modeling_openpangu_dense.py

#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#                       This file was automatically generated from modular_openpangu_dense.py.
#               Do NOT edit this file manually as any edits will be overwritten by the generation of
#             the file from the modular. If any change should be done, please apply the change to the
#                          modular_openpangu_dense.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# coding=utf-8
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from typing import Callable, Optional, Union

import torch
import torch.nn.functional as F
import torch_npu
from torch_npu.contrib import transfer_to_npu

if "910" in torch.npu.get_device_name():
    NPU_ATTN_INFR = True
    print("[INFO] torch_npu detected. Using NPU fused infer attention.")
else:
    NPU_ATTN_INFR = False
from einops import rearrange
from torch import nn

from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache
from transformers.generation import GenerationMixin
from transformers.integrations import use_kernel_forward_from_hub
from transformers.masking_utils import create_causal_mask, create_sliding_window_causal_mask
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
from transformers.modeling_layers import GradientCheckpointingLayer
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
from transformers.modeling_rope_utils import dynamic_rope_update
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from transformers.processing_utils import Unpack
from transformers.utils import LossKwargs, auto_docstring, can_return_tuple, logging

from .configuration_openpangu_dense import PanguEmbeddedConfig


logger = logging.get_logger(__name__)


def aggregate_hidden_through_time(
    input_hidden, merge_conv, sliding_window=2, decay_coeff=0.5, restore_sliding_window=False, history_cache=None
):
    """
    input_hidden.shape = (B, S, H)
    return.shape = (B, S, H)
    """
    B, S, H = input_hidden.shape

    # concat zeors to the lefe of the first token
    if history_cache is None:
        history_cache = torch.zeros((B, H, sliding_window - 1), device=input_hidden.device, dtype=input_hidden.dtype)
    else:
        history_cache = history_cache.permute(0, 2, 1)

    conv_input = torch.cat(
        [history_cache, input_hidden.permute(0, 2, 1)],  # input_hidden (B, S, H) -> (B, H, S)
        dim=-1,
    )

    conv_output = merge_conv(conv_input)
    # (B, H, S) -> (B, S, H)
    return conv_output.permute(0, 2, 1)


class WindowBuffer:
    def __init__(self, win_size, decay_coeff, use_cache, aggregate_fn):
        self.win_size = win_size
        self.decay_coeff = decay_coeff
        self.use_cache = use_cache
        self.aggregate_fn = aggregate_fn
        self.buffer = None

    def get_aggregated_hidden(self, hidden_states):
        if not self.use_cache:
            self.buffer = None
            return aggregate_hidden_through_time(hidden_states, self.aggregate_fn, sliding_window=self.win_size)

        B, S, H = hidden_states.shape
        if S > 1:
            # prefill, generate first token
            win_input = aggregate_hidden_through_time(hidden_states, self.aggregate_fn, sliding_window=self.win_size)
            self.buffer = hidden_states[:, -(self.win_size - 1) :]
        else:
            # decode stage
            win_input = aggregate_hidden_through_time(
                hidden_states, self.aggregate_fn, sliding_window=self.win_size, history_cache=self.buffer
            )
            if self.win_size > 2:
                self.buffer = torch.cat([self.buffer[:, -(self.win_size - 2) :], hidden_states], dim=1)
            else:
                self.buffer = hidden_states

        return win_input


@use_kernel_forward_from_hub("RMSNorm")
class PanguEmbeddedRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        PanguEmbeddedRMSNorm is equivalent to T5LayerNorm
        """
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

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

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


class PanguEmbeddedRotaryEmbedding(nn.Module):
    def __init__(self, config: PanguEmbeddedConfig, device=None):
        super().__init__()

        base_dim = config.head_dim

        rotary_percent = config.rotary_percent

        dim = base_dim
        if rotary_percent < 1.0:
            dim = int(dim * rotary_percent)
            if dim % 2 != 0:
                dim += 1

        rotary_base = config.rope_theta
        inv_freq = 1.0 / (rotary_base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))

        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
        else:
            self.rope_type = "default"

        self.max_seq_len_cached = config.max_position_embeddings
        self.original_max_seq_len = config.max_position_embeddings

        self.config = config

        self.attention_scaling = 1.0

        if device is not None:
            inv_freq = inv_freq.to(device)

        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self.original_inv_freq = self.inv_freq

        self.dim = dim

    @torch.no_grad()
    @dynamic_rope_update
    def forward(self, x, position_ids):
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
        position_ids_expanded = position_ids[:, None, :].float()

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

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


class PanguEmbeddedMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
        self.act_fn = ACT2FN[config.hidden_act]

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


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


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


def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: Optional[torch.Tensor],
    scaling: float,
    dropout: float = 0.0,
    **kwargs,
):
    key_states = repeat_kv(key, module.num_key_value_groups)
    value_states = repeat_kv(value, module.num_key_value_groups)

    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
    if attention_mask is not None:
        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
        attn_weights = attn_weights + causal_mask

    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
    attn_output = torch.matmul(attn_weights, value_states)
    attn_output = attn_output.transpose(1, 2).contiguous()

    return attn_output, attn_weights


def apply_rotary_pos_emb(
    q: torch.Tensor, k: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor, unsqueeze_dim: int = 1
):
    """
    Applies Rotary Position Embedding to the query and key tensors,
    handling cases where rotary_percent < 1.0 by only rotating a subset of the dimensions.

    ATTENTION: This version assumes cos/sin tensors are already the full rotation dimension (D_rot),
    consistent with some Megatron/Fusion implementations, rather than the standard HF (D_rot/2) format.

    Args:
        q (`torch.Tensor`): The query tensor [Batch, Heads, Seq, Head_Dim].
        k (`torch.Tensor`): The key tensor [Batch, Heads, Seq, Head_Dim].
        cos (`torch.Tensor`): The cosine part of the rotary embedding [Batch, Seq, Head_Dim_Rotary]. <--- FULL D_ROT
        sin (`torch.Tensor`): The sine part of the rotary embedding [Batch, Seq, Head_Dim_Rotary]. <--- FULL D_ROT
        unsqueeze_dim (`int`, *optional*, defaults to 1): The dimension to unsqueeze cos/sin for broadcasting (usually the Heads dimension).

    Returns:
        `tuple(torch.Tensor)` comprising of the rotated query and key tensors.
    """
    rot_dim = cos.shape[-1]

    q_rope, q_pass = q[..., :rot_dim], q[..., rot_dim:]
    k_rope, k_pass = k[..., :rot_dim], k[..., rot_dim:]

    cos_broad = cos.unsqueeze(unsqueeze_dim)  # [B, 1, S, Dim]
    sin_broad = sin.unsqueeze(unsqueeze_dim)  # [B, 1, S, Dim]

    q_embed_rope = (q_rope * cos_broad) + (rotate_half(q_rope) * sin_broad)
    k_embed_rope = (k_rope * cos_broad) + (rotate_half(k_rope) * sin_broad)

    q_embed = torch.cat((q_embed_rope, q_pass), dim=-1)
    k_embed = torch.cat((k_embed_rope, k_pass), dim=-1)

    return q_embed, k_embed


class PanguEmbeddedAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(self, config: PanguEmbeddedConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.head_dim = config.head_dim
        self.num_key_value_groups = config.num_key_value_groups
        self.scaling = self.head_dim**-0.5
        self.attention_dropout = config.attention_dropout
        self.is_causal = True

        self.q_proj = nn.Linear(config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.bias)
        self.k_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.bias)
        self.v_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.bias)
        self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.bias)
        if layer_idx is None:
            logger.warning_once(
                f"Instantiating {self.__class__.__name__} without passing `layer_idx` is not recommended and will "
                "to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
                "when creating this class."
            )

        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.qk_nope_dim = config.qk_nope_dim
        self.qk_rope_dim = config.qk_rope_dim
        self.v_channels = config.v_channels
        self.num_key_value_heads = config.num_key_value_heads

        self.max_position_embeddings = config.max_position_embeddings
        self.rope_theta = config.rope_theta
        self.attn_groupnorm = config.attn_groupnorm
        self.attn_elementwise_gate = config.attn_elementwise_gate

        self.param_sink_number = config.param_sink_number
        self.param_sink_with_value = config.param_sink_with_value
        self.num_attention_heads = config.num_attention_heads

        self.rotary_emb = PanguEmbeddedRotaryEmbedding(config=config)
        if self.param_sink_number > 0:
            self.param_sink_query = torch.zeros(
                (self.param_sink_number, self.num_heads, self.head_dim), dtype=config.torch_dtype
            )

            self.param_sink_num_heads_per_partition = self.num_key_value_heads
            self.param_sink_key = torch.nn.Parameter(
                torch.empty(
                    (self.param_sink_number, self.param_sink_num_heads_per_partition, self.head_dim),
                    dtype=config.torch_dtype,
                )
            )
            if self.param_sink_with_value:
                self.param_sink_value = torch.nn.Parameter(
                    torch.empty(
                        (self.param_sink_number, self.param_sink_num_heads_per_partition, self.v_channels),
                        dtype=config.torch_dtype,
                    )
                )
            else:
                self.param_sink_value = torch.zeros(
                    (self.param_sink_number, self.param_sink_num_heads_per_partition, self.v_channels),
                    dtype=config.torch_dtype,
                )

        if self.attn_groupnorm:
            self.groupnorm = PanguEmbeddedRMSNorm(hidden_size=self.head_dim, eps=config.rms_norm_eps)

        if self.attn_elementwise_gate:
            self.attention_gate = nn.Linear(config.hidden_size, config.num_attention_heads * self.head_dim, bias=False)

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor],
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Cache] = None,
        output_attentions: bool = False,
        use_cache: bool = False,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[Cache]]:
        attention_interface: Callable = eager_attention_forward
        if self.config._attn_implementation != "eager":
            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.head_dim)

        bsz, q_len, _ = hidden_states.size()
        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)

        if self.attn_elementwise_gate:
            gate_score = self.attention_gate(hidden_states)
        else:
            gate_score = None

        kv_seq_len = q_len
        is_prefill = past_key_value.get_usable_length(kv_seq_len, self.layer_idx) == 0
        if past_key_value is not None:
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)

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

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

            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)

        kv_seq_len = key_states.shape[-2]

        if self.param_sink_number > 0:
            batch_size = query_states.shape[0]
            if is_prefill:
                param_sink_query = (
                    self.param_sink_query.permute(1, 0, 2)
                    .unsqueeze(0)
                    .expand(batch_size, -1, -1, -1)
                    .to(query_states.device)
                )
                query_states = torch.cat([param_sink_query, query_states], dim=2)
                q_len += self.param_sink_number

            param_sink_key = (
                self.param_sink_key.permute(1, 0, 2).unsqueeze(0).expand(batch_size, -1, -1, -1).to(key_states.device)
            )
            param_sink_value = (
                self.param_sink_value.permute(1, 0, 2)
                .unsqueeze(0)
                .expand(batch_size, -1, -1, -1)
                .to(value_states.device)
            )

            key_states = torch.cat([param_sink_key, key_states], dim=2)
            value_states = torch.cat([param_sink_value, value_states], dim=2)

            kv_seq_len += self.param_sink_number

        if not self.training and NPU_ATTN_INFR:
            q_len_current = query_states.shape[2]
            kv_len_current = key_states.shape[2]
            param_sink_number = self.config.param_sink_number

            # Causal Mask
            if is_prefill:
                causal_mask_npu = (
                    torch.triu(torch.ones([q_len_current, kv_len_current]), diagonal=1)
                    .bool()
                    .unsqueeze(0)
                    .unsqueeze(0)
                    .to(query_states.device)
                )
                original_mask = ~attention_mask.bool()
                expanded_mask = F.pad(
                    original_mask.float(), (param_sink_number, 0, param_sink_number, 0), mode="constant", value=1.0
                ).bool()
                attention_mask_npu = (expanded_mask) & (~causal_mask_npu)
            else:
                original_mask = ~attention_mask.bool()
                attention_mask_npu = F.pad(
                    original_mask.float(), (param_sink_number, 0, 0, 0), mode="constant", value=1.0
                ).bool()

            attention_mask_npu = ~attention_mask_npu.bool()

            attn_output, _ = torch_npu.npu_fused_infer_attention_score(
                query_states,
                key_states,
                value_states,
                num_heads=self.num_heads,
                num_key_value_heads=self.num_key_value_heads,
                input_layout="BNSD",
                atten_mask=attention_mask_npu,
                scale=self.scaling,
            )
            attn_output = attn_output.transpose(1, 2)  # (bsz, q_len, num_heads * head_dim)
            attn_weights = None
        else:
            attn_output, attn_weights = attention_interface(
                self,
                query_states,
                key_states,
                value_states,
                attention_mask,
                dropout=0.0 if not self.training else self.attention_dropout,
                scaling=self.scaling,
                sliding_window=self.sliding_window,
                position_ids=position_ids,
            )

        if self.param_sink_number > 0 and is_prefill:
            # (bsz, q_len_original, hidden_dim)
            attn_output = attn_output[:, self.param_sink_number :, :]

        if self.attn_groupnorm:
            attn_output = self.groupnorm(attn_output)
        if self.attn_elementwise_gate:
            core_attn_out_reshaped = rearrange(attn_output, "s b h d -> s b (h d)", h=self.num_attention_heads)
            core_attn_out_reshaped = core_attn_out_reshaped * F.sigmoid(gate_score)
            attn_output = rearrange(core_attn_out_reshaped, "s b (h d) -> s b h d", h=self.num_attention_heads)

        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value


class PanguEmbeddedDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: PanguEmbeddedConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = PanguEmbeddedAttention(config=config, layer_idx=layer_idx)
        self.mlp = PanguEmbeddedMLP(config)
        self.input_layernorm = PanguEmbeddedRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = PanguEmbeddedRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.attention_type = config.layer_types[layer_idx]
        if layer_idx == 0 or layer_idx == config.num_hidden_layers - 1:
            self.start_end = True
        else:
            self.start_end = False
        if self.start_end:
            self.router_sliding_window = config.router_sliding_window
            self.router_win_decay = config.router_win_decay
            self.merge_conv = torch.nn.Conv1d(
                config.hidden_size,
                config.hidden_size,
                self.router_sliding_window,
                groups=config.hidden_size,
                bias=False,
            )
            self.window_buffer = WindowBuffer(
                self.router_sliding_window, self.router_win_decay, True, self.merge_conv.forward
            )

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Cache] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = False,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]:
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)

        # Self Attention
        hidden_states, self_attn_weights, _ = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_value=past_key_value,
            output_attentions=output_attentions,
            use_cache=use_cache,
            cache_position=cache_position,
            **kwargs,
        )
        hidden_states = residual + hidden_states

        # Fully Connected
        residual = hidden_states

        if self.start_end and self.router_sliding_window:
            win_input = self.window_buffer.get_aggregated_hidden(hidden_states)
        else:
            win_input = hidden_states

        hidden_states = self.post_attention_layernorm(win_input)

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

        outputs = (hidden_states,)
        if output_attentions:
            outputs += (self_attn_weights,)

        return outputs


@auto_docstring
class PanguEmbeddedPreTrainedModel(PreTrainedModel):
    config_class = PanguEmbeddedConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["PanguEmbeddedDecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn_3 = True
    _supports_flash_attn_2 = True
    _supports_sdpa = True
    _supports_flex_attn = True
    _supports_cache_class = True
    _supports_quantized_cache = True
    _supports_static_cache = True
    _supports_attention_backend = True
    _keys_to_ignore_on_load_unexpected = [r"model\.layers\.27.*"]

    def _init_weights(self, module):
        std = self.config.initializer_range
        if isinstance(module, nn.Linear):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()
        elif isinstance(module, PanguEmbeddedRMSNorm):
            module.weight.data.fill_(1.0)


@auto_docstring
class PanguEmbeddedModel(PanguEmbeddedPreTrainedModel):
    def __init__(self, config: PanguEmbeddedConfig):
        super().__init__(config)
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
        self.layers = nn.ModuleList(
            [PanguEmbeddedDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.rotary_emb = PanguEmbeddedRotaryEmbedding(config=config)
        self.gradient_checkpointing = False
        self.norms = nn.ModuleList(
            [
                PanguEmbeddedRMSNorm(config.hidden_size, eps=config.rms_norm_eps),
                PanguEmbeddedRMSNorm(config.hidden_size, eps=config.rms_norm_eps),
            ]
        )

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

    def get_input_embeddings(self):
        return self.embed_tokens

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

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
    ) -> BaseModelOutputWithPast:
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache

        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

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

        if not isinstance(past_key_values, (type(None), Cache)):
            raise ValueError("The `past_key_values` should be either a `Cache` object or `None`.")
        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids)

        if use_cache and past_key_values is None:
            past_key_values = DynamicCache()

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

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

        if not isinstance(causal_mask_mapping := attention_mask, dict):
            mask_kwargs = {
                "config": self.config,
                "input_embeds": inputs_embeds,
                "attention_mask": attention_mask,
                "cache_position": cache_position,
                "past_key_values": past_key_values,
                "position_ids": position_ids,
            }
            causal_mask_mapping = {
                "full_attention": create_causal_mask(**mask_kwargs),
                "sliding_attention": create_sliding_window_causal_mask(**mask_kwargs),
            }

        hidden_states = inputs_embeds

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

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

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

            layer_outputs = decoder_layer(
                hidden_states,
                attention_mask=causal_mask_mapping[decoder_layer.attention_type],
                position_ids=position_ids,
                past_key_value=past_key_values,
                output_attentions=output_attentions,
                use_cache=use_cache,
                cache_position=cache_position,
                position_embeddings=position_embeddings,
                **flash_attn_kwargs,
            )

            hidden_states = layer_outputs[0]

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

        hidden_states = self.norms[0](hidden_states)

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

        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values if use_cache else None,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
        )


class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ...


@auto_docstring
class PanguEmbeddedForCausalLM(PanguEmbeddedPreTrainedModel, GenerationMixin):
    _tied_weights_keys = ["lm_head.weight"]
    _tp_plan = {"lm_head": "colwise_rep"}
    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}

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

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

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

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

    def get_output_embeddings(self):
        return self.lm_head

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

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

    def get_decoder(self):
        return self.model

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        logits_to_keep: Union[int, torch.Tensor] = 0,
        **kwargs: Unpack[KwargsForCausalLM],
    ) -> CausalLMOutputWithPast:
        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]`.

        Example:

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

        >>> model = PanguEmbeddedForCausalLM.from_pretrained("meta-PanguEmbedded/PanguEmbedded-2-7b-hf")
        >>> tokenizer = AutoTokenizer.from_pretrained("meta-PanguEmbedded/PanguEmbedded-2-7b-hf")

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

        >>> # Generate
        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
        ```"""
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )

        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs: BaseModelOutputWithPast = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            cache_position=cache_position,
            **kwargs,
        )

        hidden_states = outputs.last_hidden_state
        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
        logits = self.lm_head(hidden_states[:, slice_indices, :])

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

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


__all__ = ["PanguEmbeddedForCausalLM", "PanguEmbeddedModel", "PanguEmbeddedPreTrainedModel"]