deepseek_v2.py

# SPDX-License-Identifier: Apache-2.0

# 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
from typing import Any, Dict, Iterable, Optional, Set, Tuple, Union

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
from vllm.distributed import (get_pp_group,
                              get_tensor_model_parallel_world_size,
                              tensor_model_parallel_all_reduce)
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                               MergedColumnParallelLinear,
                                               ReplicatedLinear,
                                               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.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 vllm.utils import W8a8GetCacheJSON

from .interfaces import 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.model_executor.layers.quantization.utils.int8_utils import (
    block_dequant as int8_block_dequant,
)


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__()
        self.gate_up_proj = MergedColumnParallelLinear(
            hidden_size, [intermediate_size] * 2,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.gate_up_proj")
        self.down_proj = RowParallelLinear(intermediate_size,
                                           hidden_size,
                                           bias=False,
                                           quant_config=quant_config,
                                           reduce_results=reduce_results,
                                           prefix=f"{prefix}.down_proj")
        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):
        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 = "",
    ):
        super().__init__()
        self.tp_size = get_tensor_model_parallel_world_size()
        self.routed_scaling_factor = config.routed_scaling_factor
        self.n_shared_experts = 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

        self.experts = FusedMoE(
            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,)

        if config.n_shared_experts is not None:
            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",
            )
        from vllm.two_batch_overlap.two_batch_overlap import tbo_all_reduce, is_enable_tbo
        self.tbo_all_reduce = tbo_all_reduce
        self.enable_tbo = is_enable_tbo()

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        num_tokens, hidden_dim = hidden_states.shape
        hidden_states = hidden_states.view(-1, hidden_dim)
        if self.n_shared_experts is not None:
            shared_output = self.shared_experts(hidden_states)
        # router_logits: (num_tokens, n_experts)
        router_logits, _ = self.gate(hidden_states)
        # 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)
        
        final_hidden_states = self.experts(
            hidden_states=hidden_states,
            router_logits=router_logits) * self.routed_scaling_factor
        if shared_output is not None:
            final_hidden_states = final_hidden_states + shared_output
            
        # 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 self.enable_tbo:
                final_hidden_states = self.tbo_all_reduce(final_hidden_states)
            else:
                final_hidden_states = tensor_model_parallel_all_reduce(
                    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 = "",
    ) -> 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")
        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 = "",
    ) -> 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")
        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")

        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,
            rotary_emb=self.rotary_emb,
            q_proj=self.q_proj if self.q_lora_rank is None else self.q_b_proj,
            kv_b_proj=self.kv_b_proj,
            o_proj=self.o_proj,
        )

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

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        if self.q_lora_rank is not None:
            ckq = self.q_a_proj(hidden_states)[0]
            hidden_states_or_q_c = self.q_a_layernorm(ckq)
        else:
            hidden_states_or_q_c = hidden_states
        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)
        kv_c_normed = self.kv_a_layernorm(kv_c.contiguous())
        return self.mla_attn(hidden_states_or_q_c,
                             kv_c_normed,
                             k_pe,
                             output_shape=hidden_states.shape)


class DeepseekV2DecoderLayer(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        prefix: str,
        model_config: ModelConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = 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)
        # 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
        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",
        )

        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",
            )
        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.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

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        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,
        )

        # 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:
        #         # The residual is shared by all layers, we only scale it on
        #         # first layer.
        #         residual *= 1. / self.routed_scaling_factor

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

        # 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


@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
        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,
            ),
            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))

    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)
        return hidden_states


class DeepseekV2ForCausalLM(nn.Module, SupportsPP):

    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.tritonsingleton= W8a8GetCacheJSON()  
        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.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
        self.use_awq_pad = os.environ.get('AWQ_PAD') == '1'
        
    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),
        ]

        # 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)

        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

            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
                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
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                for mapping in expert_params_mapping:
                    param_name, weight_name, expert_id, shard_id = mapping
                    if weight_name not in name:
                        continue
                    name = name.replace(weight_name, param_name)

                    if is_pp_missing_parameter(name, self):
                        continue

                    param = params_dict[name]
                    weight_loader = param.weight_loader
                    weight_loader(param,
                                  loaded_weight,
                                  name,
                                  shard_id=shard_id,
                                  expert_id=expert_id)
                    break
                else:
                    # 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

                    param = params_dict[name]
                    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)

        if hasattr(self.config, "quantization_config") and self.config.quantization_config["quant_method"] == "awq" and not envs.VLLM_USE_TRITON_AWQ:
            lay_key_words = [
                "self_attn.q_a_proj.qweight",
                "self_attn.q_b_proj.qweight",
                "self_attn.kv_b_proj.qweight",
                "self_attn.kv_a_proj_with_mqa.qweight",
                "self_attn.o_proj.qweight",
                "mlp.gate_up_proj.qweight",
                "mlp.down_proj.qweight",
                "mlp.shared_experts.gate_up_proj.qweight",
                "mlp.shared_experts.down_proj.qweight"
            ]
            combined_words = "|".join(lay_key_words)
            # moe_gather_sz
            moe_key_words = ["mlp.experts.w13_qweight", "mlp.experts.w2_qweight"]
            moe_combined_words = "|".join(moe_key_words)
            
            for layername in loaded_params:
                weight = params_dict[layername]
                
                matches = re.findall(combined_words, layername)

                        
                if self.use_w4a16_moe_sz:
                    matches_moe = re.findall(moe_combined_words, layername)
                    # sz.shape == s.shape.T
                    if matches_moe:
                        qzeros=params_dict[layername.replace("qweight", "qzeros")]
                        scales=params_dict[layername.replace("qweight", "scales")]
                        sz_tensor = self.restore_qzeros_tensor(qzeros, scales)
                        scales.data = sz_tensor
            
        if hasattr(self.config, "quantization_config") and self.config.quantization_config["quant_method"] == "blockwise_int8":
            lay_key_words = [
                "self_attn.q_a_proj.weight",
                "self_attn.q_b_proj.weight",
                "self_attn.kv_b_proj.weight",
                "self_attn.kv_a_proj_with_mqa.weight",
                "self_attn.o_proj.weight",
                "mlp.gate_up_proj.weight",
                "mlp.down_proj.weight",
                "mlp.shared_experts.gate_up_proj.weight",
                "mlp.shared_experts.down_proj.weight"
            ]
            combined_words = "|".join(lay_key_words)
            weight_shapes=[]
            all_json={}
            matched_key_words=set()
            
            for layername, weight in params_dict.items():  
                matches = re.findall(combined_words, layername)
                if matches and "scale" not in layername:
                    weight_data =params_dict[layername]
                    n=weight_data.shape[0]
                    
                    if len(matched_key_words) < 9 and matches[0] not in matched_key_words:
                        matched_key_words.add(matches[0])
                        k=weight_data.shape[1]
                        weight_shapes.append({n,k})
                        #print("n:{},k:{}".format(n,k))
                
                        json_file=self.tritonsingleton.get_blockint8json_name(n,k,128,128)
                        configs_dict=self.tritonsingleton.get_blockint8_triton_cache(json_file,n,k,128,128)
                        if configs_dict:
                            all_json.update(configs_dict)
                            
            self.tritonsingleton.triton_json_dict.append(all_json) 
            #print("self.tritonsingleton.triton_json_dict[0].shape:",len(self.tritonsingleton.triton_json_dict[0]))    
            for key, value in all_json.items():
                m=int(key.split('_')[0])
                n=int(key.split('_')[1])
                k=int(key.split('_')[2])
                # ops.triton_int8_gemm_helper(m=m,n=n,k=k,per_token_act_quant=True,per_out_channel_weight_quant=True,use_bias=False,best_config=value)

        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