deepseek_v2.py

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

# Adapted from
# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py
# 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 DeepseekV2/DeepseekV3 model."""
import os
import re
import vllm.envs as envs

import typing
from collections.abc import Callable, Iterable
from typing import Any, Optional, Union, Tuple

import torch
from torch import nn
from transformers import PretrainedConfig

from vllm.attention import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import (CacheConfig, ModelConfig, VllmConfig,
                         get_current_vllm_config)
from vllm.distributed import (get_ep_group, get_pp_group, get_dp_group,
                              get_tensor_model_parallel_world_size,
                              tensor_model_parallel_all_gather,
                              get_tensor_model_parallel_rank,
                              tensor_model_parallel_reduce_scatter)
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.fused_moe import FusedMoE, SharedFusedMoE

from vllm.model_executor.layers.fused_moe.utils import EPSharedExperts
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                               MergedColumnParallelLinear,
                                               ReplicatedLinear,
                                               FusedQuantedReplicatedLinear,
                                               RowParallelLinear)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
    ParallelLMHead, VocabParallelEmbedding)
from vllm.model_executor.layers.dp_attention import (dp_gather, dp_reduce_scatter_tensor, 
    get_moe_tp_size, get_moe_tp_rank, get_attention_tp_size)
from vllm.model_executor.model_loader.weight_utils import (
    default_weight_loader, maybe_remap_kv_scale_name)
from vllm.model_executor.sampling_metadata import SamplingMetadata
from vllm.sequence import IntermediateTensors

from .interfaces import MixtureOfExperts, SupportsPP
from .utils import (PPMissingLayer, is_pp_missing_parameter,
                    make_empty_intermediate_tensors_factory, make_layers,
                    maybe_prefix)
from vllm import _custom_ops as ops
from vllm.utils import W8a8GetCacheJSON
from lmslim.quantize.quant_ops import lm_faster_rmsquant

class DeepseekV2MLP(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
        quant_config: Optional[QuantizationConfig] = None,
        reduce_results: bool = True,
        prefix: str = "",
    ) -> None:
        super().__init__()
        
        vllm_config = get_current_vllm_config()
        enable_dp_attention = vllm_config.parallel_config.enable_dp_attention
        self.gate_up_proj = MergedColumnParallelLinear(
            hidden_size, [intermediate_size] * 2,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.gate_up_proj",
            enable_dp_attn_moe=enable_dp_attention)
        self.down_proj = RowParallelLinear(intermediate_size,
                                           hidden_size,
                                           bias=False,
                                           quant_config=quant_config,
                                           reduce_results=reduce_results,
                                           prefix=f"{prefix}.down_proj",
                                           enable_dp_attn_moe=enable_dp_attention)
        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,
                xqxs: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
                iqis: Optional[tuple[torch.Tensor, torch.Tensor]] = None
                ) -> torch.Tensor:
        if envs.USE_FUSED_RMS_QUANT:
            i_q, i_s = None, None
            if iqis is not None:
                i_q, i_s = iqis
            gate_up,  _   = self.gate_up_proj(x, iqis=iqis)
            if envs.USE_FUSED_SILU_MUL_QUANT:
                x, _ = self.down_proj(gate_up, use_fused_silu_mul_quant=True)
            else:
                x = self.act_fn(gate_up)
                x, _ = self.down_proj(x)
            return x
        elif envs.USE_FUSED_CUSTOM_ALL_REDUCE_RMS_QUANT and xqxs is not None:
            gate_up, _ = self.gate_up_proj(x, xqxs=xqxs)
            if envs.USE_FUSED_SILU_MUL_QUANT:
                x, _ = self.down_proj(gate_up, use_fused_silu_mul_quant=True)
            else:
                x = self.act_fn(gate_up)
                x, _ = self.down_proj(x)
            return x
        else:
            gate_up, _ = self.gate_up_proj(x)
            x = self.act_fn(gate_up)
            x, _ = self.down_proj(x)
            return x


class DeepseekV2MoE(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        enable_eplb: bool = False,
    ):
        super().__init__()
        self.tp_size = get_tensor_model_parallel_world_size()
        self.routed_scaling_factor = config.routed_scaling_factor

        self.ep_group = get_ep_group().device_group
        self.ep_rank = self.ep_group.rank()
        self.ep_size = self.ep_group.size()
        self.n_routed_experts: int = config.n_routed_experts
        self.n_shared_experts: int = config.n_shared_experts

        if config.hidden_act != "silu":
            raise ValueError(f"Unsupported activation: {config.hidden_act}. "
                             "Only silu is supported for now.")

        self.gate = ReplicatedLinear(config.hidden_size,
                                     config.n_routed_experts,
                                     bias=False,
                                     quant_config=None,
                                     prefix=f"{prefix}.gate")
        if config.topk_method == "noaux_tc":
            self.gate.e_score_correction_bias = nn.Parameter(
                torch.empty(config.n_routed_experts))
        else:
            self.gate.e_score_correction_bias = None

        # Load balancing settings.
        vllm_config = get_current_vllm_config()
        parallel_config = vllm_config.parallel_config
        self.enable_eplb = enable_eplb

        self.n_redundant_experts = parallel_config.num_redundant_experts
        self.n_logical_experts = self.n_routed_experts
        self.n_physical_experts = (self.n_logical_experts +
                                   self.n_redundant_experts)
        self.n_local_physical_experts = self.n_physical_experts // self.ep_size

        self.physical_expert_start = (self.ep_rank *
                                      self.n_local_physical_experts)
        self.physical_expert_end = (self.physical_expert_start +
                                    self.n_local_physical_experts)

        dp_size = get_dp_group().world_size
        self.enable_expert_parallel = parallel_config.enable_expert_parallel
        self.use_deepep = dp_size > 1 and parallel_config.enable_expert_parallel and \
            (envs.VLLM_ALL2ALL_BACKEND == "deepep_high_throughput" or \
             envs.VLLM_ALL2ALL_BACKEND == "deepep_low_latency" or \
             envs.VLLM_ALL2ALL_BACKEND == "deepep_auto")
        
        self.enable_shared_experts_overlap = False
        self.enable_shared_experts_fusion = (self.n_shared_experts != 0 and envs.VLLM_ENABLE_SHARED_EXPERTS_FUSION)

        if not self.use_deepep:
            if config.n_shared_experts is not None and not envs.VLLM_ENABLE_SHARED_EXPERTS_FUSION:
                intermediate_size = (config.moe_intermediate_size *
                                    config.n_shared_experts)
                self.shared_experts = DeepseekV2MLP(
                    hidden_size=config.hidden_size,
                    intermediate_size=intermediate_size,
                    hidden_act=config.hidden_act,
                    quant_config=quant_config,
                    reduce_results = False,
                    prefix=f"{prefix}.shared_experts",
                )
                self.enable_shared_experts_overlap = (not envs.VLLM_DISABLE_SHARED_EXPERTS_STREAM
                    and not envs.VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD
                    and config.n_shared_experts is not None)

            if self.enable_shared_experts_overlap:
                self.experts = SharedFusedMoE(
                    shared_experts=self.shared_experts,
                    num_experts=config.n_routed_experts,
                    top_k=config.num_experts_per_tok,
                    hidden_size=config.hidden_size,
                    intermediate_size=config.moe_intermediate_size,
                    reduce_results=False,
                    renormalize=config.norm_topk_prob,
                    quant_config=quant_config,
                    use_grouped_topk=True,
                    num_expert_group=config.n_group,
                    topk_group=config.topk_group,
                    prefix=f"{prefix}.experts",
                    scoring_func=config.scoring_func,
                    e_score_correction_bias=self.gate.e_score_correction_bias,
                    enable_eplb=self.enable_eplb,
                    num_redundant_experts=self.n_redundant_experts,
                    routed_scaling_factor=self.routed_scaling_factor)
            else:
                num_fake_experts = (
                        config.n_shared_experts 
                        if self.enable_shared_experts_fusion
                        else 0)
                self.experts = FusedMoE(
                    num_experts=config.n_routed_experts + (num_fake_experts),
                    top_k=config.num_experts_per_tok + min(num_fake_experts, 1),
                    hidden_size=config.hidden_size,
                    intermediate_size=config.moe_intermediate_size,
                    reduce_results=False,
                    renormalize=config.norm_topk_prob,
                    quant_config=quant_config,
                    use_grouped_topk=True,
                    num_expert_group=config.n_group,
                    topk_group=config.topk_group,
                    prefix=f"{prefix}.experts",
                    scoring_func=config.scoring_func,
                    e_score_correction_bias=self.gate.e_score_correction_bias,
                    enable_eplb=self.enable_eplb,
                    num_redundant_experts=self.n_redundant_experts,
                    routed_scaling_factor=self.routed_scaling_factor)
        else:
            if config.n_shared_experts is not None:
                intermediate_size = (config.moe_intermediate_size *
                                    config.n_shared_experts)
                self.shared_experts = EPSharedExperts(
                    hidden_size=config.hidden_size,
                    intermediate_size=intermediate_size,
                    hidden_act=config.hidden_act,
                    quant_config=quant_config,
                    reduce_results=False,
                    prefix=f"{prefix}.shared_experts",
                )
            self.experts = SharedFusedMoE(
                num_experts=config.n_routed_experts,
                top_k=config.num_experts_per_tok,
                hidden_size=config.hidden_size,
                intermediate_size=config.moe_intermediate_size,
                reduce_results=False,
                renormalize=config.norm_topk_prob,
                quant_config=quant_config,
                use_grouped_topk=True,
                num_expert_group=config.n_group,
                topk_group=config.topk_group,
                prefix=f"{prefix}.experts",
                scoring_func=config.scoring_func,
                e_score_correction_bias=self.gate.e_score_correction_bias,
                enable_eplb=self.enable_eplb,
                num_redundant_experts=self.n_redundant_experts,
                routed_scaling_factor=self.routed_scaling_factor,
                shared_experts=self.shared_experts)

        self.run_shared_expert_singlely = (
                self.n_shared_experts is not None
                and not self.enable_shared_experts_overlap
                and not envs.VLLM_ENABLE_SHARED_EXPERTS_FUSION
                )

        from vllm.two_batch_overlap.two_batch_overlap import tbo_all_reduce
        self.tbo_all_reduce = tbo_all_reduce

    def forward(self, hidden_states: torch.Tensor,
                residual: Optional[torch.Tensor] = None,
                xqxs: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
                iqis: Optional[tuple[torch.Tensor, torch.Tensor]] = None  # iq = input quant, is = input scale
                ) -> Union[torch.Tensor,
                           Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]]:
        num_tokens, hidden_dim = hidden_states.shape
        hidden_states = hidden_states.view(-1, hidden_dim)

        # For shared experts overlap optimization.
        def shared_exprts_overlap_pass(
                    hidden_states: torch.Tensor, router_logits: torch.Tensor, 
                    iqis: Optional[tuple[torch.Tensor, torch.Tensor]] = None
                ) -> tuple[torch.Tensor, torch.Tensor]:
            i_q, i_s = None, None
            if envs.USE_FUSED_RMS_QUANT:
                assert iqis is not None
                i_q, i_s = iqis[0], iqis[1]
            hidden_states_copy = hidden_states.clone()
            return self.experts(
                    hidden_states = hidden_states,
                    router_logits = router_logits,
                    hidden_states_copy = hidden_states_copy,
                    i_q = i_q,
                    i_s = i_s)
        
        # For shared experts fusion optimization.
        def shared_exprts_fusion_pass(
                    hidden_states: torch.Tensor, router_logits: torch.Tensor, 
                    iqis: Optional[tuple[torch.Tensor, torch.Tensor]] = None
                ) -> torch.Tensor:
            i_q, i_s = None, None
            if envs.USE_FUSED_RMS_QUANT:
                assert iqis is not None
                i_q, i_s = iqis[0], iqis[1]
            return self.experts(
                    hidden_states = hidden_states,
                    router_logits = router_logits,
                    i_q = i_q,
                    i_s = i_s)


        if envs.USE_FUSED_CUSTOM_ALL_REDUCE_RMS_QUANT and xqxs is not None:
            if self.n_shared_experts is not None and not self.enable_shared_experts_overlap:
                shared_output = self.shared_experts(hidden_states, xqxs=xqxs)

            router_logits, _ = self.gate(hidden_states)

            if envs.VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD:
                final_hidden_states = self.experts(
                    hidden_states=hidden_states,
                    router_logits=router_logits,
                    shared_output=shared_output)
            else:
                if self.enable_shared_experts_fusion:
                    final_hidden_states = shared_exprts_fusion_pass(hidden_states, router_logits)
                elif self.enable_shared_experts_overlap:
                    assert self.shared_experts is not None
                    shared_output, final_hidden_states = shared_exprts_overlap_pass(hidden_states, router_logits)
                    # Fix FP16 overflow
                    # See DeepseekV2DecoderLayer for more details.
                    if hidden_states.dtype != torch.float16:
                        final_hidden_states *= self.routed_scaling_factor
                        final_hidden_states += shared_output
                    else:
                        assert shared_output is not None
                        final_hidden_states += (shared_output * (1.0 / self.routed_scaling_factor))
                else:
                    if hidden_states.dtype != torch.float16:
                        final_hidden_states = self.experts(
                            hidden_states=hidden_states,
                            router_logits=router_logits) * self.routed_scaling_factor
                    else:
                        # Fix FP16 overflow
                        # See DeepseekV2DecoderLayer for more details.
                        final_hidden_states = self.experts(hidden_states=hidden_states,
                                                        router_logits=router_logits)

                    if shared_output is not None:
                        if hidden_states.dtype != torch.float16:
                            final_hidden_states = final_hidden_states + shared_output
                        else:
                            # Fix FP16 overflow
                            # See DeepseekV2DecoderLayer for more details.
                            final_hidden_states = final_hidden_states + shared_output \
                                * (1. / self.routed_scaling_factor)
        else: # RQ
            if not self.enable_expert_parallel:
                assert iqis is not None
                i_q, i_s = iqis
                if self.run_shared_expert_singlely:
                    if envs.USE_FUSED_RMS_QUANT:
                        shared_output = self.shared_experts(hidden_states, iqis=iqis)
                    else:
                        shared_output = self.shared_experts(hidden_states)

                router_logits, _ = self.gate(hidden_states)

                if self.enable_shared_experts_fusion:
                    final_hidden_states = shared_exprts_fusion_pass(hidden_states, router_logits, iqis = iqis)
                elif self.enable_shared_experts_overlap:
                    assert self.shared_experts is not None
                    shared_output, final_hidden_states = shared_exprts_overlap_pass(hidden_states, router_logits, iqis = iqis)
                    # Fix FP16 overflow
                    # See DeepseekV2DecoderLayer for more details.
                    if hidden_states.dtype != torch.float16:
                        final_hidden_states *= self.routed_scaling_factor
                        final_hidden_states += shared_output
                    else:
                        assert shared_output is not None
                        final_hidden_states += (shared_output * (1.0 / self.routed_scaling_factor))
                else:
                    if envs.VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD:
                        final_hidden_states = self.experts(
                            hidden_states=hidden_states,
                            router_logits=router_logits,
                            shared_output=shared_output,
                            i_q=i_q, i_s=i_s)
                    else:
                        if hidden_states.dtype != torch.float16:
                            final_hidden_states = self.experts(
                                hidden_states=hidden_states,
                                router_logits=router_logits,
                                i_q=i_q, i_s=i_s) * self.routed_scaling_factor
                        else:
                            # Fix FP16 overflow
                            # See DeepseekV2DecoderLayer for more details.
                            # fp16 mode not fused quant
                            if i_q is not None:
                                i_q=iqis[0]
                                i_s=iqis[1]
                            final_hidden_states = self.experts(hidden_states=hidden_states,
                                                            router_logits=router_logits,
                                                            i_q=i_q, i_s=i_s)
                    
                        if shared_output is not None:
                            if hidden_states.dtype != torch.float16:
                                final_hidden_states = final_hidden_states + shared_output
                            else:
                                # Fix FP16 overflow
                                # See DeepseekV2DecoderLayer for more details.
                                final_hidden_states = final_hidden_states + shared_output \
                                    * (1. / self.routed_scaling_factor)
            else:  # EP
                router_logits, _ = self.gate(hidden_states)
                if self.use_deepep:
                    shared_output, final_hidden_states = self.experts(hidden_states=hidden_states,
                                                router_logits=router_logits)

                    if shared_output is not None:
                        if hidden_states.dtype != torch.float16:
                            final_hidden_states = final_hidden_states + shared_output
                        else:
                            # Fix FP16 overflow
                            # See DeepseekV2DecoderLayer for more details.
                            final_hidden_states = final_hidden_states + shared_output \
                                * (1. / self.routed_scaling_factor)
                else:
                    if self.run_shared_expert_singlely:
                        if envs.USE_FUSED_RMS_QUANT:
                            shared_output = self.shared_experts(hidden_states, iqis=iqis)
                        else:
                            shared_output = self.shared_experts(hidden_states)

                    if self.enable_shared_experts_overlap:
                        assert self.shared_experts is not None
                        shared_output, final_hidden_states = shared_exprts_overlap_pass(hidden_states, router_logits)
                        # Fix FP16 overflow
                        # See DeepseekV2DecoderLayer for more details.
                        if hidden_states.dtype != torch.float16:
                            final_hidden_states += shared_output
                        else:
                            assert shared_output is not None
                            final_hidden_states += (shared_output * (1. / self.routed_scaling_factor))
                    else:
                        if i_q is not None:
                            i_q=iqis[0]
                            i_s=iqis[1]
                        final_hidden_states = self.experts(hidden_states=hidden_states,
                                                           router_logits=router_logits,
                                                           i_q=i_q, i_s=i_s)

                        if shared_output is not None:
                            if hidden_states.dtype != torch.float16:
                                final_hidden_states = final_hidden_states + shared_output
                            else:
                                # Fix FP16 overflow
                                # See DeepseekV2DecoderLayer for more details.
                                final_hidden_states = final_hidden_states + shared_output \
                                    * (1. / self.routed_scaling_factor)

        if self.tp_size > 1:
            if envs.VLLM_ENABLE_TBO:
                final_hidden_states = self.tbo_all_reduce(final_hidden_states)
            else:
                final_hidden_states = (
                    self.experts.maybe_all_reduce_tensor_model_parallel(
                        final_hidden_states))

            return final_hidden_states.view(num_tokens, hidden_dim)


def yarn_get_mscale(scale: float = 1, mscale: float = 1) -> float:
    import math
    if scale <= 1:
        return 1.0
    return 0.1 * mscale * math.log(scale) + 1.0


class DeepseekV2Attention(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        hidden_size: int,
        num_heads: int,
        qk_nope_head_dim: int,
        qk_rope_head_dim: int,
        v_head_dim: int,
        q_lora_rank: int,
        kv_lora_rank: int,
        rope_theta: float = 10000,
        rope_scaling: Optional[dict[str, Any]] = None,
        max_position_embeddings: int = 8192,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        reduce_results: bool = True,
    ) -> None:
        super().__init__()
        self.hidden_size = hidden_size
        self.qk_nope_head_dim = qk_nope_head_dim
        self.qk_rope_head_dim = qk_rope_head_dim
        self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim
        self.v_head_dim = v_head_dim
        self.q_lora_rank = q_lora_rank
        self.kv_lora_rank = kv_lora_rank
        self.num_heads = num_heads
        tp_size = get_tensor_model_parallel_world_size()
        assert num_heads % tp_size == 0
        self.num_local_heads = num_heads // tp_size
        self.scaling = self.qk_head_dim**-0.5
        self.rope_theta = rope_theta
        self.max_position_embeddings = max_position_embeddings

        if self.q_lora_rank is not None:
            self.q_a_proj = ReplicatedLinear(self.hidden_size,
                                             self.q_lora_rank,
                                             bias=False,
                                             quant_config=quant_config,
                                             prefix=f"{prefix}.q_a_proj")
            self.q_a_layernorm = RMSNorm(self.q_lora_rank,
                                         eps=config.rms_norm_eps)
            self.q_b_proj = ColumnParallelLinear(q_lora_rank,
                                                 self.num_heads *
                                                 self.qk_head_dim,
                                                 bias=False,
                                                 quant_config=quant_config,
                                                 prefix=f"{prefix}.q_b_proj")
        else:
            self.q_proj = ColumnParallelLinear(self.hidden_size,
                                               self.num_heads *
                                               self.qk_head_dim,
                                               bias=False,
                                               quant_config=quant_config,
                                               prefix=f"{prefix}.q_proj")

        self.kv_a_proj_with_mqa = ReplicatedLinear(
            self.hidden_size,
            self.kv_lora_rank + self.qk_rope_head_dim,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.kv_a_proj_with_mqa")
        self.kv_a_layernorm = RMSNorm(self.kv_lora_rank,
                                      eps=config.rms_norm_eps)
        self.kv_b_proj = ColumnParallelLinear(
            self.kv_lora_rank,
            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.kv_b_proj")
        # O projection.
        self.o_proj = RowParallelLinear(self.num_heads * self.v_head_dim,
                                        self.hidden_size,
                                        bias=False,
                                        quant_config=quant_config,
                                        prefix=f"{prefix}.o_proj",
                                        reduce_results=reduce_results)
        if rope_scaling:
            rope_scaling["rope_type"] = 'deepseek_yarn'

        self.rotary_emb = get_rope(qk_rope_head_dim,
                                   rotary_dim=qk_rope_head_dim,
                                   max_position=max_position_embeddings,
                                   base=rope_theta,
                                   rope_scaling=rope_scaling,
                                   is_neox_style=False)

        if rope_scaling:
            mscale_all_dim = rope_scaling.get("mscale_all_dim", False)
            scaling_factor = rope_scaling["factor"]
            mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
            self.scaling = self.scaling * mscale * mscale

        self.attn = Attention(self.num_local_heads,
                              self.qk_head_dim,
                              self.scaling,
                              num_kv_heads=self.num_local_heads,
                              cache_config=cache_config,
                              quant_config=quant_config,
                              prefix=f"{prefix}.attn")

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        if self.q_lora_rank is not None:
            q = self.q_a_proj(hidden_states)[0]
            q = self.q_a_layernorm(q)
            q = self.q_b_proj(q)[0].view(-1, self.num_local_heads,
                                         self.qk_head_dim)
        else:
            q = self.q_proj(hidden_states)[0].view(-1, self.num_local_heads,
                                                   self.qk_head_dim)
        q_nope, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim],
                               dim=-1)
        latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
        kv_a, _ = latent_cache.split(
            [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
        latent_cache = latent_cache.unsqueeze(1)
        kv_a = self.kv_a_layernorm(kv_a.contiguous())
        kv = self.kv_b_proj(kv_a)[0]
        kv = kv.view(-1, self.num_local_heads,
                     self.qk_nope_head_dim + self.v_head_dim)
        k_nope, v = kv.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)
        k_pe = latent_cache[:, :, self.kv_lora_rank:]

        q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)

        q[..., self.qk_nope_head_dim:] = q_pe
        k = torch.empty_like(q)
        k[..., :self.qk_nope_head_dim] = k_nope
        k[..., self.qk_nope_head_dim:] = k_pe
        # padding value to qk_head_dim for alignment
        v = torch.nn.functional.pad(
            v, [0, self.qk_head_dim - self.v_head_dim],
            value=0).view(-1, self.num_local_heads * self.qk_head_dim)
        attn_output = self.attn(q, k, v)
        attn_output = attn_output.view(
            -1, self.num_local_heads,
            self.qk_head_dim)[..., :self.v_head_dim].reshape(
                -1, self.num_local_heads * self.v_head_dim)
        output, _ = self.o_proj(attn_output)
        return output


class DeepseekV2MLAAttention(nn.Module):
    """
    Main reference: DeepseekV2 paper, and FlashInfer Implementation
    (https://arxiv.org/abs/2405.04434 and https://github.com/flashinfer-ai/flashinfer/pull/551).

    For more info see MLACommonImpl in: vllm/attention/backends/mla/utils.py
    """

    def __init__(
        self,
        config: PretrainedConfig,
        hidden_size: int,
        num_heads: int,
        qk_nope_head_dim: int,
        qk_rope_head_dim: int,
        v_head_dim: int,
        q_lora_rank: Optional[int],
        kv_lora_rank: int,
        rope_theta: float = 10000,
        rope_scaling: Optional[dict[str, Any]] = None,
        max_position_embeddings: int = 8192,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        reduce_results: bool = True,
    ) -> None:
        super().__init__()
        self.hidden_size = hidden_size
        self.qk_nope_head_dim = qk_nope_head_dim
        self.qk_rope_head_dim = qk_rope_head_dim
        self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim
        self.v_head_dim = v_head_dim

        self.q_lora_rank = q_lora_rank
        self.kv_lora_rank = kv_lora_rank

        self.num_heads = num_heads
        tp_size = get_tensor_model_parallel_world_size()
        assert num_heads % tp_size == 0
        self.num_local_heads = num_heads // tp_size

        self.scaling = self.qk_head_dim**-0.5
        self.rope_theta = rope_theta
        self.max_position_embeddings = max_position_embeddings

        if self.q_lora_rank is not None:
            if envs.USE_FUSED_RMS_QUANT:
                if envs.VLLM_USE_FUSED_QA_KVA_GEMM:
                    self.qa_kva_proj = FusedQuantedReplicatedLinear(self.hidden_size,
                                                             self.q_lora_rank,
                                                             self.kv_lora_rank,
                                                             self.qk_rope_head_dim,
                                                             bias=False,
                                                             quant_config=quant_config,
                                                             prefix=f"{prefix}.qa_kva_proj")
                else:
                    self.q_a_proj = ReplicatedLinear(self.hidden_size,
                                                self.q_lora_rank,
                                                bias=False,
                                                quant_config=quant_config,
                                                prefix=f"{prefix}.q_a_proj")
                self.q_b_proj = ColumnParallelLinear(q_lora_rank,
                                                 self.num_heads *
                                                 self.qk_head_dim,
                                                 bias=False,
                                                 quant_config=quant_config,
                                                 eps=config.rms_norm_eps,
                                                 prefix=f"{prefix}.q_b_proj")
            else:
                self.q_a_proj = ReplicatedLinear(self.hidden_size,
                                             self.q_lora_rank,
                                             bias=False,
                                             quant_config=quant_config,
                                             prefix=f"{prefix}.q_a_proj")
                self.q_b_proj = ColumnParallelLinear(q_lora_rank,
                                                 self.num_heads *
                                                 self.qk_head_dim,
                                                 bias=False,
                                                 quant_config=quant_config,
                                                 prefix=f"{prefix}.q_b_proj")

            self.q_a_layernorm = RMSNorm(self.q_lora_rank,
                                         eps=config.rms_norm_eps)

        else:
            self.q_proj = ColumnParallelLinear(self.hidden_size,
                                               self.num_heads *
                                               self.qk_head_dim,
                                               bias=False,
                                               quant_config=quant_config,
                                               prefix=f"{prefix}.q_proj")
        if not envs.VLLM_USE_FUSED_QA_KVA_GEMM:
            self.kv_a_proj_with_mqa = ReplicatedLinear(
                self.hidden_size,
                self.kv_lora_rank + self.qk_rope_head_dim,
                bias=False,
                quant_config=quant_config,
                prefix=f"{prefix}.kv_a_proj_with_mqa")
        self.kv_a_layernorm = RMSNorm(self.kv_lora_rank,
                                      eps=config.rms_norm_eps)
        self.kv_b_proj = ColumnParallelLinear(
            self.kv_lora_rank,
            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.kv_b_proj")
        self.o_proj = RowParallelLinear(self.num_heads * self.v_head_dim,
                                        self.hidden_size,
                                        bias=False,
                                        quant_config=quant_config,
                                        prefix=f"{prefix}.o_proj",
                                        reduce_results=reduce_results)

        if rope_scaling:
            rope_scaling["rope_type"] = 'deepseek_yarn'
        self.rotary_emb = get_rope(qk_rope_head_dim,
                                   rotary_dim=qk_rope_head_dim,
                                   max_position=max_position_embeddings,
                                   base=rope_theta,
                                   rope_scaling=rope_scaling,
                                   is_neox_style=False)
        if rope_scaling:
            mscale_all_dim = rope_scaling.get("mscale_all_dim", False)
            scaling_factor = rope_scaling["factor"]
            mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
            self.scaling = self.scaling * mscale * mscale

        # In the MLA backend, kv_cache includes both k_c and
        # pe (i.e. decoupled position embeddings). In particular,
        # the concat_and_cache_mla op requires
        #     k_c.size(1) + k_pe.size(1) == kv_cache.size(2)
        # i.e.
        #     kv_lora_rank + qk_rope_head_dim == head_size
        self.mla_attn = Attention(
            num_heads=self.num_local_heads,
            head_size=self.kv_lora_rank + self.qk_rope_head_dim,
            scale=self.scaling,
            num_kv_heads=1,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.attn",
            use_mla=True,
            # MLA Args
            q_lora_rank=self.q_lora_rank,
            kv_lora_rank=self.kv_lora_rank,
            qk_nope_head_dim=self.qk_nope_head_dim,
            qk_rope_head_dim=self.qk_rope_head_dim,
            qk_head_dim=self.qk_head_dim,
            v_head_dim=self.v_head_dim,
            kv_b_proj=self.kv_b_proj,
        )

        self.prefix = prefix
        self.debug_layer_idx = int(self.prefix.split(".")[-2])

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        pa_rms_weight: Optional[torch.Tensor] = None,
        pa_residual: Optional[torch.Tensor] = None,
        pa_rms_eps: Optional[float] = 1e-6,
        pa_quant_dtype: Optional[torch.dtype] = torch.int8,
        update_input: Optional[bool] = True,
        iqis: Optional[tuple[torch.Tensor, torch.Tensor]] = None
            
    ) -> Union[torch.Tensor,
               Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]]:
        if envs.USE_FUSED_RMS_QUANT and iqis is not None:
            if envs.VLLM_USE_FUSED_QA_KVA_GEMM:
                if self.q_lora_rank is not None:
                    qc_kvc_kpe, _bias = self.qa_kva_proj(hidden_states,  iqis)
                    q_c = qc_kvc_kpe[:, :self.q_lora_rank]
                    kvc_kpe = qc_kvc_kpe[:, self.q_lora_rank:]
                    q, _, _ = self.q_b_proj(q_c, rms_weight=self.q_a_layernorm.weight.data, residual=None, update_hd=False)
                    
                else:
                    q = self.q_proj(hidden_states)[0]
                kv_c, k_pe = kvc_kpe.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
            else:
                if self.q_lora_rank is not None:
                    q_c, _ = self.q_a_proj(hidden_states, iqis=iqis)
                    q, _, _ = self.q_b_proj(q_c, rms_weight=self.q_a_layernorm.weight.data, residual=None, update_hd=False)
                    
                else:
                    q = self.q_proj(hidden_states)[0]
                kvc_kpe = self.kv_a_proj_with_mqa(hidden_states, iqis=iqis)[0]
                kv_c, k_pe = kvc_kpe.split(
                                [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)

            if not envs.VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT:
                if envs.VLLM_USE_LIGHTOP:
                    kv_c_normed = self.kv_a_layernorm.forward_cuda_opt(kv_c)
                else:
                    kv_c_normed = self.kv_a_layernorm(kv_c.contiguous())

                q = q.view(-1, self.num_local_heads, self.qk_head_dim)
                # Add head dim of 1 to k_pe
                k_pe = k_pe.unsqueeze(1)

                q[..., self.qk_nope_head_dim:], k_pe = self.rotary_emb(
                    positions, q[..., self.qk_nope_head_dim:], k_pe)

                attn_out = self.mla_attn(
                    q,
                    kv_c_normed,
                    k_pe,
                    output_shape=(hidden_states.shape[0],
                                self.num_local_heads * self.v_head_dim))
            else:
                q = q.view(-1, self.num_local_heads, self.qk_head_dim)
                # Add head dim of 1 to k_pe
                k_pe = k_pe.unsqueeze(1)
                weight = self.kv_a_layernorm.weight
                cos_sin_cache = self.rotary_emb.cos_sin_cache
                if cos_sin_cache.device != positions.device or cos_sin_cache.device != q.dtype:
                    cos_sin_cache = cos_sin_cache.to(positions.device, dtype=q.dtype)
                kv_c_normed = torch.empty(kv_c.shape, dtype=kv_c.dtype, device=kv_c.device)
                attn_out = self.mla_attn(
                    q[..., self.qk_nope_head_dim:],
                    kv_c,
                    k_pe,
                    output_shape=(hidden_states.shape[0],
                                self.num_local_heads * self.v_head_dim),
                    q_ori=q,
                    key_normed=kv_c_normed,
                    positions=positions,
                    weight=weight,
                    cos_sin_cache=cos_sin_cache)
            return self.o_proj(attn_out)[0]
        elif envs.USE_FUSED_CUSTOM_ALL_REDUCE_RMS_QUANT and pa_rms_weight is not None and pa_residual is not None:
            if self.q_lora_rank is not None:
                q_c = self.q_a_proj(hidden_states)[0]
                q_c = self.q_a_layernorm(q_c)
                q = self.q_b_proj(q_c)[0]
            else:
                q = self.q_proj(hidden_states)[0]
            kv_c, k_pe = self.kv_a_proj_with_mqa(hidden_states)[0].split(
                [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
            if not envs.VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT:
                if envs.VLLM_USE_LIGHTOP:
                    kv_c_normed = self.kv_a_layernorm.forward_cuda_opt(kv_c)
                else:
                    kv_c_normed = self.kv_a_layernorm(kv_c.contiguous())

                q = q.view(-1, self.num_local_heads, self.qk_head_dim)
                k_pe = k_pe.unsqueeze(1)

                q[..., self.qk_nope_head_dim:], k_pe = self.rotary_emb(
                    positions, q[..., self.qk_nope_head_dim:], k_pe)

                attn_out = self.mla_attn(
                    q,
                    kv_c_normed,
                    k_pe,
                    output_shape=(hidden_states.shape[0],
                                self.num_local_heads * self.v_head_dim))
            else:
                q = q.view(-1, self.num_local_heads, self.qk_head_dim)
                # Add head dim of 1 to k_pe
                k_pe = k_pe.unsqueeze(1)
                weight = self.kv_a_layernorm.weight
                cos_sin_cache = self.rotary_emb.cos_sin_cache
                if cos_sin_cache.device != positions.device or cos_sin_cache.device != q.dtype:
                    cos_sin_cache = cos_sin_cache.to(positions.device, dtype=q.dtype)
                kv_c_normed = torch.empty(kv_c.shape, dtype=kv_c.dtype, device=kv_c.device)
                attn_out = self.mla_attn(
                    q[..., self.qk_nope_head_dim:],
                    kv_c,
                    k_pe,
                    output_shape=(hidden_states.shape[0],
                                self.num_local_heads * self.v_head_dim),
                    q_ori=q,
                    key_normed=kv_c_normed,
                    positions=positions,
                    weight=weight,
                    cos_sin_cache=cos_sin_cache)
            packages_ = self.o_proj(attn_out,
                                   pa_rms_weight=pa_rms_weight,
                                   pa_residual=pa_residual,
                                   pa_rms_eps=pa_rms_eps,
                                   pa_quant_dtype=pa_quant_dtype,
                                   update_input=update_input)[:4]
            assert len(packages_) == 4
            hs, resi, xq, xs = packages_
            assert xq is not None and xs is not None
            return hs, resi, xq, xs

        else:
            if self.q_lora_rank is not None:
                q_c = self.q_a_proj(hidden_states)[0]
                q_c = self.q_a_layernorm(q_c)
                q = self.q_b_proj(q_c)[0]
            else:
                q = self.q_proj(hidden_states)[0]
            kv_c, k_pe = self.kv_a_proj_with_mqa(hidden_states)[0].split(
                [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
            if not envs.VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT:
                if envs.VLLM_USE_LIGHTOP:
                    kv_c_normed = self.kv_a_layernorm.forward_cuda_opt(kv_c)
                else:
                    kv_c_normed = self.kv_a_layernorm(kv_c.contiguous())

                q = q.view(-1, self.num_local_heads, self.qk_head_dim)
                # Add head dim of 1 to k_pe
                k_pe = k_pe.unsqueeze(1)

                q[..., self.qk_nope_head_dim:], k_pe = self.rotary_emb(
                    positions, q[..., self.qk_nope_head_dim:], k_pe)

                attn_out = self.mla_attn(
                    q,
                    kv_c_normed,
                    k_pe,
                    output_shape=(hidden_states.shape[0],
                                self.num_local_heads * self.v_head_dim))
            else:
                q = q.view(-1, self.num_local_heads, self.qk_head_dim)
                # Add head dim of 1 to k_pe
                k_pe = k_pe.unsqueeze(1)
                weight = self.kv_a_layernorm.weight
                cos_sin_cache = self.rotary_emb.cos_sin_cache
                if cos_sin_cache.device != positions.device or cos_sin_cache.device != q.dtype:
                    cos_sin_cache = cos_sin_cache.to(positions.device, dtype=q.dtype)
                kv_c_normed = torch.empty(kv_c.shape, dtype=kv_c.dtype, device=kv_c.device)
                attn_out = self.mla_attn(
                    q[..., self.qk_nope_head_dim:],
                    kv_c,
                    k_pe,
                    output_shape=(hidden_states.shape[0],
                                self.num_local_heads * self.v_head_dim),
                    q_ori=q,
                    key_normed=kv_c_normed,
                    positions=positions,
                    weight=weight,
                    cos_sin_cache=cos_sin_cache)
            return self.o_proj(attn_out)[0]


class DeepseekV2DecoderLayer(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        prefix: str,
        model_config: ModelConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        enable_eplb: bool = False,
    ) -> 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)
        # DecoderLayers are created with `make_layers` which passes the prefix
        # with the layer's index.
        layer_idx = int(prefix.split(sep='.')[-1])
        self.layer_idx = layer_idx

        self.dp_size = get_dp_group().world_size
        vllm_config = get_current_vllm_config()
        parallel_config = vllm_config.parallel_config
        self.use_deepep = self.dp_size > 1 and parallel_config.enable_expert_parallel and \
            (envs.VLLM_ALL2ALL_BACKEND == "deepep_high_throughput" or \
             envs.VLLM_ALL2ALL_BACKEND == "deepep_low_latency" or \
             envs.VLLM_ALL2ALL_BACKEND == "deepep_auto")
        self.tp_size = get_tensor_model_parallel_world_size()
        self.config = config
        self.tp_rank = get_tensor_model_parallel_rank()
        self.enable_dp_attention = vllm_config.parallel_config.enable_dp_attention
        self.dp_rank = vllm_config.parallel_config.data_parallel_rank

        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 = DeepseekV2MoE(
                config=config,
                quant_config=quant_config,
                prefix=f"{prefix}.mlp",
                enable_eplb=enable_eplb,
            )
        else:
            self.mlp = DeepseekV2MLP(
                hidden_size=config.hidden_size,
                intermediate_size=config.intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
                prefix=f"{prefix}.mlp",
            )

        self.enable_ep_sp = isinstance(self.mlp,
                        DeepseekV2MoE) and self.use_deepep and self.tp_size > 1
        self.is_mtp_layer = False
        if self.layer_idx == config.num_hidden_layers:
            self.is_mtp_layer = True
        reduce_results = True
        if isinstance(self.mlp,
                        DeepseekV2MoE) and self.use_deepep and \
                            self.tp_size > 1 and not self.is_mtp_layer:
            reduce_results = False
        else:
            if self.enable_dp_attention:
                reduce_results = False

        if model_config.use_mla:
            attn_cls = DeepseekV2MLAAttention
        else:
            attn_cls = DeepseekV2Attention
        self.self_attn = attn_cls(
            config=config,
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            qk_nope_head_dim=config.qk_nope_head_dim,
            qk_rope_head_dim=config.qk_rope_head_dim,
            v_head_dim=config.v_head_dim,
            q_lora_rank=config.q_lora_rank
            if hasattr(config, "q_lora_rank") else None,
            kv_lora_rank=config.kv_lora_rank,
            rope_theta=rope_theta,
            rope_scaling=rope_scaling,
            max_position_embeddings=max_position_embeddings,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.self_attn",
            reduce_results=reduce_results
        )

        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)
        self.routed_scaling_factor = config.routed_scaling_factor
        self.use_fused_rms_quant = envs.USE_FUSED_RMS_QUANT
        self.use_fused_custom_all_reduce = envs.USE_FUSED_CUSTOM_ALL_REDUCE_RMS_QUANT
        self._eps = config.rms_norm_eps



    def forward_fused_RQ(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor]
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        # Fix residual FP16 overflow
        residual_fix_overflow = False

        assert self.input_layernorm.has_weight is True
        if residual is None:
            residual = hidden_states
            residual_fix_overflow = True
            i_q, i_s = lm_faster_rmsquant(input=hidden_states,
                                         rms_weight = self.input_layernorm.weight.data,
                                         epsilon=self._eps,
                                         quant_dtype=torch.int8,
                                         residual=None,
                                         update_input=False)
        else:
            i_q, i_s = lm_faster_rmsquant(input=hidden_states,
                                         rms_weight = self.input_layernorm.weight.data,
                                         epsilon=self._eps,
                                         quant_dtype=torch.int8,
                                         residual=residual,
                                         update_input=False)
            
        hidden_states = self.self_attn(positions=positions,
                                       hidden_states = hidden_states, # get attr
                                       iqis=(i_q, i_s))
            
        if hidden_states.dtype == torch.float16:
            # rmsnorm, and rmsnorm result would not affect by scale.
            hidden_states *= 1. / self.routed_scaling_factor
            if self.layer_idx == 0 or residual_fix_overflow:
                # The residual is shared by all layers, we only scale it on
                # first layer.
                residual *= 1. / self.routed_scaling_factor
        
        update_hs = True if isinstance(self.mlp, DeepseekV2MoE) else False
        _i_q, _i_s = lm_faster_rmsquant(input=hidden_states,
                                        rms_weight=self.post_attention_layernorm.weight.data,
                                        epsilon=self._eps,
                                        quant_dtype=torch.int8,
                                        residual=residual,
                                        update_input=update_hs)
        new_resi = residual
        hidden_states = self.mlp(hidden_states, 
                                 iqis=(_i_q, _i_s))

        if isinstance(self.mlp,
                    DeepseekV2MLP) and hidden_states.dtype == torch.float16:
            # Fix FP16 overflow
            # Scaling the DeepseekV2MLP output, it is the input of
            # input_layernorm of next decoder layer.
            # The scaling of DeepseekV2MOE output would be done in the forward
            # of DeepseekV2MOE
            hidden_states *= 1. / self.routed_scaling_factor
        return hidden_states, new_resi

    def forward_fused_CRQ(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor]
        ) -> Tuple[torch.Tensor, torch.Tensor]:
        residual_fix_overflow = False
        if residual is None:
            residual = hidden_states
            hidden_states = self.input_layernorm(hidden_states)
            residual_fix_overflow = True
        else:
            hidden_states, resi_new = self.input_layernorm(
                hidden_states, residual)
            residual = resi_new
        new_hs, new_resi, xq, xs = self.self_attn(
            positions=positions,
            hidden_states=hidden_states,
            pa_rms_weight=self.post_attention_layernorm.weight.data,
            pa_residual=residual,
            pa_rms_eps=self.post_attention_layernorm.variance_epsilon,
            pa_quant_dtype = torch.int8,
            update_input=True
        )


        assert xq is not None and xs is not None
        if new_hs.dtype == torch.float16: # overflow处理逻辑
            new_hs *= 1. / self.routed_scaling_factor
            if self.layer_idx == 0 or residual_fix_overflow:
                new_resi *= 1. / self.routed_scaling_factor

        hidden_states = self.mlp(new_hs, xqxs=(xq, xs))

        if isinstance(self.mlp,
                    DeepseekV2MLP) and hidden_states.dtype == torch.float16:
            hidden_states *= 1. / self.routed_scaling_factor
        return hidden_states, new_resi

    def forward_default(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor]
        ) -> Tuple[torch.Tensor, torch.Tensor]:
        # Self Attention
        # Fix residual FP16 overflow
        residual_fix_overflow = False
        if residual is None:
            residual = hidden_states
            hidden_states = self.input_layernorm(hidden_states)
            residual_fix_overflow = True
        else:
            hidden_states, residual = self.input_layernorm(
                hidden_states, residual)

        if not self.is_mtp_layer and self.enable_ep_sp and \
            self.layer_idx > self.config.first_k_dense_replace:
            hidden_states = tensor_model_parallel_all_gather(hidden_states, dim=0)

        hidden_states = self.self_attn(
            positions=positions,
            hidden_states=hidden_states,
        )

        if not self.is_mtp_layer and self.enable_ep_sp:
            if self.layer_idx == self.config.first_k_dense_replace:
                residual = residual.tensor_split(self.tp_size)[self.tp_rank]

            hidden_states = tensor_model_parallel_reduce_scatter(hidden_states, dim=0)
                
        if self.enable_dp_attention:
            if self.tp_rank == 0:
                hidden_states += residual
            hidden_states = dp_gather(hidden_states)

        if hidden_states.dtype == torch.float16:
            # Fix FP16 overflow
            # We scale both hidden_states and residual before
            # rmsnorm, and rmsnorm result would not affect by scale.
            hidden_states *= 1. / self.routed_scaling_factor
            if self.layer_idx == 0 or residual_fix_overflow:
                # The residual is shared by all layers, we only scale it on
                # first layer.
                residual *= 1. / self.routed_scaling_factor

        # Fully Connected
        if not self.enable_dp_attention:
            hidden_states, residual = self.post_attention_layernorm(
                hidden_states, residual)
        else:
            num_tokens = hidden_states.shape[0]
            new_bs = num_tokens // get_moe_tp_size() * get_attention_tp_size()
            residual = hidden_states[self.dp_rank*new_bs: (self.dp_rank+1)*new_bs, :]
            hidden_states = self.post_attention_layernorm(hidden_states)

        if self.is_mtp_layer and self.enable_ep_sp:
            ori_bs = hidden_states.shape[0]
            pad_size = (ori_bs + self.tp_size - 1) // self.tp_size * self.tp_size - ori_bs
            if pad_size > 0:
                hidden_states = torch.nn.functional.pad(hidden_states, [0, 0, 0, pad_size], value=0)
            new_bs = (ori_bs+pad_size) // self.tp_size
            hidden_states = hidden_states[self.tp_rank*new_bs: (self.tp_rank+1)*new_bs, :]


        hidden_states = self.mlp(hidden_states)
        
        if self.enable_dp_attention:
            hidden_states = dp_reduce_scatter_tensor(hidden_states)

        if self.is_mtp_layer and self.enable_ep_sp:
            hidden_states = tensor_model_parallel_all_gather(hidden_states, dim=0)
            hidden_states = hidden_states[:ori_bs, :]

        if isinstance(self.mlp,
                    DeepseekV2MLP) and hidden_states.dtype == torch.float16:
            # Fix FP16 overflow
            # Scaling the DeepseekV2MLP output, it is the input of
            # input_layernorm of next decoder layer.
            # The scaling of DeepseekV2MOE output would be done in the forward
            # of DeepseekV2MOE
            hidden_states *= 1. / self.routed_scaling_factor

        return hidden_states, residual

    def choose_forward(self):
        if self.use_fused_rms_quant:
            return self.forward_fused_RQ

        elif self.use_fused_custom_all_reduce:
            return self.forward_fused_CRQ
        else:
            return self.forward_default

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor]
    )  -> Tuple[torch.Tensor, torch.Tensor]:
        forward_func = self.choose_forward()
        return forward_func(positions=positions, hidden_states=hidden_states, residual=residual )


@support_torch_compile
class DeepseekV2Model(nn.Module):

    fall_back_to_pt_during_load = False

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()

        config = vllm_config.model_config.hf_config
        model_config = vllm_config.model_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
        enable_eplb = vllm_config.parallel_config.enable_eplb
        self.config = config

        self.vocab_size = config.vocab_size

        if get_pp_group().is_first_rank:
            self.embed_tokens = VocabParallelEmbedding(
                config.vocab_size,
                config.hidden_size,
                quant_config=quant_config,
                prefix=f"{prefix}.embed_tokens")
        else:
            self.embed_tokens = PPMissingLayer()

        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: DeepseekV2DecoderLayer(
                config,
                prefix,
                model_config=model_config,
                cache_config=cache_config,
                quant_config=quant_config,
                enable_eplb=enable_eplb,
            ),
            prefix=f"{prefix}.layers")

        if get_pp_group().is_last_rank:
            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        else:
            self.norm = PPMissingLayer()
        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(
                ["hidden_states", "residual"], config.hidden_size))

        self.dp_size = get_dp_group().world_size
        vllm_config = get_current_vllm_config()
        parallel_config = vllm_config.parallel_config
        self.use_deepep = self.dp_size > 1 and parallel_config.enable_expert_parallel and \
            (envs.VLLM_ALL2ALL_BACKEND == "deepep_high_throughput" or \
             envs.VLLM_ALL2ALL_BACKEND == "deepep_low_latency" or \
             envs.VLLM_ALL2ALL_BACKEND == "deepep_auto")
        self.tp_size = get_tensor_model_parallel_world_size()

    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.embed_tokens(input_ids)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors],
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        if get_pp_group().is_first_rank:
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
                hidden_states = self.get_input_embeddings(input_ids)
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

        for layer in self.layers[self.start_layer:self.end_layer]:
            hidden_states, residual = layer(positions, hidden_states, residual)

        if not get_pp_group().is_last_rank:
            return IntermediateTensors({
                "hidden_states": hidden_states,
                "residual": residual
            })

        hidden_states, _ = self.norm(hidden_states, residual)

        if self.use_deepep and self.tp_size > 1:
            hidden_states = tensor_model_parallel_all_gather(hidden_states, dim=0)

        return hidden_states


class DeepseekV2ForCausalLM(nn.Module, SupportsPP, MixtureOfExperts):

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config

        self.quant_method = None
        if quant_config is not None:
            self.quant_method = quant_config.get_name()
            os.environ['LLAMA_NN'] = '0'
            os.environ['LM_NN'] = '0'

        self.use_w4a16_moe_sz = os.environ.get('AWQ_MOE_SZ') == '1'
        self.config = config
        self.quant_config = quant_config

        self.model = DeepseekV2Model(vllm_config=vllm_config,
                                     prefix=maybe_prefix(prefix, "model"))
        if get_pp_group().is_last_rank:
            self.lm_head = ParallelLMHead(config.vocab_size,
                                          config.hidden_size,
                                          quant_config=quant_config)
        else:
            self.lm_head = PPMissingLayer()
        self.logits_processor = LogitsProcessor(config.vocab_size)
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors)
        self.expert_weights = []

        # Set MoE hyperparameters
        self.num_moe_layers = (config.num_hidden_layers -
                               config.first_k_dense_replace)
        self.num_expert_groups = config.n_group

        self.moe_layers: list[FusedMoE] = []
        example_moe = None
        for layer in self.model.layers:
            if isinstance(layer, PPMissingLayer):
                continue

            assert isinstance(layer, DeepseekV2DecoderLayer)
            if isinstance(layer.mlp, DeepseekV2MoE):
                example_moe = layer.mlp
                self.moe_layers.append(layer.mlp.experts)

        # Pick last one layer since the first ones may be dense layers.
        self.num_logical_experts = example_moe.n_logical_experts
        self.num_physical_experts = example_moe.n_physical_experts
        self.num_local_physical_experts = example_moe.n_local_physical_experts
        self.num_routed_experts = example_moe.n_routed_experts
        self.num_shared_experts = example_moe.n_shared_experts
        self.num_redundant_experts = example_moe.n_redundant_experts

        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
        self.use_awq_pad = os.environ.get('AWQ_PAD') == '1'
        self.tritonsingleton= W8a8GetCacheJSON()
        self.tritonsingleton.topk = config.num_experts_per_tok
        self.tritonsingleton.quant_method=self.quant_method

    def set_eplb_state(
        self,
        expert_load_view: torch.Tensor,
        logical_to_physical_map: torch.Tensor,
        logical_replica_count: torch.Tensor,
    ) -> None:
        for layer_idx, layer in enumerate(self.moe_layers):
            # Register the expert weights.
            self.expert_weights.append(layer.get_expert_weights())
            layer.set_eplb_state(
                moe_layer_idx=layer_idx,
                expert_load_view=expert_load_view,
                logical_to_physical_map=logical_to_physical_map,
                logical_replica_count=logical_replica_count,
            )

    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.get_input_embeddings(input_ids)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        hidden_states = self.model(input_ids, positions, intermediate_tensors,
                                   inputs_embeds)
        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
        logits = self.logits_processor(self.lm_head, hidden_states,
                                       sampling_metadata)
        return logits

    def make_empty_intermediate_tensors(
            self, batch_size: int, dtype: torch.dtype,
            device: torch.device) -> IntermediateTensors:
        return IntermediateTensors({
            "hidden_states":
            torch.zeros((batch_size, self.config.hidden_size),
                        dtype=dtype,
                        device=device),
            "residual":
            torch.zeros((batch_size, self.config.hidden_size),
                        dtype=dtype,
                        device=device),
        })

    def restore_qzeros_tensor(self, qzeros, qscales):

        low_bits = qzeros & 0x0F
        high_bits = qzeros >> 4

        zeors_tensor = torch.stack([low_bits, high_bits], dim=2).view(qzeros.shape[0], -1 , qzeros.shape[-1])
        zeors_int16 = zeors_tensor.to(torch.int16)
        assert zeors_int16.shape == qscales.shape

        uint16_tensor1 = zeors_int16.view(torch.uint16)
        uint16_tensor2 = qscales.view(torch.uint16)

        uint32_tensor1 = uint16_tensor1.to(torch.int32) << 16
        uint32_tensor2 = uint16_tensor2.to(torch.int32)

        result_tensor = uint32_tensor1 + uint32_tensor2
        result_tensor =result_tensor.view(torch.uint32)
        result_tensor = result_tensor.transpose(1, 2).contiguous()
        return result_tensor

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
        ]
        if envs.USE_FUSED_RMS_QUANT and envs.VLLM_USE_FUSED_QA_KVA_GEMM:
            fused_params_mapping = [
                ("qa_kva_proj", "q_a_proj", 0),
                ("qa_kva_proj", "kv_a_proj_with_mqa", 1)
            ]
            stacked_params_mapping += fused_params_mapping

        # Params for weights, fp8 weight scales, fp8 activation scales
        # (param_name, weight_name, expert_id, shard_id)
        expert_params_mapping = FusedMoE.make_expert_params_mapping(
            ckpt_gate_proj_name="gate_proj",
            ckpt_down_proj_name="down_proj",
            ckpt_up_proj_name="up_proj",
            num_experts=self.config.n_routed_experts
            + ( 
                self.num_shared_experts
                if (envs.VLLM_ENABLE_SHARED_EXPERTS_FUSION and self.num_shared_experts > 0)
                else 0
            ),
            num_redundant_experts=self.num_redundant_experts)

        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
            if "rotary_emb.inv_freq" in name:
                continue

            spec_layer = get_spec_layer_idx_from_weight_name(self.config, name)
            if spec_layer is not None:
                continue  # skip spec decode layers for main model

            # Assuame num of shared experts is only one.
            if envs.VLLM_ENABLE_SHARED_EXPERTS_FUSION \
                    and self.num_shared_experts > 0 \
                    and "mlp.shared_experts" in name:
                name = name.replace(
                    "mlp.shared_experts",
                    f"mlp.experts.{self.config.n_routed_experts}",
                )

            for (param_name, weight_name, shard_id) in stacked_params_mapping:
                # Skip non-stacked layers and experts (experts handled below).
                if weight_name not in name:
                    continue
                # We have mlp.experts[0].gate_proj in the checkpoint.
                # Since we handle the experts below in expert_params_mapping,
                # we need to skip here BEFORE we update the name, otherwise
                # name will be updated to mlp.experts[0].gate_up_proj, which
                # will then be updated below in expert_params_mapping
                # for mlp.experts[0].gate_gate_up_proj, which breaks load.
                if (("mlp.experts." in name) and name not in params_dict):
                    continue
                old_weight_name = name
                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

                if is_pp_missing_parameter(name, self):
                    continue

                param = params_dict[name]
                weight_loader = param.weight_loader

                if envs.USE_FUSED_RMS_QUANT and envs.VLLM_USE_FUSED_QA_KVA_GEMM and (("q_a_proj"  in old_weight_name) or ("kv_a_proj_with_mqa" in old_weight_name)):
                    weight_loader(param, loaded_weight, old_weight_name)
                else:
                    weight_loader(param, loaded_weight, shard_id)
                break
            else:
                is_expert_weight = False
                for mapping in expert_params_mapping:
                    param_name, weight_name, expert_id, shard_id = mapping
                    if weight_name not in name:
                        continue

                    # Anyway, this is an expert weight and should not be
                    # attempted to load as other weights later
                    is_expert_weight = True

                    # Do not modify `name` since the loop may continue here
                    # Instead, create a new variable
                    name_mapped = name.replace(weight_name, param_name)

                    if is_pp_missing_parameter(name_mapped, self):
                        continue

                    param = params_dict[name_mapped]
                    # We should ask the weight loader to return success or not
                    # here since otherwise we may skip experts with other
                    # available replicas.
                    weight_loader = typing.cast(Callable[..., bool],
                                                param.weight_loader)
                    success = weight_loader(param,
                                            loaded_weight,
                                            name_mapped,
                                            shard_id=shard_id,
                                            expert_id=expert_id,
                                            return_success=True)
                    if success:
                        name = name_mapped
                        break
                else:
                    if is_expert_weight:
                        # We've checked that this is an expert weight
                        # However it's not mapped locally to this rank
                        # So we simply skip it
                        continue

                    # Skip loading extra bias for GPTQ models.
                    if name.endswith(".bias") and name not in params_dict:
                        continue

                    # Remapping the name of FP8 kv-scale.
                    name = maybe_remap_kv_scale_name(name, params_dict)
                    if name is None:
                        continue

                    if is_pp_missing_parameter(name, self):
                        continue

                    try:
                        param = params_dict[name]
                    except Exception as e:
                        continue
                    weight_loader = getattr(param, "weight_loader",
                                            default_weight_loader)
                    weight_loader(param, loaded_weight)
            loaded_params.add(name)

        if self.use_llama_nn and self.quant_method is None:
            lay_key_words = [
                "self_attn.q_proj.weight",
                "self_attn.q_a_proj.weight",
                "self_attn.q_b_proj.weight",
                "self_attn.kv_a_proj_with_mqa.weight",
                "self_attn.kv_b_proj.weight",
                "self_attn.o_proj.weight",
                "mlp.gate_up_proj.weight",
                "mlp.down_proj.weight",
                "mlp.gate.weight",
                "shared_experts.gate_up_proj.weight",
                "shared_experts.down_proj.weight",
                "lm_head.weight",
            ]

            combined_words = "|".join(lay_key_words)

            for layername in loaded_params:
                weight = params_dict[layername]
                matches = re.findall(combined_words, layername)
                if matches:
                    _weight = torch.zeros_like(weight.data)
                    ori_shape =_weight.shape

                    ops.trans_w16_gemm(_weight, weight.data, _weight.shape[0], _weight.shape[1])
                    weight.data.copy_(_weight)

                    weight.data=weight.data.reshape(ori_shape[1],-1)

        return loaded_params


class DeepseekV3ForCausalLM(DeepseekV2ForCausalLM):
    pass


def get_spec_layer_idx_from_weight_name(config: PretrainedConfig,
                                        weight_name: str) -> Optional[int]:
    if hasattr(config,
               "num_nextn_predict_layers") and (config.num_nextn_predict_layers
                                                > 0):
        layer_idx = config.num_hidden_layers
        for i in range(config.num_nextn_predict_layers):
            if weight_name.startswith(f"model.layers.{layer_idx+i}."):
                return layer_idx + i
    return None