_attention
            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
            key_states = torch.cat([past_key_value[0], key_states], dim=2)
            value_states = torch.cat([past_key_value[1], value_states], dim=2)
        else:
            # self_attention
            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)

        if self.is_decoder:
            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
            # Further calls to cross_attention layer can then reuse all cross-attention
            # key/value_states (first "if" case)
            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
            # all previous decoder key/value_states. Further calls to uni-directional self-attention
            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
            # if encoder bi-directional self-attention `past_key_value` is always `None`
            past_key_value = (key_states, value_states)

        query_states = self._shape(query_states, tgt_len, bsz)

        # NOTE: SDPA with memory-efficient backend is currently (torch==2.1.2) bugged when using non-contiguous inputs and a custom attn_mask,
        # but we are fine here as `_shape` do call `.contiguous()`. Reference: https://github.com/pytorch/pytorch/issues/112577
        attn_output = F.scaled_dot_product_attention(
            query_states,
            key_states,
            value_states,
            attn_mask=attention_mask,
            dropout_p=self.dropout if self.training else 0.0,
            # The tgt_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case tgt_len == 1.
            is_causal=self.is_causal and attention_mask is None and tgt_len > 1,
        )

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

        attn_output = attn_output.transpose(1, 2)

        # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
        # partitioned across GPUs when using tensor-parallelism.
        attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)

        attn_output = self.out_proj(attn_output)

        return attn_output, None, past_key_value


INDICTRANS_ATTENTION_CLASSES = {
    "eager": IndicTransAttention,
    "sdpa": IndicTransSdpaAttention,
    "flash_attention_2": IndicTransFlashAttention2,
}

# Copied from transformers.models.mbart.modeling_mbart.MBartEncoderLayer with MBart->IndicTrans
class IndicTransEncoderLayer(nn.Module):
    def __init__(self, config: IndicTransConfig):
        super().__init__()
        self.embed_dim = config.encoder_embed_dim
        self.self_attn = INDICTRANS_ATTENTION_CLASSES[config._attn_implementation](
            embed_dim=self.embed_dim,
            num_heads=config.encoder_attention_heads,
            dropout=config.attention_dropout,
            config=config,
        )
        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
        self.dropout = config.dropout
        self.activation_fn = ACT2FN[config.activation_function]
        self.activation_dropout = config.activation_dropout
        self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim)
        self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim)
        self.final_layer_norm = nn.LayerNorm(self.embed_dim)
        self.normalize_before = config.encoder_normalize_before

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor,
        layer_head_mask: torch.Tensor,
        output_attentions: bool = False,
    ) -> torch.Tensor:
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`): attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
            layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size
                `(encoder_attention_heads,)`.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
        """
        residual = hidden_states
        if self.normalize_before:
            hidden_states = self.self_attn_layer_norm(hidden_states)
        hidden_states, attn_weights, _ = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            layer_head_mask=layer_head_mask,
            output_attentions=output_attentions,
        )
        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states
        if not self.normalize_before:
            hidden_states = self.self_attn_layer_norm(hidden_states)

        residual = hidden_states
        if self.normalize_before:
            hidden_states = self.final_layer_norm(hidden_states)
        hidden_states = self.activation_fn(self.fc1(hidden_states))
        hidden_states = F.dropout(
            hidden_states, p=self.activation_dropout, training=self.training
        )
        hidden_states = self.fc2(hidden_states)
        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states
        if not self.normalize_before:
            hidden_states = self.final_layer_norm(hidden_states)

        if hidden_states.dtype == torch.float16 and (
            torch.isinf(hidden_states).any() or torch.isnan(hidden_states).any()
        ):
            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
            hidden_states = torch.clamp(
                hidden_states, min=-clamp_value, max=clamp_value
            )

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (attn_weights,)

        return outputs


# Copied from transformers.models.mbart.modeling_mbart.MBartDecoderLayer with MBart->IndicTrans
class IndicTransDecoderLayer(nn.Module):
    def __init__(self, config: IndicTransConfig):
        super().__init__()
        self.embed_dim = config.decoder_embed_dim

        self.self_attn = INDICTRANS_ATTENTION_CLASSES[config._attn_implementation](
            embed_dim=self.embed_dim,
            num_heads=config.decoder_attention_heads,
            dropout=config.attention_dropout,
            is_decoder=True,
            is_causal=True,
            config=config,
        )
        self.dropout = config.dropout
        self.activation_fn = ACT2FN[config.activation_function]
        self.activation_dropout = config.activation_dropout

        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
        self.encoder_attn = INDICTRANS_ATTENTION_CLASSES[config._attn_implementation](
            self.embed_dim,
            config.decoder_attention_heads,
            dropout=config.attention_dropout,
            is_decoder=True,
            config=config,
        )
        self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim)
        self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim)
        self.fc2 = nn.Linear(config.decoder_ffn_dim, self.embed_dim)
        self.final_layer_norm = nn.LayerNorm(self.embed_dim)
        self.normalize_before = config.decoder_normalize_before

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
        layer_head_mask: Optional[torch.Tensor] = None,
        cross_attn_layer_head_mask: Optional[torch.Tensor] = None,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = True,
    ) -> torch.Tensor:
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`): attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
            encoder_hidden_states (`torch.FloatTensor`):
                cross attention input to the layer of shape `(batch, seq_len, embed_dim)`
            encoder_attention_mask (`torch.FloatTensor`): encoder attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
            layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size
                `(encoder_attention_heads,)`.
            cross_attn_layer_head_mask (`torch.FloatTensor`): mask for cross-attention heads in a given layer of
                size `(decoder_attention_heads,)`.
            past_key_value (`Tuple(torch.FloatTensor)`): cached past key and value projection states
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
        """
        residual = hidden_states
        if self.normalize_before:
            hidden_states = self.self_attn_layer_norm(hidden_states)

        # Self Attention
        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
        self_attn_past_key_value = (
            past_key_value[:2] if past_key_value is not None else None
        )
        # add present self-attn cache to positions 1,2 of present_key_value tuple
        hidden_states, self_attn_weights, present_key_value = self.self_attn(
            hidden_states=hidden_states,
            past_key_value=self_attn_past_key_value,
            attention_mask=attention_mask,
            layer_head_mask=layer_head_mask,
            output_attentions=output_attentions,
        )
        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states
        if not self.normalize_before:
            hidden_states = self.self_attn_layer_norm(hidden_states)

        # Cross-Attention Block
        cross_attn_present_key_value = None
        cross_attn_weights = None
        if encoder_hidden_states is not None:
            residual = hidden_states
            if self.normalize_before:
                hidden_states = self.encoder_attn_layer_norm(hidden_states)

            # cross_attn cached key/values tuple is at positions 3,4 of present_key_value tuple
            cross_attn_past_key_value = (
                past_key_value[-2:] if past_key_value is not None else None
            )
            (
                hidden_states,
                cross_attn_weights,
                cross_attn_present_key_value,
            ) = self.encoder_attn(
                hidden_states=hidden_states,
                key_value_states=encoder_hidden_states,
                attention_mask=encoder_attention_mask,
                layer_head_mask=cross_attn_layer_head_mask,
                past_key_value=cross_attn_past_key_value,
                output_attentions=output_attentions,
            )
            hidden_states = F.dropout(
                hidden_states, p=self.dropout, training=self.training
            )
            hidden_states = residual + hidden_states
            if not self.normalize_before:
                hidden_states = self.encoder_attn_layer_norm(hidden_states)

            # add cross-attn to positions 3,4 of present_key_value tuple
            present_key_value = present_key_value + cross_attn_present_key_value

        # Fully Connected
        residual = hidden_states
        if self.normalize_before:
            hidden_states = self.final_layer_norm(hidden_states)
        hidden_states = self.activation_fn(self.fc1(hidden_states))
        hidden_states = F.dropout(
            hidden_states, p=self.activation_dropout, training=self.training
        )
        hidden_states = self.fc2(hidden_states)
        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states
        if not self.normalize_before:
            hidden_states = self.final_layer_norm(hidden_states)

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights, cross_attn_weights)

        if use_cache:
            outputs += (present_key_value,)

        return outputs


# Copied from transformers.models.m2m_100.modeling_m2m_100.M2M100PretrainedModel->IndicTrans
class IndicTransPreTrainedModel(PreTrainedModel):
    config_class = IndicTransConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["IndicTransAttention"]

    def _init_weights(self, module):
        std = self.config.init_std
        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 _set_gradient_checkpointing(self, module, value=False):
        if isinstance(module, (IndicTransDecoder, IndicTransEncoder)):
            module.gradient_checkpointing = value


# Copied from transformers.models.m2m_100.modeling_m2m_100.M2M100EncoderLayer->IndicTrans
class IndicTransEncoder(IndicTransPreTrainedModel):
    """
    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
    [`IndicTransEncoderLayer`].

    Args:
        config: IndicTransConfig
        embed_tokens (nn.Embedding): output embedding
    """

    def __init__(
        self, config: IndicTransConfig, embed_tokens: Optional[nn.Embedding] = None
    ):
        super().__init__(config)

        self.dropout = config.dropout
        self.layerdrop = config.encoder_layerdrop

        embed_dim = config.encoder_embed_dim
        self.padding_idx = config.pad_token_id
        self.max_source_positions = config.max_source_positions
        self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0

        self.embed_tokens = nn.Embedding(
            config.encoder_vocab_size, embed_dim, self.padding_idx
        )

        if embed_tokens is not None:
            self.embed_tokens.weight = embed_tokens.weight

        self.embed_positions = IndicTransSinusoidalPositionalEmbedding(
            config.max_source_positions,
            embed_dim,
            self.padding_idx,
        )
        self.layers = nn.ModuleList(
            [IndicTransEncoderLayer(config) for _ in range(config.encoder_layers)]
        )
        self.layer_norm = (
            nn.LayerNorm(embed_dim) if config.encoder_normalize_before else None
        )
        self.layernorm_embedding = (
            nn.LayerNorm(embed_dim) if config.layernorm_embedding else None
        )

        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
        self._use_sdpa = config._attn_implementation == "sdpa"

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

    def forward(
        self,
        input_ids: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ):
        r"""
        Args:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
                provide it.

                Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
                [`PreTrainedTokenizer.__call__`] for details.

                [What are input IDs?](../glossary#input-ids)
            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.

                [What are attention masks?](../glossary#attention-mask)
            head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:

                - 1 indicates the head is **not masked**,
                - 0 indicates the head is **masked**.

            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
                than the model's internal embedding lookup matrix.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            output_hidden_states (`bool`, *optional*):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more detail.
            return_dict (`bool`, *optional*):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
        """
        output_attentions = (
            output_attentions
            if output_attentions is not None
            else self.config.output_attentions
        )
        output_hidden_states = (
            output_hidden_states
            if output_hidden_states is not None
            else self.config.output_hidden_states
        )
        return_dict = (
            return_dict if return_dict is not None else self.config.use_return_dict
        )

        # retrieve input_ids and inputs_embeds
        if input_ids is not None and inputs_embeds is not None:
            raise ValueError(
                "You cannot specify both input_ids and inputs_embeds at the same time"
            )
        elif input_ids is not None:
            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
            input_shape = input_ids.size()
            input_ids = input_ids.view(-1, input_shape[-1])
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

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

        embed_pos = self.embed_positions(input_ids, inputs_embeds)
        embed_pos = embed_pos.to(inputs_embeds.device)

        hidden_states = inputs_embeds + embed_pos
        if self.layernorm_embedding is not None:
            hidden_states = self.layernorm_embedding(hidden_states)
        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)

        if attention_mask is not None:
            if self._use_flash_attention_2:
                attention_mask = attention_mask if 0 in attention_mask else None
            elif self._use_sdpa and head_mask is None and not output_attentions:
                # output_attentions=True & head_mask can not be supported when using SDPA, fall back to
                # the manual implementation that requires a 4D causal mask in all cases.
                # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
                attention_mask = _prepare_4d_attention_mask_for_sdpa(attention_mask, inputs_embeds.dtype)
            else:
                # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
                attention_mask = _prepare_4d_attention_mask(attention_mask, inputs_embeds.dtype)


        encoder_states = () if output_hidden_states else None
        all_attentions = () if output_attentions else None

        # check if head_mask has a correct number of layers specified if desired
        if head_mask is not None:
            if head_mask.size()[0] != len(self.layers):
                raise ValueError(
                    f"The head_mask should be specified for {len(self.layers)} layers, but it is for"
                    f" {head_mask.size()[0]}."
                )
        deepspeed_zero3_is_enabled = is_deepspeed_zero3_enabled()

        for idx, encoder_layer in enumerate(self.layers):
            if output_hidden_states:
                encoder_states = encoder_states + (hidden_states,)

            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
            dropout_probability = torch.rand([])

            skip_the_layer = (
                True
                if self.training and (dropout_probability < self.layerdrop)
                else False
            )
            if not skip_the_layer or deepspeed_zero3_is_enabled:
                # under deepspeed zero3 all gpus must run in sync

                if self.gradient_checkpointing and self.training:
                    # create gradient checkpointing function
                    def create_custom_forward(module):
                        def custom_forward(*inputs):
                            return module(*inputs, output_attentions)

                        return custom_forward

                    layer_outputs = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(encoder_layer),
                        hidden_states,
                        attention_mask,
                        (head_mask[idx] if head_mask is not None else None),
                    )
                else:
                    layer_outputs = encoder_layer(
                        hidden_states,
                        attention_mask,
                        layer_head_mask=(
                            head_mask[idx] if head_mask is not None else None
                        ),
                        output_attentions=output_attentions,
                    )

                hidden_states = layer_outputs[0]

            if skip_the_layer:
                layer_outputs = (None, None)

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

        if self.layer_norm is not None:
            hidden_states = self.layer_norm(hidden_states)

        if output_hidden_states:
            encoder_states = encoder_states + (hidden_states,)

        if not return_dict:
            return tuple(
                v
                for v in [hidden_states, encoder_states, all_attentions]
                if v is not None
            )
        return BaseModelOutput(
            last_hidden_state=hidden_states,
            hidden_states=encoder_states,
            attentions=all_attentions,
        )


# Copied from transformers.models.m2m_100.modeling_m2m_100.M2M100DecoderLayer->IndicTrans
class IndicTransDecoder(IndicTransPreTrainedModel):
    """
    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`IndicTransDecoderLayer`]

    Args:
        config: IndicTransConfig
        embed_tokens (nn.Embedding): output embedding
    """

    def __init__(
        self, config: IndicTransConfig, embed_tokens: Optional[nn.Embedding] = None
    ):
        super().__init__(config)
        self.dropout = config.dropout
        self.layerdrop = config.decoder_layerdrop

        embed_dim = config.encoder_embed_dim
        self.padding_idx = config.pad_token_id
        self.max_target_positions = config.max_target_positions
        self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0

        self.embed_tokens = nn.Embedding(
            config.decoder_vocab_size, embed_dim, self.padding_idx
        )

        if embed_tokens is not None:
            self.embed_tokens.weight = embed_tokens.weight

        self.embed_positions = IndicTransSinusoidalPositionalEmbedding(
            config.max_target_positions,
            embed_dim,
            self.padding_idx,
        )
        self.layers = nn.ModuleList(
            [IndicTransDecoderLayer(config) for _ in range(config.decoder_layers)]
        )
        self.layer_norm = (
            nn.LayerNorm(embed_dim) if config.decoder_normalize_before else None
        )
        self.layernorm_embedding = (
            nn.LayerNorm(embed_dim) if config.layernorm_embedding else None
        )

        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
        self._use_sdpa = config._attn_implementation == "sdpa"

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

    def forward(
        self,
        input_ids: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        cross_attn_head_mask: Optional[torch.Tensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ):
        r"""
        Args:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
                provide it.

                Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
                [`PreTrainedTokenizer.__call__`] for details.

                [What are input IDs?](../glossary#input-ids)
            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.

                [What are attention masks?](../glossary#attention-mask)
            encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*):
                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
                of the decoder.
            encoder_attention_mask (`torch.LongTensor` of shape `(batch_size, encoder_sequence_length)`, *optional*):
                Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. Mask values
                selected in `[0, 1]`:

                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.

                [What are attention masks?](../glossary#attention-mask)
            head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:

                - 1 indicates the head is **not masked**,
                - 0 indicates the head is **masked**.

            cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
                Mask to nullify selected heads of the cross-attention modules in the decoder to avoid performing
                cross-attention on hidden heads. Mask values selected in `[0, 1]`:

                - 1 indicates the head is **not masked**,
                - 0 indicates the head is **masked**.

            past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
                Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
                shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
                shape `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.

                Contains pre-computed hidden-states (key and values in the self-attention blocks and in the
                cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.

                If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
                that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
                all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of
                shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing
                `input_ids` you can choose to directly pass an embedded representation. This is useful if you want more
                control over how to convert `input_ids` indices into associated vectors than the model's internal
                embedding lookup matrix.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            output_hidden_states (`bool`, *optional*):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more detail.
            return_dict (`bool`, *optional*):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
        """
        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 decoder_input_ids and decoder_inputs_embeds at the same time"
            )
        elif input_ids is not None:
            input_shape = input_ids.size()
            input_ids = input_ids.view(-1, input_shape[-1])
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
        else:
            raise ValueError(
                "You have to specify either decoder_input_ids or decoder_inputs_embeds"
            )

        # past_key_values_length
        past_key_values_length = (
            past_key_values[0][0].shape[2] if past_key_values is not None else 0
        )

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


        if self._use_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
        elif self._use_sdpa and not output_attentions and cross_attn_head_mask is None:
            # output_attentions=True & cross_attn_head_mask can not be supported when using SDPA, and we fall back on
            # the manual implementation that requires a 4D causal mask in all cases.
            attention_mask = _prepare_4d_causal_attention_mask_for_sdpa(
                attention_mask,
                input_shape,
                inputs_embeds,
                past_key_values_length,
            )
        else:
            # 4d mask is passed through the layers
            attention_mask = _prepare_4d_causal_attention_mask(
                attention_mask, input_shape, inputs_embeds, past_key_values_length
            )

        # expand encoder attention mask
        if encoder_hidden_states is not None and encoder_attention_mask is not None:
            if self._use_flash_attention_2:
                encoder_attention_mask = encoder_attention_mask if 0 in encoder_attention_mask else None
            elif self._use_sdpa and cross_attn_head_mask is None and not output_attentions:
                # output_attentions=True & cross_attn_head_mask can not be supported when using SDPA, and we fall back on
                # the manual implementation that requires a 4D causal mask in all cases.
                # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
                encoder_attention_mask = _prepare_4d_attention_mask_for_sdpa(
                    encoder_attention_mask,
                    inputs_embeds.dtype,
                    tgt_len=input_shape[-1],
                )
            else:
                # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
                encoder_attention_mask = _prepare_4d_attention_mask(
                    encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]
                )

        # embed positions
        positions = self.embed_positions(
            input_ids, inputs_embeds, past_key_values_length
        )
        positions = positions.to(inputs_embeds.device)

        hidden_states = inputs_embeds + positions
        if self.layernorm_embedding is not None:
            hidden_states = self.layernorm_embedding(hidden_states)

        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)

        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

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

        # check if head_mask/cross_attn_head_mask has a correct number of layers specified if desired
        for attn_mask, mask_name in zip(
            [head_mask, cross_attn_head_mask], ["head_mask", "cross_attn_head_mask"]
        ):
            if attn_mask is not None:
                if attn_mask.size()[0] != len(self.layers):
                    raise ValueError(
                        f"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for"
                        f" {head_mask.size()[0]}."
                    )
        deepspeed_zero3_is_enabled = is_deepspeed_zero3_enabled()

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

            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
            dropout_probability = torch.rand([])

            skip_the_layer = (
                True
                if self.training and (dropout_probability < self.layerdrop)
                else False
            )
            if not skip_the_layer or deepspeed_zero3_is_enabled:
                # under deepspeed zero3 all gpus must run in sync

                past_key_value = (
                    past_key_values[idx] if past_key_values is not None else None
                )

                if self.gradient_checkpointing and self.training:

                    def create_custom_forward(module):
                        def custom_forward(*inputs):
                            # None for past_key_value
                            return module(*inputs, output_attentions, use_cache)

                        return custom_forward

                    layer_outputs = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(decoder_layer),
                        hidden_states,
                        attention_mask,
                        encoder_hidden_states,
                        encoder_attention_mask,
                        head_mask[idx] if head_mask is not None else None,
                        cross_attn_head_mask[idx]
                        if cross_attn_head_mask is not None
                        else None,
                        None,
                    )
                else:
                    layer_outputs = decoder_layer(
                        hidden_states,
                        attention_mask=attention_mask,
                        encoder_hidden_states=encoder_hidden_states,
                        encoder_attention_mask=encoder_attention_mask,
                        layer_head_mask=(
                            head_mask[idx] if head_mask is not None else None
                        ),
                        cross_attn_layer_head_mask=(
                            cross_attn_head_mask[idx]
                            if cross_attn_head_mask is not None
                            else None
                        ),
                        past_key_value=past_key_value,
                        output_attentions=output_attentions,
                        use_cache=use_cache,
                    )

                hidden_states = layer_outputs[0]

            if skip_the_layer:
                continue

            if use_cache:
                next_decoder_cache += (layer_outputs[3 if output_attentions else 1],)

            if output_attentions:
                all_self_attns += (layer_outputs[1],)
                all_cross_attentions += (layer_outputs[2],)

        if self.layer_norm is not None:
            hidden_states = self.layer_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,
                    all_cross_attentions,
                ]
                if v is not None
            )
        return BaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=next_cache,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
            cross_attentions=all_cross_attentions,
        )


# Copied from transformers.models.m2m_100.modeling_m2m_100.M2M100Model->IndicTrans
class IndicTransModel(IndicTransPreTrainedModel):
    _tied_weights_keys = None

    def __init__(self, config: IndicTransConfig):
        super().__init__(config)

        self.encoder = IndicTransEncoder(config)
        self.decoder = IndicTransDecoder(config)

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

    def get_encoder(self):
        return self.encoder

    def get_decoder(self):
        return self.decoder

    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        decoder_input_ids: Optional[torch.LongTensor] = None,
        decoder_attention_mask: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        decoder_head_mask: Optional[torch.Tensor] = None,
        cross_attn_head_mask: Optional[torch.Tensor] = None,
        encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        decoder_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[torch.Tensor], Seq2SeqModelOutput]:
        output_attentions = (
            output_attentions
            if output_attentions is not None
            else self.config.output_attentions
        )
        output_hidden_states = (
            output_hidden_states
            if output_hidden_states is not None
            else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = (
            return_dict if return_dict is not None else self.config.use_return_dict
        )

        if encoder_outputs is None:
            encoder_outputs = self.encoder(
                input_ids=input_ids,
                attention_mask=attention_mask,
                head_mask=head_mask,
                inputs_embeds=inputs_embeds,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                return_dict=return_dict,
            )
        # If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput when return_dict=True
        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
            encoder_outputs = BaseModelOutput(
                last_hidden_state=encoder_outputs[0],
                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
            )

        # decoder outputs consists of (dec_features, past_key_value, dec_hidden, dec_attn)
        decoder_outputs = self.decoder(
            input_ids=decoder_input_ids,
            attention_mask=decoder_attention_mask,
            encoder_hidden_states=encoder_outputs[0],
            encoder_attention_mask=attention_mask,
            head_mask=decoder_head_mask,
            cross_attn_head_mask=cross_attn_head_mask,
            past_key_values=past_key_values,
            inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        if not return_dict:
            return decoder_outputs + encoder_outputs

        return Seq2SeqModelOutput(
            last_hidden_state=decoder_outputs.last_hidden_state,
            past_key_values=decoder_outputs.past_key_values,
            decoder_hidden_states=decoder_outputs.hidden_states,
            decoder_attentions=decoder_outputs.attentions,
            cross_attentions=decoder_outputs.cross_attentions,
            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
            encoder_hidden_states=encoder_outputs.hidden_states,
            encoder_attentions=encoder_outputs.attentions,
        )


# Copied from transformers.models.m2m_100.modeling_m2m_100.M2M100ForConditionalGeneration->IndicTrans
class IndicTransForConditionalGeneration(IndicTransPreTrainedModel, GenerationMixin):
    base_model_prefix = "model"
    _tied_weights_keys = ["decoder.embed_tokens.weight", "lm_head.weight"]
    _label_smoothing = 0.0

    def __init__(self, config: IndicTransConfig):
        super().__init__(config)
        self.model = IndicTransModel(config)
        self.lm_head = nn.Linear(
            config.decoder_embed_dim, config.decoder_vocab_size, bias=False
        )

        self.post_init()

    def tie_weights(self):
        if self.config.share_decoder_input_output_embed:
            self._tie_or_clone_weights(self.model.decoder.embed_tokens, self.lm_head)

    def get_encoder(self):
        return self.model.encoder

    def get_decoder(self):
        return self.model.decoder

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

    def get_output_embeddings(self):
        return self.lm_head

    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings
    
    def set_label_smoothing(self, label_smoothing):
        self._label_smoothing = label_smoothing

    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        decoder_input_ids: Optional[torch.LongTensor] = None,
        decoder_attention_mask: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        decoder_head_mask: Optional[torch.Tensor] = None,
        cross_attn_head_mask: Optional[torch.Tensor] = None,
        encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        decoder_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[torch.Tensor], Seq2SeqLMOutput]:
        r"""
        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]`.

        Returns:
        """
        return_dict = (
            return_dict if return_dict is not None else self.config.use_return_dict
        )

        if labels is not None:
            if decoder_input_ids is None:
                decoder_input_ids = shift_tokens_right(
                    labels, self.config.pad_token_id, self.config.decoder_start_token_id
                )

        outputs = self.model(
            input_ids,
            attention_mask=attention_mask,
            decoder_input_ids=decoder_input_ids,
            encoder_outputs=encoder_outputs,
            decoder_attention_mask=decoder_attention_mask,
            head_mask=head_mask,
            decoder_head_mask=decoder_head_mask,
            cross_attn_head_mask=cross_attn_head_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            decoder_inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        lm_logits = self.lm_head(outputs[0])

        masked_lm_loss = None
        if labels is not None:
            # move labels to the correct device to enable PP
            labels = labels.to(lm_logits.device)
            masked_lm_loss = F.cross_entropy(
                input=lm_logits.view(-1, self.config.decoder_vocab_size),
                target=labels.view(-1),
                ignore_index=-100,
                label_smoothing=self._label_smoothing,
            )

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

        return Seq2SeqLMOutput(
            loss=masked_lm_loss,
            logits=lm_logits,
            past_key_values=outputs.past_key_values,
            decoder_hidden_states=outputs.decoder_hidden_states,
            decoder_attentions=outputs.decoder_attentions,
            cross_attentions=outputs.cross_attentions,
            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
            encoder_hidden_states=outputs.encoder_hidden_states,
            encoder_attentions=outputs.encoder_attentions,
        )

    def prepare_inputs_for_generation(
        self,
        decoder_input_ids,
        past_key_values=None,
        attention_mask=None,
        head_mask=None,
        decoder_head_mask=None,
        cross_attn_head_mask=None,
        use_cache=None,
        encoder_outputs=None,
        **kwargs,
    ):
        # cut decoder_input_ids if past is used
        if past_key_values is not None:
            decoder_input_ids = decoder_input_ids[:, -1:]

        return {
            "input_ids": None,  # encoder_outputs is defined. input_ids not needed
            "encoder_outputs": encoder_outputs,
            "past_key_values": past_key_values,
            "decoder_input_ids": decoder_input_ids,
            "attention_mask": attention_mask,
            "head_mask": head_mask,
            "decoder_head_mask": decoder_head_mask,
            "cross_attn_head_mask": cross_attn_head_mask,
            "use_cache": use_cache,  # change this to avoid caching (presumably for debugging)
        }

    @staticmethod
    def _reorder_cache(past_key_values, beam_idx):
        reordered_past = ()
        for layer_past in past_key_values:
            reordered_past += (
                tuple(
                    past_state.index_select(0, beam_idx) for past_state in layer_past
                ),
            )
        return reordered_past