# coding=utf-8
# Adapted from
# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py
# Copyright 2023 The PygmalionAI team.
# Copyright 2023 The vLLM team.
# Copyright 2023 DeepSeek-AI 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.
"""Inference-only Deepseek model."""
from typing import Any, Dict, List, Optional, Tuple

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

from aphrodite.modeling.metadata import InputMetadata
from aphrodite.modeling.layers.activation import SiluAndMul
from aphrodite.modeling.layers.attention import PagedAttention
from aphrodite.modeling.layers.triton_kernel.fused_moe import fused_moe
from aphrodite.modeling.layers.layernorm import RMSNorm
from aphrodite.modeling.layers.linear import (
    LinearMethodBase, MergedColumnParallelLinear, ReplicatedLinear,
    QKVParallelLinear, RowParallelLinear, ColumnParallelLinear)
from aphrodite.modeling.layers.rotary_embedding import get_rope
from aphrodite.modeling.layers.sampler import Sampler
from aphrodite.modeling.layers.vocab_parallel_embedding import (
    VocabParallelEmbedding, ParallelLMHead)
from aphrodite.modeling.megatron.communication_op import (
    tensor_model_parallel_all_reduce)
from aphrodite.modeling.megatron.parallel_state import (
    get_tensor_model_parallel_rank, get_tensor_model_parallel_world_size)
from aphrodite.modeling.sampling_metadata import SamplingMetadata
from aphrodite.modeling.hf_downloader import (default_weight_loader,
                                              hf_model_weights_iterator)
from aphrodite.common.sequence import SamplerOutput

KVCache = Tuple[torch.Tensor, torch.Tensor]


class DeepseekMLP(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
        linear_method: Optional[LinearMethodBase] = None,
        reduce_results: bool = True,
    ) -> None:
        super().__init__()
        if linear_method is not None and not linear_method.quant_config.merge_weight(
        ):
            self.merge_weight = False
            self.gate_proj = ColumnParallelLinear(hidden_size,
                                                  intermediate_size,
                                                  bias=False,
                                                  linear_method=linear_method)
            self.up_proj = ColumnParallelLinear(hidden_size,
                                                intermediate_size,
                                                bias=False,
                                                linear_method=linear_method)
        else:
            self.merge_weight = True
            self.gate_up_proj = MergedColumnParallelLinear(
                hidden_size, [intermediate_size] * 2,
                bias=False,
                linear_method=linear_method)
        self.down_proj = RowParallelLinear(intermediate_size,
                                           hidden_size,
                                           bias=False,
                                           linear_method=linear_method,
                                           reduce_results=reduce_results)
        if hidden_act != "silu":
            raise ValueError(f"Unsupported activation: {hidden_act}. "
                             "Only silu is supported for now.")
        self.act_fn = SiluAndMul()

    def forward(self, x):
        if self.merge_weight:
            gate_up, _ = self.gate_up_proj(x)
        else:
            up, _ = self.up_proj(x)
            gate, _ = self.gate_proj(x)
            gate_up = torch.cat([gate, up], dim=-1)
        x = self.act_fn(gate_up)
        x, _ = self.down_proj(x)
        return x


class DeepseekMoE(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        linear_method: Optional[LinearMethodBase] = None,
    ):
        super().__init__()
        self.config = config
        self.rank = get_tensor_model_parallel_rank()
        self.tp_size = get_tensor_model_parallel_world_size()
        self.n_routed_experts = config.n_routed_experts
        self.top_k = config.num_experts_per_tok
        if self.tp_size > self.n_routed_experts:
            raise ValueError(
                f"Tensor parallel size {self.tp_size} is greater than "
                f"the number of experts {self.n_routed_experts}.")

        self.experts = nn.ModuleList([
            DeepseekMLP(hidden_size=config.hidden_size,
                        intermediate_size=config.moe_intermediate_size,
                        hidden_act=config.hidden_act,
                        linear_method=linear_method,
                        reduce_results=False)
            for idx in range(self.n_routed_experts)
        ])
        self.pack_params()

        self.gate = ReplicatedLinear(config.hidden_size,
                                     self.n_routed_experts,
                                     bias=False,
                                     linear_method=None)

        if config.n_shared_experts is not None:
            intermediate_size = config.moe_intermediate_size * config.n_shared_experts
            self.shared_experts = DeepseekMLP(
                hidden_size=config.hidden_size,
                intermediate_size=intermediate_size,
                hidden_act=config.hidden_act,
                linear_method=linear_method,
                reduce_results=False,
            )

    def pack_params(self):
        w1 = []
        w2 = []
        for expert in self.experts:
            w1.append(expert.gate_up_proj.weight)
            w2.append(expert.down_proj.weight)
        self.w1 = torch._utils._flatten_dense_tensors(w1)
        w1s = torch._utils._unflatten_dense_tensors(self.w1, w1)
        for data, param in zip(w1s, w1):
            param.data = data
        self.w1 = self.w1.view(len(w1), *w1s[0].shape)

        self.w2 = torch._utils._flatten_dense_tensors(w2)
        w2s = torch._utils._unflatten_dense_tensors(self.w2, w2)
        for data, param in zip(w2s, w2):
            param.data = data

        self.w2 = self.w2.view(len(w2), *w2s[0].shape)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        batch_size, sequence_length, hidden_dim = hidden_states.shape
        hidden_states = hidden_states.view(-1, hidden_dim)
        if self.config.n_shared_experts is not None:
            shared_output = self.shared_experts(hidden_states)
        # router_logits: (batch * sequence_length, n_experts)
        router_logits, _ = self.gate(hidden_states)

        routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
        routing_weights, selected_experts = torch.topk(routing_weights,
                                                       self.top_k,
                                                       dim=-1)

        if self.config.norm_topk_prob:
            routing_weights /= routing_weights.sum(dim=-1, keepdim=True)

        final_hidden_states = fused_moe(hidden_states,
                                        self.w1,
                                        self.w2,
                                        routing_weights,
                                        selected_experts,
                                        inplace=True)

        if self.config.n_shared_experts is not None:
            final_hidden_states = final_hidden_states + shared_output
        final_hidden_states = tensor_model_parallel_all_reduce(
            final_hidden_states)

        return final_hidden_states.view(batch_size, sequence_length,
                                        hidden_dim)


class DeepseekAttention(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        rope_theta: float = 10000,
        rope_scaling: Optional[Dict[str, Any]] = None,
        max_position_embeddings: int = 8192,
        linear_method: Optional[LinearMethodBase] = None,
    ) -> None:
        super().__init__()
        self.hidden_size = hidden_size
        tp_size = get_tensor_model_parallel_world_size()
        self.total_num_heads = num_heads
        assert self.total_num_heads % tp_size == 0
        self.num_heads = self.total_num_heads // tp_size
        self.total_num_kv_heads = num_kv_heads
        if self.total_num_kv_heads >= tp_size:
            # Number of KV heads is greater than TP size, so we partition
            # the KV heads across multiple tensor parallel GPUs.
            assert self.total_num_kv_heads % tp_size == 0
        else:
            # Number of KV heads is less than TP size, so we replicate
            # the KV heads across multiple tensor parallel GPUs.
            assert tp_size % self.total_num_kv_heads == 0
        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
        self.head_dim = hidden_size // self.total_num_heads
        self.q_size = self.num_heads * self.head_dim
        self.kv_size = self.num_kv_heads * self.head_dim
        self.scaling = self.head_dim**-0.5
        self.rope_theta = rope_theta
        self.max_position_embeddings = max_position_embeddings

        if linear_method is not None and not linear_method.quant_config.merge_weight(
        ):
            self.merge_weight = False
            self.q_proj = ColumnParallelLinear(hidden_size,
                                               self.q_size,
                                               bias=False,
                                               linear_method=linear_method)
            self.k_proj = ColumnParallelLinear(hidden_size,
                                               self.kv_size,
                                               bias=False,
                                               linear_method=linear_method)
            self.v_proj = ColumnParallelLinear(hidden_size,
                                               self.kv_size,
                                               bias=False,
                                               linear_method=linear_method)
        else:
            self.merge_weight = True
            self.qkv_proj = QKVParallelLinear(
                hidden_size,
                self.head_dim,
                self.total_num_heads,
                self.total_num_kv_heads,
                bias=False,
                linear_method=linear_method,
            )

        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            hidden_size,
            bias=False,
            linear_method=linear_method,
        )

        is_neox_style = True if linear_method is None or linear_method.quant_config.rope_style(
        ) is None else linear_method.quant_config.rope_style()
        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=max_position_embeddings,
            base=rope_theta,
            rope_scaling=rope_scaling,
            is_neox_style=is_neox_style,
        )
        self.attn = PagedAttention(self.num_heads,
                                   self.head_dim,
                                   self.scaling,
                                   num_kv_heads=self.num_kv_heads)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        kv_cache: KVCache,
        input_metadata: InputMetadata,
    ) -> torch.Tensor:
        if self.merge_weight:
            qkv, _ = self.qkv_proj(hidden_states)
            q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size],
                                dim=-1)
        else:
            q, _ = self.q_proj(hidden_states)
            k, _ = self.k_proj(hidden_states)
            v, _ = self.v_proj(hidden_states)
        q, k = self.rotary_emb(positions, q, k)
        k_cache, v_cache = kv_cache
        attn_output = self.attn(q, k, v, k_cache, v_cache, input_metadata)
        output, _ = self.o_proj(attn_output)
        return output


class DeepseekDecoderLayer(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        layer_idx: int,
        linear_method: Optional[LinearMethodBase] = None,
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        rope_theta = getattr(config, "rope_theta", 10000)
        rope_scaling = getattr(config, "rope_scaling", None)
        max_position_embeddings = getattr(config, "max_position_embeddings",
                                          8192)
        self.self_attn = DeepseekAttention(
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            num_kv_heads=config.num_key_value_heads,
            rope_theta=rope_theta,
            rope_scaling=rope_scaling,
            max_position_embeddings=max_position_embeddings,
            linear_method=linear_method,
        )
        if (config.n_routed_experts is not None and  \
            layer_idx >= config.first_k_dense_replace and layer_idx % config.moe_layer_freq == 0):
            self.mlp = DeepseekMoE(config=config, linear_method=linear_method)
        else:
            self.mlp = DeepseekMLP(
                hidden_size=config.hidden_size,
                intermediate_size=config.intermediate_size,
                hidden_act=config.hidden_act,
                linear_method=linear_method,
            )
        self.input_layernorm = RMSNorm(config.hidden_size,
                                       eps=config.rms_norm_eps)
        self.post_attention_layernorm = RMSNorm(config.hidden_size,
                                                eps=config.rms_norm_eps)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        kv_cache: KVCache,
        input_metadata: InputMetadata,
        residual: Optional[torch.Tensor],
    ) -> torch.Tensor:
        # Self Attention
        if residual is None:
            residual = hidden_states
            hidden_states = self.input_layernorm(hidden_states)
        else:
            hidden_states, residual = self.input_layernorm(
                hidden_states, residual)
        hidden_states = self.self_attn(
            positions=positions,
            hidden_states=hidden_states,
            kv_cache=kv_cache,
            input_metadata=input_metadata,
        )

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


class DeepseekModel(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        linear_method: Optional[LinearMethodBase] = None,
    ) -> None:
        super().__init__()
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        self.embed_tokens = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
            linear_method=linear_method,
        )
        self.layers = nn.ModuleList([
            DeepseekDecoderLayer(config,
                                 layer_idx,
                                 linear_method=linear_method)
            for layer_idx in range(config.num_hidden_layers)
        ])
        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        kv_caches: List[KVCache],
        input_metadata: InputMetadata,
    ) -> torch.Tensor:
        hidden_states = self.embed_tokens(input_ids)
        residual = None
        for i in range(len(self.layers)):
            layer = self.layers[i]
            hidden_states, residual = layer(positions, hidden_states,
                                            kv_caches[i], input_metadata,
                                            residual)
        hidden_states, _ = self.norm(hidden_states, residual)
        return hidden_states


class DeepseekForCausalLM(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        linear_method: Optional[LinearMethodBase] = None,
    ) -> None:
        super().__init__()
        self.config = config
        self.linear_method = linear_method
        self.model = DeepseekModel(config, linear_method)
        self.lm_head = ParallelLMHead(config.vocab_size,
                                      config.hidden_size,
                                      linear_method=linear_method)
        self.sampler = Sampler(config.vocab_size)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        kv_caches: List[KVCache],
        input_metadata: InputMetadata,
    ) -> torch.Tensor:
        hidden_states = self.model(input_ids, positions, kv_caches,
                                   input_metadata)
        return hidden_states

    def sample(
        self,
        hidden_states: Optional[torch.Tensor],
        sampling_metadata: SamplingMetadata,
    ) -> Optional[SamplerOutput]:
        next_tokens = self.sampler(self.lm_head(hidden_states),
                                   sampling_metadata)
        return next_tokens

    def load_weights(self,
                     model_name_or_path: str,
                     cache_dir: Optional[str] = None,
                     load_format: str = "auto",
                     revision: Optional[str] = None):
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("qkv_proj", "q_proj", "q"),
            ("qkv_proj", "k_proj", "k"),
            ("qkv_proj", "v_proj", "v"),
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
        ]
        if self.linear_method is not None and not self.linear_method.quant_config.merge_weight(
        ):
            stacked_params_mapping = []
        params_dict = dict(self.named_parameters())
        for name, loaded_weight in hf_model_weights_iterator(
                model_name_or_path,
                cache_dir,
                load_format,
                revision,
                fall_back_to_pt=False):
            if "rotary_emb.inv_freq" in name:
                continue
            for (param_name, weight_name, shard_id) in stacked_params_mapping:
                if weight_name not in name:
                    continue
                name = name.replace(weight_name, param_name)
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
                # Skip experts that are not assigned to this worker.
                if (("mlp.experts." in name or "mlp.shared_experts." in name)
                        and name not in params_dict):
                    continue
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
                # Skip experts that are not assigned to this worker.
                if (("mlp.experts." in name or "mlp.shared_experts." in name)
                        and name not in params_dict):
                    continue
                param = params_dict[name]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                weight_loader(param, loaded_weight)