modeling_chatglm.py

""" PyTorch ChatGLM model. """

import math
import sys
import torch
import torch.utils.checkpoint
import torch.nn.functional as F
from torch import nn
from torch.nn import CrossEntropyLoss, LayerNorm, MSELoss, BCEWithLogitsLoss
from torch.nn.utils import skip_init
from typing import Optional, Tuple, Union, List, Dict, Any

from transformers.modeling_outputs import (
    BaseModelOutputWithPast,
    CausalLMOutputWithPast,
    SequenceClassifierOutputWithPast,
)
from transformers.modeling_utils import PreTrainedModel
from transformers.utils import logging, is_torch_npu_available
from transformers.generation.logits_process import LogitsProcessor
from transformers.generation.utils import ModelOutput

from .configuration_chatglm import ChatGLMConfig

try:
    from transformers.utils import is_flash_attn_greater_or_equal_2_10, is_flash_attn_2_available

    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

from nltk.tokenize import PunktSentenceTokenizer
import re

# flags required to enable jit fusion kernels

if sys.platform != 'darwin' and not is_torch_npu_available():
    torch._C._jit_set_profiling_mode(False)
    torch._C._jit_set_profiling_executor(False)
    torch._C._jit_override_can_fuse_on_cpu(True)
    torch._C._jit_override_can_fuse_on_gpu(True)

logger = logging.get_logger(__name__)

_CHECKPOINT_FOR_DOC = "THUDM/ChatGLM"
_CONFIG_FOR_DOC = "ChatGLMConfig"


def default_init(cls, *args, **kwargs):
    return cls(*args, **kwargs)


class InvalidScoreLogitsProcessor(LogitsProcessor):
    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
        if torch.isnan(scores).any() or torch.isinf(scores).any():
            scores.zero_()
            scores[..., 198] = 5e4
        return scores


def split_tensor_along_last_dim(
        tensor: torch.Tensor,
        num_partitions: int,
        contiguous_split_chunks: bool = False,
) -> List[torch.Tensor]:
    """Split a tensor along its last dimension.
    Arguments:
        tensor: input tensor.
        num_partitions: number of partitions to split the tensor
        contiguous_split_chunks: If True, make each chunk contiguous
                                 in memory.
    Returns:
        A list of Tensors
    """
    # Get the size and dimension.
    last_dim = tensor.dim() - 1
    last_dim_size = tensor.size()[last_dim] // num_partitions
    # Split.
    tensor_list = torch.split(tensor, last_dim_size, dim=last_dim)
    # Note: torch.split does not create contiguous tensors by default.
    if contiguous_split_chunks:
        return tuple(chunk.contiguous() for chunk in tensor_list)

    return tensor_list


class RotaryEmbedding(nn.Module):
    def __init__(self, dim, rope_ratio=1, original_impl=False, device=None, dtype=None):
        super().__init__()
        inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2, device=device).to(dtype=dtype) / dim))
        self.register_buffer("inv_freq", inv_freq)
        self.dim = dim
        self.original_impl = original_impl
        self.rope_ratio = rope_ratio

    def forward_impl(
            self, seq_len: int, n_elem: int, dtype: torch.dtype, device: torch.device, base: int = 10000
    ):
        """Enhanced Transformer with Rotary Position Embedding.
        Derived from: https://github.com/labmlai/annotated_deep_learning_paper_implementations/blob/master/labml_nn/
        transformers/rope/__init__.py. MIT License:
        https://github.com/labmlai/annotated_deep_learning_paper_implementations/blob/master/license.
        """
        # $\Theta = {\theta_i = 10000^{\frac{2(i-1)}{d}}, i \in [1, 2, ..., \frac{d}{2}]}$
        base = base * self.rope_ratio
        theta = 1.0 / (base ** (torch.arange(0, n_elem, 2, dtype=torch.float, device=device) / n_elem))

        # Create position indexes `[0, 1, ..., seq_len - 1]`
        seq_idx = torch.arange(seq_len, dtype=torch.float, device=device)

        # Calculate the product of position index and $\theta_i$
        idx_theta = torch.outer(seq_idx, theta).float()

        cache = torch.stack([torch.cos(idx_theta), torch.sin(idx_theta)], dim=-1)

        # this is to mimic the behaviour of complex32, else we will get different results
        if dtype in (torch.float16, torch.bfloat16, torch.int8):
            cache = cache.bfloat16() if dtype == torch.bfloat16 else cache.half()
        return cache

    def forward(self, max_seq_len, offset=0):
        return self.forward_impl(
            max_seq_len, self.dim, dtype=self.inv_freq.dtype, device=self.inv_freq.device
        )


@torch.jit.script
def apply_rotary_pos_emb(x: torch.Tensor, rope_cache: torch.Tensor) -> torch.Tensor:
    # x: [b, np, sq, hn]
    b, np, sq, hn = x.size(0), x.size(1), x.size(2), x.size(3)
    rot_dim = rope_cache.shape[-2] * 2
    x, x_pass = x[..., :rot_dim], x[..., rot_dim:]
    # truncate to support variable sizes
    rope_cache = rope_cache[:, :sq]
    xshaped = x.reshape(b, np, sq, rot_dim // 2, 2)
    rope_cache = rope_cache.view(-1, 1, sq, xshaped.size(3), 2)
    x_out2 = torch.stack(
        [
            xshaped[..., 0] * rope_cache[..., 0] - xshaped[..., 1] * rope_cache[..., 1],
            xshaped[..., 1] * rope_cache[..., 0] + xshaped[..., 0] * rope_cache[..., 1],
        ],
        -1,
    )
    x_out2 = x_out2.flatten(3)
    return torch.cat((x_out2, x_pass), dim=-1)


class RMSNorm(torch.nn.Module):
    def __init__(self, normalized_shape, eps=1e-5, device=None, dtype=None, **kwargs):
        super().__init__()
        self.weight = torch.nn.Parameter(torch.empty(normalized_shape, device=device, dtype=dtype))
        self.eps = eps

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

        return (self.weight * hidden_states).to(input_dtype)


class CoreAttention(torch.nn.Module):
    def __init__(self, config: ChatGLMConfig, layer_number):
        super(CoreAttention, self).__init__()
        self.config = config
        self.apply_query_key_layer_scaling = config.apply_query_key_layer_scaling
        self.attention_softmax_in_fp32 = config.attention_softmax_in_fp32
        if self.apply_query_key_layer_scaling:
            self.attention_softmax_in_fp32 = True
        self.layer_number = max(1, layer_number)
        self.is_causal = True

        projection_size = config.kv_channels * config.num_attention_heads

        # Per attention head and per partition values.
        self.hidden_size_per_partition = projection_size
        self.hidden_size_per_attention_head = projection_size // config.num_attention_heads
        self.num_attention_heads_per_partition = config.num_attention_heads

        coeff = None
        self.norm_factor = math.sqrt(self.hidden_size_per_attention_head)
        if self.apply_query_key_layer_scaling:
            coeff = self.layer_number
            self.norm_factor *= coeff
        self.coeff = coeff

        self.attention_dropout = torch.nn.Dropout(config.attention_dropout)

    def forward(self, query_layer, key_layer, value_layer, attention_mask):
        # [b, np, sq, sk]
        output_size = (query_layer.size(0), query_layer.size(1), query_layer.size(2), key_layer.size(2))

        # [b, np, sq, hn] -> [b * np, sq, hn]
        query_layer = query_layer.view(output_size[0] * output_size[1], output_size[2], -1)
        # [b, np, sk, hn] -> [b * np, sk, hn]
        key_layer = key_layer.view(output_size[0] * output_size[1], output_size[3], -1)

        # preallocting input tensor: [b * np, sq, sk]
        matmul_input_buffer = torch.empty(
            output_size[0] * output_size[1], output_size[2], output_size[3], dtype=query_layer.dtype,
            device=query_layer.device
        )

        # Raw attention scores. [b * np, sq, sk]
        matmul_result = torch.baddbmm(
            matmul_input_buffer,
            query_layer,  # [b * np, sq, hn]
            key_layer.transpose(1, 2),  # [b * np, hn, sk]
            beta=0.0,
            alpha=(1.0 / self.norm_factor),
        )

        # change view to [b, np, sq, sk]
        attention_scores = matmul_result.view(*output_size)

        # ===========================
        # Attention probs and dropout
        # ===========================

        # attention scores and attention mask [b, np, sq, sk]
        if self.attention_softmax_in_fp32:
            attention_scores = attention_scores.float()
        if self.coeff is not None:
            attention_scores = attention_scores * self.coeff
        if attention_mask is None and attention_scores.shape[2] == attention_scores.shape[3]:
            attention_mask = torch.ones(output_size[0], 1, output_size[2], output_size[3],
                                        device=attention_scores.device, dtype=torch.bool)
            attention_mask.tril_()
            attention_mask = ~attention_mask
        if attention_mask is not None:
            attention_scores = attention_scores.masked_fill(attention_mask, float("-inf"))
        attention_probs = F.softmax(attention_scores, dim=-1)
        attention_probs = attention_probs.type_as(value_layer)

        # This is actually dropping out entire tokens to attend to, which might
        # seem a bit unusual, but is taken from the original Transformer paper.
        attention_probs = self.attention_dropout(attention_probs)

        # query layer shape: [b * np, sq, hn]
        # value layer shape: [b, np, sk, hn]
        # attention shape: [b, np, sq, sk]
        # context layer shape: [b, np, sq, hn]
        output_size = (value_layer.size(0), value_layer.size(1), query_layer.size(1), value_layer.size(3))
        # change view [b * np, sk, hn]
        value_layer = value_layer.view(output_size[0] * output_size[1], value_layer.size(2), -1)
        # change view [b * np, sq, sk]
        attention_probs = attention_probs.view(output_size[0] * output_size[1], output_size[2], -1)
        # matmul: [b * np, sq, hn]
        context_layer = torch.bmm(attention_probs, value_layer)
        # change view [b, np, sq, hn]
        context_layer = context_layer.view(*output_size)
        # [b, np, sq, hn] --> [b, sq, np, hn]
        context_layer = context_layer.transpose(1, 2).contiguous()
        # [b, sq, np, hn] --> [b, sq, hp]
        splited_context_layer_shape = context_layer.size()[:-2] + (self.hidden_size_per_partition,)
        context_layer = context_layer.reshape(*splited_context_layer_shape)

        return context_layer


class SdpaAttention(CoreAttention):
    def forward(self, query_layer, key_layer, value_layer, attention_mask):
        if attention_mask is None and query_layer.shape[2] == key_layer.shape[2]:
            context_layer = torch.nn.functional.scaled_dot_product_attention(query_layer, key_layer, value_layer,
                                                                             is_causal=True,
                                                                             dropout_p=self.config.attention_dropout if self.training else 0.0)
        else:
            if attention_mask is not None:
                attention_mask = ~attention_mask
            context_layer = torch.nn.functional.scaled_dot_product_attention(query_layer, key_layer, value_layer,
                                                                             attention_mask,
                                                                             dropout_p=self.config.attention_dropout if self.training else 0.0)
        context_layer = context_layer.transpose(1, 2).contiguous()
        splited_context_layer_shape = context_layer.size()[:-2] + (self.hidden_size_per_partition,)
        context_layer = context_layer.reshape(*splited_context_layer_shape)
        return context_layer


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.llama.modeling_llama.LlamaFlashAttention2
class FlashAttention2(CoreAttention):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()

    def forward(self, query_states, key_states, value_states, attention_mask):
        query_states = query_states.transpose(1, 2)
        key_states = key_states.transpose(1, 2)
        value_states = value_states.transpose(1, 2)
        batch_size, query_length = query_states.shape[:2]
        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
        dropout = self.config.attention_dropout if self.training else 0.0
        # Contains at least one padding token in the sequence
        if attention_mask is not None:
            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=None,
                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=None, causal=causal
            )
        attn_output = attn_output.reshape(batch_size, query_length, self.hidden_size_per_partition).contiguous()
        return attn_output

    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_attention_heads_per_partition, 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),
        )


CORE_ATTENTION_CLASSES = {
    "eager": CoreAttention,
    "sdpa": SdpaAttention,
    "flash_attention_2": FlashAttention2
}


class SelfAttention(torch.nn.Module):
    """Parallel self-attention layer abstract class.
    Self-attention layer takes input with size [s, b, h]
    and returns output of the same size.
    """

    def __init__(self, config: ChatGLMConfig, layer_number, device=None):
        super(SelfAttention, self).__init__()
        self.layer_number = max(1, layer_number)

        self.projection_size = config.kv_channels * config.num_attention_heads

        # Per attention head and per partition values.
        self.hidden_size_per_attention_head = self.projection_size // config.num_attention_heads
        self.num_attention_heads_per_partition = config.num_attention_heads

        self.multi_query_attention = config.multi_query_attention
        self.qkv_hidden_size = 3 * self.projection_size
        if self.multi_query_attention:
            self.num_multi_query_groups_per_partition = config.multi_query_group_num
            self.qkv_hidden_size = (
                    self.projection_size + 2 * self.hidden_size_per_attention_head * config.multi_query_group_num
            )
        self.query_key_value = nn.Linear(config.hidden_size, self.qkv_hidden_size,
                                         bias=config.add_bias_linear or config.add_qkv_bias,
                                         device=device, **_config_to_kwargs(config)
                                         )

        self.core_attention = CORE_ATTENTION_CLASSES[config._attn_implementation](config, self.layer_number)

        # Output.
        self.dense = nn.Linear(self.projection_size, config.hidden_size, bias=config.add_bias_linear,
                               device=device, **_config_to_kwargs(config)
                               )

    def _allocate_memory(self, inference_max_sequence_len, batch_size, device=None, dtype=None):
        if self.multi_query_attention:
            num_attention_heads = self.num_multi_query_groups_per_partition
        else:
            num_attention_heads = self.num_attention_heads_per_partition
        return torch.empty(
            inference_max_sequence_len,
            batch_size,
            num_attention_heads,
            self.hidden_size_per_attention_head,
            dtype=dtype,
            device=device,
        )

    def forward(
            self, hidden_states, attention_mask, rotary_pos_emb, kv_cache=None, use_cache=True
    ):
        # hidden_states: [b, sq, h]

        # =================================================
        # Pre-allocate memory for key-values for inference.
        # =================================================
        # =====================
        # Query, Key, and Value
        # =====================

        # Attention heads [b, sq, h] --> [b, sq, (np * 3 * hn)]
        mixed_x_layer = self.query_key_value(hidden_states)

        if self.multi_query_attention:
            (query_layer, key_layer, value_layer) = mixed_x_layer.split(
                [
                    self.num_attention_heads_per_partition * self.hidden_size_per_attention_head,
                    self.num_multi_query_groups_per_partition * self.hidden_size_per_attention_head,
                    self.num_multi_query_groups_per_partition * self.hidden_size_per_attention_head,
                ],
                dim=-1,
            )
            query_layer = query_layer.view(
                query_layer.size()[:-1] + (self.num_attention_heads_per_partition, self.hidden_size_per_attention_head)
            )
            key_layer = key_layer.view(
                key_layer.size()[:-1] + (self.num_multi_query_groups_per_partition, self.hidden_size_per_attention_head)
            )
            value_layer = value_layer.view(
                value_layer.size()[:-1]
                + (self.num_multi_query_groups_per_partition, self.hidden_size_per_attention_head)
            )
        else:
            new_tensor_shape = mixed_x_layer.size()[:-1] + \
                               (self.num_attention_heads_per_partition,
                                3 * self.hidden_size_per_attention_head)
            mixed_x_layer = mixed_x_layer.view(*new_tensor_shape)

            # [b, sq, np, 3 * hn] --> 3 [b, sq, np, hn]
            (query_layer, key_layer, value_layer) = split_tensor_along_last_dim(mixed_x_layer, 3)

        # [b, sq, np, hn] -> [b, np, sq, hn]
        query_layer, key_layer, value_layer = [k.transpose(1, 2) for k in [query_layer, key_layer, value_layer]]

        # apply relative positional encoding (rotary embedding)
        if rotary_pos_emb is not None:
            query_layer = apply_rotary_pos_emb(query_layer, rotary_pos_emb)
            key_layer = apply_rotary_pos_emb(key_layer, rotary_pos_emb)

        # adjust key and value for inference
        if kv_cache is not None:
            cache_k, cache_v = kv_cache
            key_layer = torch.cat((cache_k, key_layer), dim=2)
            value_layer = torch.cat((cache_v, value_layer), dim=2)
        if use_cache:
            if kv_cache is None:
                kv_cache = torch.cat((key_layer.unsqueeze(0).unsqueeze(0), value_layer.unsqueeze(0).unsqueeze(0)),
                                     dim=1)
            else:
                kv_cache = (key_layer, value_layer)
        else:
            kv_cache = None

        if self.multi_query_attention:
            key_layer = key_layer.unsqueeze(2)
            key_layer = key_layer.expand(
                -1, -1, self.num_attention_heads_per_partition // self.num_multi_query_groups_per_partition, -1, -1
            )
            key_layer = key_layer.contiguous().view(
                key_layer.size()[:1] + (self.num_attention_heads_per_partition,) + key_layer.size()[3:]
            )
            value_layer = value_layer.unsqueeze(2)
            value_layer = value_layer.expand(
                -1, -1, self.num_attention_heads_per_partition // self.num_multi_query_groups_per_partition, -1, -1
            )
            value_layer = value_layer.contiguous().view(
                value_layer.size()[:1] + (self.num_attention_heads_per_partition,) + value_layer.size()[3:]
            )

        # ==================================
        # core attention computation
        # ==================================

        context_layer = self.core_attention(query_layer, key_layer, value_layer, attention_mask)

        # =================
        # Output. [sq, b, h]
        # =================

        output = self.dense(context_layer)

        return output, kv_cache


def _config_to_kwargs(args):
    common_kwargs = {
        "dtype": args.torch_dtype,
    }
    return common_kwargs


class MLP(torch.nn.Module):
    """MLP.
    MLP will take the input with h hidden state, project it to 4*h
    hidden dimension, perform nonlinear transformation, and project the
    state back into h hidden dimension.
    """

    def __init__(self, config: ChatGLMConfig, device=None):
        super(MLP, self).__init__()

        self.add_bias = config.add_bias_linear

        # Project to 4h. If using swiglu double the output width, see https://arxiv.org/pdf/2002.05202.pdf
        self.dense_h_to_4h = nn.Linear(
            config.hidden_size,
            config.ffn_hidden_size * 2,
            bias=self.add_bias,
            device=device,
            **_config_to_kwargs(config)
        )

        def swiglu(x):
            x = torch.chunk(x, 2, dim=-1)
            return F.silu(x[0]) * x[1]

        self.activation_func = swiglu

        # Project back to h.
        self.dense_4h_to_h = nn.Linear(
            config.ffn_hidden_size,
            config.hidden_size,
            bias=self.add_bias,
            device=device,
            **_config_to_kwargs(config)
        )

    def forward(self, hidden_states):
        # [s, b, 4hp]
        intermediate_parallel = self.dense_h_to_4h(hidden_states)
        intermediate_parallel = self.activation_func(intermediate_parallel)
        # [s, b, h]
        output = self.dense_4h_to_h(intermediate_parallel)
        return output


class GLMBlock(torch.nn.Module):
    """A single transformer layer.
    Transformer layer takes input with size [s, b, h] and returns an
    output of the same size.
    """

    def __init__(self, config: ChatGLMConfig, layer_number, device=None):
        super(GLMBlock, self).__init__()
        self.layer_number = layer_number

        self.apply_residual_connection_post_layernorm = config.apply_residual_connection_post_layernorm

        self.fp32_residual_connection = config.fp32_residual_connection

        LayerNormFunc = RMSNorm if config.rmsnorm else LayerNorm
        # Layernorm on the input data.
        self.input_layernorm = LayerNormFunc(config.hidden_size, eps=config.layernorm_epsilon, device=device,
                                             dtype=config.torch_dtype)

        # Self attention.
        self.self_attention = SelfAttention(config, layer_number, device=device)
        self.hidden_dropout = config.hidden_dropout

        # Layernorm on the attention output
        self.post_attention_layernorm = LayerNormFunc(config.hidden_size, eps=config.layernorm_epsilon, device=device,
                                                      dtype=config.torch_dtype)

        # MLP
        self.mlp = MLP(config, device=device)

    def forward(
            self, hidden_states, attention_mask, rotary_pos_emb, kv_cache=None, use_cache=True,
    ):
        # hidden_states: [s, b, h]

        # Layer norm at the beginning of the transformer layer.
        layernorm_output = self.input_layernorm(hidden_states)
        # Self attention.
        attention_output, kv_cache = self.self_attention(
            layernorm_output,
            attention_mask,
            rotary_pos_emb,
            kv_cache=kv_cache,
            use_cache=use_cache
        )

        # Residual connection.
        if self.apply_residual_connection_post_layernorm:
            residual = layernorm_output
        else:
            residual = hidden_states

        layernorm_input = torch.nn.functional.dropout(attention_output, p=self.hidden_dropout, training=self.training)
        layernorm_input = residual + layernorm_input

        # Layer norm post the self attention.
        layernorm_output = self.post_attention_layernorm(layernorm_input)

        # MLP.
        mlp_output = self.mlp(layernorm_output)

        # Second residual connection.
        if self.apply_residual_connection_post_layernorm:
            residual = layernorm_output
        else:
            residual = layernorm_input

        output = torch.nn.functional.dropout(mlp_output, p=self.hidden_dropout, training=self.training)
        output = residual + output

        return output, kv_cache


class GLMTransformer(torch.nn.Module):
    """Transformer class."""

    def __init__(self, config: ChatGLMConfig, device=None):
        super(GLMTransformer, self).__init__()

        self.fp32_residual_connection = config.fp32_residual_connection
        self.post_layer_norm = config.post_layer_norm

        # Number of layers.
        self.num_layers = config.num_layers

        # Transformer layers.
        def build_layer(layer_number):
            return GLMBlock(config, layer_number, device=device)

        self.layers = torch.nn.ModuleList([build_layer(i + 1) for i in range(self.num_layers)])

        if self.post_layer_norm:
            LayerNormFunc = RMSNorm if config.rmsnorm else LayerNorm
            # Final layer norm before output.
            self.final_layernorm = LayerNormFunc(config.hidden_size, eps=config.layernorm_epsilon, device=device,
                                                 dtype=config.torch_dtype)

        self.gradient_checkpointing = False

    def _get_layer(self, layer_number):
        return self.layers[layer_number]

    def forward(
            self, hidden_states, attention_mask, rotary_pos_emb, kv_caches=None,
            use_cache: Optional[bool] = True,
            output_hidden_states: Optional[bool] = False,
    ):
        if not kv_caches:
            kv_caches = [None for _ in range(self.num_layers)]
        presents = () if use_cache else None
        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

        all_self_attentions = None
        all_hidden_states = () if output_hidden_states else None
        for index in range(self.num_layers):
            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

            layer = self._get_layer(index)
            if self.gradient_checkpointing and self.training:
                layer_ret = torch.utils.checkpoint.checkpoint(
                    layer,
                    hidden_states,
                    attention_mask,
                    rotary_pos_emb,
                    kv_caches[index],
                    use_cache,
                    use_reentrant=False
                )
            else:
                layer_ret = layer(
                    hidden_states,
                    attention_mask,
                    rotary_pos_emb,
                    kv_cache=kv_caches[index],
                    use_cache=use_cache
                )
            hidden_states, kv_cache = layer_ret
            if use_cache:
                # token by token decoding, use tuple format
                if kv_caches[0] is not None:
                    presents = presents + (kv_cache,)
                # prefilling in decoding, use tensor format to save cuda memory
                else:
                    if len(presents) == 0:
                        presents = kv_cache
                    else:
                        presents = torch.cat((presents, kv_cache.to(presents.device)), dim=0)

        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        # Final layer norm.
        if self.post_layer_norm:
            hidden_states = self.final_layernorm(hidden_states)

        return hidden_states, presents, all_hidden_states, all_self_attentions


class ChatGLMPreTrainedModel(PreTrainedModel):
    """
    An abstract class to handle weights initialization and
    a simple interface for downloading and loading pretrained models.
    """

    is_parallelizable = False
    supports_gradient_checkpointing = True
    config_class = ChatGLMConfig
    base_model_prefix = "transformer"
    _no_split_modules = ["GLMBlock"]
    _supports_flash_attn_2 = True
    _supports_sdpa = True

    def _init_weights(self, module: nn.Module):
        """Initialize the weights."""
        return

    def get_masks(self, input_ids, past_key_values, padding_mask=None):
        if self.config._attn_implementation == "flash_attention_2":
            if padding_mask is not None and not padding_mask.all():
                return padding_mask
            return None
        batch_size, seq_length = input_ids.shape
        full_attention_mask = torch.ones(batch_size, seq_length, seq_length, device=input_ids.device)
        full_attention_mask.tril_()
        past_length = 0
        if past_key_values:
            past_length = past_key_values[0][0].shape[2]
        if past_length:
            full_attention_mask = torch.cat((torch.ones(batch_size, seq_length, past_length,
                                                        device=input_ids.device), full_attention_mask), dim=-1)
        if padding_mask is not None:
            full_attention_mask = full_attention_mask * padding_mask.unsqueeze(1)
        if not past_length and padding_mask is not None:
            full_attention_mask -= padding_mask.unsqueeze(-1) - 1
        full_attention_mask = (full_attention_mask < 0.5).bool()
        full_attention_mask.unsqueeze_(1)
        return full_attention_mask

    def get_position_ids(self, input_ids, device):
        batch_size, seq_length = input_ids.shape
        position_ids = torch.arange(seq_length, dtype=torch.long, device=device).unsqueeze(0).repeat(batch_size, 1)
        return position_ids

class Embedding(torch.nn.Module):
    """Language model embeddings."""

    def __init__(self, config: ChatGLMConfig, device=None):
        super(Embedding, self).__init__()

        self.hidden_size = config.hidden_size
        # Word embeddings (parallel).
        self.word_embeddings = nn.Embedding(
            config.padded_vocab_size,
            self.hidden_size,
            dtype=config.torch_dtype,
            device=device
        )
        self.fp32_residual_connection = config.fp32_residual_connection

    def forward(self, input_ids):
        # Embeddings.
        words_embeddings = self.word_embeddings(input_ids)
        embeddings = words_embeddings
        # If the input flag for fp32 residual connection is set, convert for float.
        if self.fp32_residual_connection:
            embeddings = embeddings.float()
        return embeddings


class ChatGLMModel(ChatGLMPreTrainedModel):
    def __init__(self, config: ChatGLMConfig, device=None, empty_init=True):
        super().__init__(config)
        if empty_init:
            init_method = skip_init
        else:
            init_method = default_init
        init_kwargs = {}
        if device is not None:
            init_kwargs["device"] = device
        self.embedding = init_method(Embedding, config, **init_kwargs)
        self.num_layers = config.num_layers
        self.multi_query_group_num = config.multi_query_group_num
        self.kv_channels = config.kv_channels

        # Rotary positional embeddings
        self.seq_length = config.seq_length
        rotary_dim = (
            config.hidden_size // config.num_attention_heads if config.kv_channels is None else config.kv_channels
        )

        self.rotary_pos_emb = RotaryEmbedding(rotary_dim // 2, rope_ratio=config.rope_ratio,
                                              original_impl=config.original_rope,
                                              device=device, dtype=config.torch_dtype)
        self.encoder = init_method(GLMTransformer, config, **init_kwargs)
        self.output_layer = init_method(nn.Linear, config.hidden_size, config.padded_vocab_size, bias=False,
                                        dtype=config.torch_dtype, **init_kwargs)

    def get_input_embeddings(self):
        return self.embedding.word_embeddings

    def set_input_embeddings(self, value):
        self.embedding.word_embeddings = value

    def forward(
            self,
            input_ids,
            position_ids: Optional[torch.Tensor] = None,
            attention_mask: Optional[torch.BoolTensor] = None,
            full_attention_mask: Optional[torch.BoolTensor] = None,
            past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = 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,
    ):
        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

        batch_size, seq_length = input_ids.shape

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

        if full_attention_mask is None:
            if (attention_mask is not None and not attention_mask.all()) or (past_key_values and seq_length != 1):
                full_attention_mask = self.get_masks(input_ids, past_key_values, padding_mask=attention_mask)

        # Rotary positional embeddings
        rotary_pos_emb = self.rotary_pos_emb(self.seq_length)
        if position_ids is not None:
            rotary_pos_emb = rotary_pos_emb[position_ids]
        else:
            rotary_pos_emb = rotary_pos_emb[None, :seq_length]

        # Run encoder.
        hidden_states, presents, all_hidden_states, all_self_attentions = self.encoder(
            inputs_embeds, full_attention_mask, rotary_pos_emb=rotary_pos_emb,
            kv_caches=past_key_values, use_cache=use_cache, output_hidden_states=output_hidden_states
        )
        if presents is not None and type(presents) is torch.Tensor:
            presents = presents.split(1, dim=0)
            presents = list(presents)
            presents = [list(x.squeeze(0).split(1, dim=0)) for x in presents]
            presents = [tuple([x.squeeze(0) for x in y]) for y in presents]
            presents = tuple(presents)

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

        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=presents,
            hidden_states=all_hidden_states,
            attentions=all_self_attentions,
        )


class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
    def __init__(self, config: ChatGLMConfig, empty_init=True, device=None):
        super().__init__(config)

        self.max_sequence_length = config.max_length
        self.transformer = ChatGLMModel(config, empty_init=empty_init, device=device)
        self.config = config

    def _update_model_kwargs_for_generation(
            self,
            outputs: ModelOutput,
            model_kwargs: Dict[str, Any],
            is_encoder_decoder: bool = False,
    ) -> Dict[str, Any]:
        # update past_key_values
        cache_name, cache = self._extract_past_from_model_output(outputs)
        model_kwargs[cache_name] = cache

        # update attention mask
        if "attention_mask" in model_kwargs:
            attention_mask = model_kwargs["attention_mask"]
            model_kwargs["attention_mask"] = torch.cat(
                [attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1
            )

        # update position ids
        if "position_ids" in model_kwargs:
            position_ids = model_kwargs["position_ids"]
            new_position_id = position_ids[..., -1:].clone()
            new_position_id += 1
            model_kwargs["position_ids"] = torch.cat(
                [position_ids, new_position_id], dim=-1
            )

        model_kwargs["is_first_forward"] = False
        return model_kwargs

    def prepare_inputs_for_generation(
            self,
            input_ids: torch.LongTensor,
            past_key_values: Optional[torch.Tensor] = None,
            attention_mask: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.Tensor] = None,
            use_cache: Optional[bool] = None,
            is_first_forward: bool = True,
            **kwargs
    ) -> dict:
        # only last token for input_ids if past is not None
        if position_ids is None:
            position_ids = self.get_position_ids(input_ids, device=input_ids.device)
        if not is_first_forward:
            if past_key_values is not None:
                position_ids = position_ids[..., -1:]
                input_ids = input_ids[:, -1:]
        return {
            "input_ids": input_ids,
            "past_key_values": past_key_values,
            "position_ids": position_ids,
            "attention_mask": attention_mask,
            "return_last_logit": True,
            "use_cache": use_cache
        }

    def forward(
            self,
            input_ids: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.Tensor] = None,
            attention_mask: Optional[torch.Tensor] = None,
            past_key_values: Optional[Tuple[torch.FloatTensor]] = None,
            inputs_embeds: Optional[torch.Tensor] = None,
            labels: Optional[torch.Tensor] = None,
            use_cache: Optional[bool] = None,
            output_attentions: Optional[bool] = None,
            output_hidden_states: Optional[bool] = None,
            return_dict: Optional[bool] = None,
            return_last_logit: Optional[bool] = False,
    ):
        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

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

        hidden_states = transformer_outputs[0]
        if return_last_logit:
            hidden_states = hidden_states[:, -1:]
        lm_logits = self.transformer.output_layer(hidden_states)

        loss = None
        if labels is not None:
            lm_logits = lm_logits.to(torch.float32)

            # Shift so that tokens < n predict n
            shift_logits = lm_logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = CrossEntropyLoss(ignore_index=-100)
            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))

            lm_logits = lm_logits.to(hidden_states.dtype)
            loss = loss.to(hidden_states.dtype)

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

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

    @staticmethod
    def _reorder_cache(
            past: Tuple[Tuple[torch.Tensor, torch.Tensor], ...], beam_idx: torch.LongTensor
    ) -> Tuple[Tuple[torch.Tensor, torch.Tensor], ...]:
        """
        This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or
        [`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
        beam_idx at every generation step.
        Output shares the same memory storage as `past`.
        """
        return tuple(
            (
                layer_past[0].index_select(0, beam_idx.to(layer_past[0].device)),
                layer_past[1].index_select(0, beam_idx.to(layer_past[1].device)),
            )
            for layer_past in past
        )

    @torch.inference_mode()
    def chat(self, tokenizer, query: str, history: List[Dict] = None, role: str = "user",
             max_length: int = 131072, num_beams=1, do_sample=True, top_p=0.7, temperature=0.95,
             **kwargs):
        if history is None:
            history = []
        gen_kwargs = {"max_length": max_length, "num_beams": num_beams, "do_sample": do_sample, "top_p": top_p,
                      "temperature": temperature, **kwargs}
        inputs = tokenizer.build_chat_input(query, history=history, role=role)
        inputs = inputs.to(self.device)
        eos_token_id = [tokenizer.eos_token_id, tokenizer.get_command("<|user|>"),
                        tokenizer.get_command("<|observation|>")]
        outputs = self.generate(**inputs, **gen_kwargs, eos_token_id=eos_token_id)
        outputs = outputs.tolist()[0][len(inputs["input_ids"][0]):-1]
        response = tokenizer.decode(outputs).strip()
        history.append({"role": role, "content": query})
        return response, history

    def query_longcite(self, context, query, tokenizer, max_input_length=128000, max_new_tokens=1024, temperature=0.95):
        
        def text_split_by_punctuation(original_text, return_dict=False):
            # text = re.sub(r'([a-z])\.([A-Z])', r'\1. \2', original_text)  # separate period without space
            text = original_text
            custom_sent_tokenizer = PunktSentenceTokenizer(text)
            punctuations = r"([。；！？])"  # For Chinese support

            separated = custom_sent_tokenizer.tokenize(text)
            separated = sum([re.split(punctuations, s) for s in separated], [])
            # Put the punctuations back to the sentence
            for i in range(1, len(separated)):
                if re.match(punctuations, separated[i]):
                    separated[i-1] += separated[i]
                    separated[i] = ''

            separated = [s for s in separated if s != ""]
            if len(separated) == 1:
                separated = original_text.split('\n\n')
            separated = [s.strip() for s in separated if s.strip() != ""]
            if not return_dict:
                return separated
            else:
                pos = 0
                res = []
                for i, sent in enumerate(separated):
                    st = original_text.find(sent, pos)
                    assert st != -1, sent
                    ed = st + len(sent)
                    res.append(
                        {
                            'c_idx': i,
                            'content': sent,
                            'start_idx': st,
                            'end_idx': ed,
                        }
                    )
                    pos = ed
                return res 

        def get_prompt(context, question):
            sents = text_split_by_punctuation(context, return_dict=True)
            splited_context = ""
            for i, s in enumerate(sents):
                st, ed = s['start_idx'], s['end_idx']
                assert s['content'] == context[st:ed], s
                ed = sents[i+1]['start_idx'] if i < len(sents)-1 else len(context)
                sents[i] = {
                    'content': context[st:ed],
                    'start': st,
                    'end': ed,
                    'c_idx': s['c_idx'],
                }
                splited_context += f"<C{i}>"+context[st:ed]
            prompt = '''Please answer the user's question based on the following document. When a sentence S in your response uses information from some chunks in the document (i.e., <C{s1}>-<C_{e1}>, <C{s2}>-<C{e2}>, ...), please append these chunk numbers to S in the format "<statement>{S}<cite>[{s1}-{e1}][{s2}-{e2}]...</cite></statement>". You must answer in the same language as the user's question.\n\n[Document Start]\n%s\n[Document End]\n\n%s''' % (splited_context, question)
            return prompt, sents, splited_context
        
        def get_citations(statement, sents):
            c_texts = re.findall(r'<cite>(.*?)</cite>', statement, re.DOTALL)
            spans = sum([re.findall(r"\[([0-9]+\-[0-9]+)\]", c_text, re.DOTALL) for c_text in c_texts], [])
            statement = re.sub(r'<cite>(.*?)</cite>', '', statement, flags=re.DOTALL)
            merged_citations = []
            for i, s in enumerate(spans):
                try:
                    st, ed = [int(x) for x in s.split('-')]
                    if st > len(sents) - 1 or ed < st:
                        continue
                    st, ed = max(0, st), min(ed, len(sents)-1)
                    assert st <= ed, str(c_texts) + '\t' + str(len(sents))
                    if len(merged_citations) > 0 and st == merged_citations[-1]['end_sentence_idx'] + 1:
                        merged_citations[-1].update({
                            "end_sentence_idx": ed,
                            'end_char_idx': sents[ed]['end'],
                            'cite': ''.join([x['content'] for x in sents[merged_citations[-1]['start_sentence_idx']:ed+1]]),
                        })
                    else:
                        merged_citations.append({
                            "start_sentence_idx": st,
                            "end_sentence_idx": ed,
                            "start_char_idx":  sents[st]['start'],
                            'end_char_idx': sents[ed]['end'],
                            'cite': ''.join([x['content'] for x in sents[st:ed+1]]),
                        })
                except:
                    print(c_texts, len(sents), statement)
                    raise
            return statement, merged_citations[:3]
        
        def postprocess(answer, sents, splited_context):
            res = []
            pos = 0
            new_answer = ""
            while True:
                st = answer.find("<statement>", pos)
                if st == -1:
                    st = len(answer)
                ed = answer.find("</statement>", st)
                statement = answer[pos:st]
                if len(statement.strip()) > 5:
                    res.append({
                        "statement": statement,
                        "citation": []
                    })
                    new_answer += f"<statement>{statement}<cite></cite></statement>"
                else:
                    res.append({
                        "statement": statement,
                        "citation": None,
                    })
                    new_answer += statement
                
                if ed == -1:
                    break

                statement = answer[st+len("<statement>"):ed]
                if len(statement.strip()) > 0:
                    statement, citations = get_citations(statement, sents)
                    res.append({
                        "statement": statement,
                        "citation": citations
                    })
                    c_str = ''.join(['[{}-{}]'.format(c['start_sentence_idx'], c['end_sentence_idx']) for c in citations])
                    new_answer += f"<statement>{statement}<cite>{c_str}</cite></statement>"
                else:
                    res.append({
                        "statement": statement,
                        "citation": None,
                    })
                    new_answer += statement
                pos = ed + len("</statement>")
            return {
                "answer": new_answer.strip(),
                "statements_with_citations": [x for x in res if x['citation'] is not None],
                "splited_context": splited_context.strip(),
                "all_statements": res,
            }

        def truncate_from_middle(prompt, max_input_length=None, tokenizer=None):
            if max_input_length is None:
                return prompt
            else:
                assert tokenizer is not None
                tokenized_prompt = tokenizer.encode(prompt, add_special_tokens=False)
                if len(tokenized_prompt) > max_input_length:
                    half = int(max_input_length/2)
                    prompt = tokenizer.decode(tokenized_prompt[:half], skip_special_tokens=True)+tokenizer.decode(tokenized_prompt[-half:], skip_special_tokens=True)
                return prompt
        
        prompt, sents, splited_context = get_prompt(context, query)
        prompt = truncate_from_middle(prompt, max_input_length, tokenizer)
        output, _ = self.chat(tokenizer, prompt, history=[], max_new_tokens=max_new_tokens, temperature=temperature)
        result = postprocess(output, sents, splited_context)
        return result