llama.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 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Inference-only LLaMA model compatible with HuggingFace weights."""
from typing import Any, Dict, Iterable, List, Optional, Set, Tuple, Type, Union

import torch
from torch import nn
from transformers import LlamaConfig
import os
import re
import vllm.envs as envs
from vllm.attention import Attention, AttentionMetadata
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
                                               QKVParallelLinear,
                                               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.sampler import SamplerOutput, get_sampler
from vllm.model_executor.layers.vocab_parallel_embedding import (
    DEFAULT_VOCAB_PADDING_SIZE, 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 SupportsLoRA, SupportsPP
from .utils import (AutoWeightsLoader, PPMissingLayer, extract_layer_index,
                    is_pp_missing_parameter,
                    make_empty_intermediate_tensors_factory, make_layers,
                    maybe_prefix)

from vllm import _custom_ops as ops
from vllm.model_executor.utils import pad_weight, gemm_bank_conf


class LlamaMLP(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
        quant_config: Optional[QuantizationConfig] = None,
        bias: bool = False,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
            input_size=hidden_size,
            output_sizes=[intermediate_size] * 2,
            bias=bias,
            quant_config=quant_config,
            prefix=f"{prefix}.gate_up_proj",
        )
        self.down_proj = RowParallelLinear(
            input_size=intermediate_size,
            output_size=hidden_size,
            bias=bias,
            quant_config=quant_config,
            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):
        x, _ = self.gate_up_proj(x)
        x = self.act_fn(x)
        x, _ = self.down_proj(x)
        return x


class LlamaAttention(nn.Module):

    def __init__(self,
                 config: LlamaConfig,
                 hidden_size: int,
                 num_heads: int,
                 num_kv_heads: int,
                 rope_theta: float = 10000,
                 rope_scaling: Optional[Dict[str, Any]] = None,
                 max_position_embeddings: int = 8192,
                 quant_config: Optional[QuantizationConfig] = None,
                 bias: bool = False,
                 bias_o_proj: bool = False,
                 cache_config: Optional[CacheConfig] = None,
                 prefix: str = "") -> None:
        super().__init__()
        layer_idx = extract_layer_index(prefix)
        self.hidden_size = hidden_size
        tp_size = get_tensor_model_parallel_world_size()
        self.total_num_heads = num_heads
        assert self.total_num_heads % tp_size == 0
        self.num_heads = self.total_num_heads // tp_size
        self.total_num_kv_heads = num_kv_heads
        if self.total_num_kv_heads >= tp_size:
            # Number of KV heads is greater than TP size, so we partition
            # the KV heads across multiple tensor parallel GPUs.
            assert self.total_num_kv_heads % tp_size == 0
        else:
            # Number of KV heads is less than TP size, so we replicate
            # the KV heads across multiple tensor parallel GPUs.
            assert tp_size % self.total_num_kv_heads == 0
        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
        # MistralConfig has an optional head_dim introduced by Mistral-Nemo
        self.head_dim = getattr(config, "head_dim",
                                self.hidden_size // self.total_num_heads)
        self.q_size = self.num_heads * self.head_dim
        self.kv_size = self.num_kv_heads * self.head_dim
        self.scaling = self.head_dim**-0.5
        self.rope_theta = rope_theta
        self.max_position_embeddings = max_position_embeddings

        self.qkv_proj = QKVParallelLinear(
            hidden_size=hidden_size,
            head_size=self.head_dim,
            total_num_heads=self.total_num_heads,
            total_num_kv_heads=self.total_num_kv_heads,
            bias=bias,
            quant_config=quant_config,
            prefix=f"{prefix}.qkv_proj",
        )

        self.o_proj = RowParallelLinear(
            input_size=self.total_num_heads * self.head_dim,
            output_size=hidden_size,
            bias=bias_o_proj,
            quant_config=quant_config,
            prefix=f"{prefix}.o_proj",
        )

        is_neox_style = True
        is_gguf = quant_config and quant_config.get_name() == "gguf"
        if is_gguf and config.model_type == "llama":
            is_neox_style = False

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

        if hasattr(config, "interleaved_sliding_window"):
            interleaved_sliding_window = config.interleaved_sliding_window
            if isinstance(interleaved_sliding_window, int):
                sliding_window = interleaved_sliding_window
            elif isinstance(interleaved_sliding_window, list):
                sw_idx = layer_idx % len(interleaved_sliding_window)
                sliding_window = interleaved_sliding_window[sw_idx]
            else:
                raise ValueError(
                    f"{type(interleaved_sliding_window)} is not supported.")
        else:
            sliding_window = None

        self.attn = Attention(
            self.num_heads,
            self.head_dim,
            self.scaling,
            num_kv_heads=self.num_kv_heads,
            cache_config=cache_config,
            quant_config=quant_config,
            per_layer_sliding_window=sliding_window,
            prefix=f"{prefix}.attn",
        )
        
        self.quant_method = None
        if quant_config is not None:
            self.quant_method=quant_config.get_name()
            self.quant_config=quant_config

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        kv_cache: torch.Tensor,
        attn_metadata: AttentionMetadata,
    ) -> torch.Tensor:
        qkv, _ = self.qkv_proj(hidden_states)
        # if os.environ.get('FA_PAD') == '1' and self.quant_method is None:
        #     qkv = qkv[...,:-32]
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
        q, k = self.rotary_emb(positions, q, k)
        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
        output, _ = self.o_proj(attn_output)
        return output


class LlamaDecoderLayer(nn.Module):

    def __init__(
        self,
        config: LlamaConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        rope_theta = getattr(config, "rope_theta", 10000)
        rope_scaling = getattr(config, "rope_scaling", None)
        if rope_scaling is not None and getattr(
                config, "original_max_position_embeddings", None):
            rope_scaling["original_max_position_embeddings"] = (
                config.original_max_position_embeddings)
        max_position_embeddings = getattr(config, "max_position_embeddings",
                                          8192)
        # Support abacusai/Smaug-72B-v0.1 with attention_bias
        # Support internlm/internlm-7b with bias
        attention_bias = getattr(config, "attention_bias", False) or getattr(
            config, "bias", False)
        bias_o_proj = attention_bias
        # support internlm/internlm3-8b with qkv_bias
        if hasattr(config, 'qkv_bias'):
            attention_bias = config.qkv_bias

        self.self_attn = LlamaAttention(
            config=config,
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            num_kv_heads=getattr(config, "num_key_value_heads",
                                 config.num_attention_heads),
            rope_theta=rope_theta,
            rope_scaling=rope_scaling,
            max_position_embeddings=max_position_embeddings,
            quant_config=quant_config,
            bias=attention_bias,
            bias_o_proj=bias_o_proj,
            cache_config=cache_config,
            prefix=f"{prefix}.self_attn",
        )
        self.mlp = LlamaMLP(
            hidden_size=self.hidden_size,
            intermediate_size=config.intermediate_size,
            hidden_act=config.hidden_act,
            quant_config=quant_config,
            bias=getattr(config, "mlp_bias", False),
            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)

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

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


@support_torch_compile
class LlamaModel(nn.Module):

    def __init__(self,
                 *,
                 vllm_config: VllmConfig,
                 prefix: str = "",
                 layer_type: Type[LlamaDecoderLayer] = LlamaDecoderLayer):
        super().__init__()

        config = vllm_config.model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
        lora_config = vllm_config.lora_config

        self.config = config
        self.quant_config = quant_config
        self.padding_idx = config.pad_token_id
        lora_vocab = (lora_config.lora_extra_vocab_size *
                      (lora_config.max_loras or 1)) if lora_config else 0
        self.vocab_size = config.vocab_size + lora_vocab
        self.org_vocab_size = config.vocab_size
        if get_pp_group().is_first_rank or (config.tie_word_embeddings
                                            and get_pp_group().is_last_rank):
            self.embed_tokens = VocabParallelEmbedding(
                self.vocab_size,
                config.hidden_size,
                org_num_embeddings=config.vocab_size,
                quant_config=quant_config,
            )
        else:
            self.embed_tokens = PPMissingLayer()
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: layer_type(config=config,
                                      cache_config=cache_config,
                                      quant_config=quant_config,
                                      prefix=prefix),
            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.quant_method = None
        if quant_config is not None:
            self.quant_method=quant_config.get_name()
            self.quant_config=quant_config
            
        self.tritonsingleton= W8a8GetCacheJSON()      
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
        # self.use_lm_nn = os.environ.get('LM_NN') == '1'
        self.use_gemm_pad = os.environ.get('GEMM_PAD') == '1'
        self.use_fa_pad = os.environ.get('FA_PAD') == '1'
        self.use_awq_pad = os.environ.get('AWQ_PAD') == '1'
        self.w8a8_strategy=int(os.getenv('W8A8_SUPPORT_METHODS', '1'))

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

    def forward(
        self,
        input_ids: Optional[torch.Tensor],
        positions: torch.Tensor,
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
        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 i in range(self.start_layer, self.end_layer):
            layer = self.layers[i]
            hidden_states, residual = layer(positions, hidden_states,
                                            kv_caches[i - self.start_layer],
                                            attn_metadata, 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

    def load_weights(self, weights: Iterable[Tuple[str,
                                                   torch.Tensor]]) -> Set[str]:
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            (".qkv_proj", ".q_proj", "q"),
            (".qkv_proj", ".k_proj", "k"),
            (".qkv_proj", ".v_proj", "v"),
            (".gate_up_proj", ".gate_proj", 0),
            (".gate_up_proj", ".up_proj", 1),
        ]
        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
            if ("rotary_emb.cos_cached" in name
                    or "rotary_emb.sin_cached" in name):
                # Models trained using ColossalAI may include these tensors in
                # the checkpoint. Skip them.
                continue
            if (self.quant_config is not None and
                (scale_name := self.quant_config.get_cache_scale(name))):
                # Loading kv cache quantization scales
                param = params_dict[scale_name]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else
                                 loaded_weight[0])
                weight_loader(param, loaded_weight)
                loaded_params.add(scale_name)
                continue
            if "scale" in name:
                # Remapping the name of FP8 kv-scale.
                name = maybe_remap_kv_scale_name(name, params_dict)
                if name is None:
                    continue
            for param_name, weight_name, shard_id in stacked_params_mapping:
                if weight_name not in name:
                    continue
                name = name.replace(weight_name, param_name)
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue

                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:
                # 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 = 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.qkv_proj.weight",
                "self_attn.o_proj.weight",
                "mlp.gate_up_proj.weight",
                "mlp.down_proj.weight",
            ]
            combined_words = "|".join(lay_key_words)
            
            # lay_qkv_words = ["self_attn.qkv_proj.weight"]   
            # qkv_words = "|".join(lay_qkv_words)          
            
            # for layername, weight in params_dict.items():
            for layername in loaded_params:
                weight = params_dict[layername]
                if "lm_head.weight" in layername and weight.shape[1] >= 4096:
                    lay_key_words.append("lm_head.weight")
                    combined_words = "|".join(lay_key_words)
                    os.environ['LM_NN'] = '1'  
                else:
                    os.environ['LM_NN'] = '0' 
                    
                matches = re.findall(combined_words, layername)
                
                if matches:         
                    # if self.use_gemm_pad and gemm_bank_conf(weight.data.shape[0]):
                    #     weight.data = pad_weight(weight.data, 32)  
                        
                    # if self.use_fa_pad and (re.findall(qkv_words, layername)):
                    #     if not gemm_bank_conf(weight.data.shape[0]):
                    #         weight.data = pad_weight(weight.data, 32)
                                 
                    _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)
        else:
            os.environ['LM_NN'] = '0'
            os.environ['LLAMA_NN'] = '0'
            
        if self.quant_method == "awq" and not envs.VLLM_USE_TRITON_AWQ:
            lay_key_words = [
                "self_attn.qkv_proj.qweight",
                "self_attn.o_proj.qweight",
                "mlp.gate_up_proj.qweight",
                "mlp.down_proj.qweight"
            ]
            combined_words = "|".join(lay_key_words)
            
            for layername in loaded_params:
                weight = params_dict[layername]
                
                matches = re.findall(combined_words, layername)
                if matches:
                    qweight =params_dict[layername]
                    qzeros=params_dict[layername.replace("qweight", "qzeros")]
                    scales=params_dict[layername.replace("qweight", "scales")]
                    zeros_and_scalse =params_dict[layername.replace("qweight", "zeros_and_scales")]
                    
                    group_size= self.quant_config.group_size 
                   
                    dim_n = scales.data.shape[1]
                    dim_k = qweight.data.shape[0]
                    pad_group=2              
                    
                    _qw, _sz=ops.convert_s4(qweight,qzeros,scales,int(group_size)) 
                    
                    sz = ops.sz_permute(_sz).reshape(-1,dim_n)       
                    
                    zeros_and_scalse.data.copy_(sz)
                    qweight.data.copy_(_qw)
                    
                    #reshape
                    zeros_and_scalse.data=zeros_and_scalse.reshape(dim_n,-1)    #[k/greop_size,n]------>[n,k/group_size]
                    qweight.data=qweight.data.reshape(dim_n,-1)                      #[k,n/8]---->[n,k/8]  
                
                    if dim_k % 4096==0 and self.use_awq_pad:
                        zeros_and_scalse_pad= torch.zeros(dim_n,pad_group,dtype=torch.int32).cuda()
                        zeros_and_scalse.data=torch.cat((zeros_and_scalse.data,zeros_and_scalse_pad),dim=1).contiguous()
                        qweight_pad= torch.zeros(dim_n,int(group_size//4),dtype=torch.int32).cuda()
                        qweight.data=torch.cat((qweight.data,qweight_pad),dim=1).contiguous()
            
        #当为triton支持推理的时候不能进行处理
        if self.quant_method == "compressed_tensors":
            lay_key_words = [
                "self_attn.qkv_proj.weight",
                "self_attn.o_proj.weight",
                "mlp.gate_up_proj.weight",
                "mlp.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]
                    # k=weight_data.shape[1]
                    
                    # #判断当前size是否在优化的范围内，假如存在则走triton，假如不存在则走rocblas
                    # json_file=self.tritonsingleton.get_w8a8json_name(n,k)
                    
                    #rocblas和cutlass目前都需要weight做处理，但是triton不用
                    if self.w8a8_strategy!=1:
                        _weight=weight_data.T.contiguous().reshape(n,-1)
                        weight_data.data.copy_(_weight)  
                    
                    #下面是针对模型记录模型出现k和n值 
                    elif len(matched_key_words) < 4 and matches[0] not in matched_key_words:
                        matched_key_words.add(matches[0])
                        k=weight_data.shape[1]
                        weight_shapes.append({n,k})
                
                        json_file=self.tritonsingleton.get_w8a8json_name(n,k)
                        configs_dict=self.tritonsingleton.get_triton_cache(json_file,n,k)
                        if configs_dict:
                            all_json.update(configs_dict)
                                              
            if self.w8a8_strategy==1:
                self.tritonsingleton.triton_json_dict.append(all_json)
                #找到的所有config都进行一次warmup
                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 LlamaForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
    packed_modules_mapping = {
        "qkv_proj": ["q_proj", "k_proj", "v_proj"],
        "gate_up_proj": ["gate_proj", "up_proj"]
    }

    # LoRA specific attributes
    supported_lora_modules = [
        "qkv_proj", "o_proj", "gate_up_proj", "down_proj", "embed_tokens",
        "lm_head"
    ]
    embedding_modules = {
        "embed_tokens": "input_embeddings",
        "lm_head": "output_embeddings"
    }
    embedding_padding_modules = ["lm_head"]

    # Mistral/Llama models can also be loaded with --load-format mistral
    # from consolidated.safetensors checkpoints
    mistral_mapping = {
        "layers": "model.layers",
        "attention": "self_attn",
        "wq": "q_proj",
        "wk": "k_proj",
        "wv": "v_proj",
        "wo": "o_proj",
        "attention_norm": "input_layernorm",
        "feed_forward": "mlp",
        "w1": "gate_proj",
        "w2": "down_proj",
        "w3": "up_proj",
        "ffn_norm": "post_attention_layernorm",
        "tok_embeddings": "model.embed_tokens",
        "output": "lm_head",
        "norm": "model.norm"
    }

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

        if get_pp_group().is_last_rank:
            self.unpadded_vocab_size = config.vocab_size
            if lora_config:
                self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
            self.lm_head = ParallelLMHead(
                self.unpadded_vocab_size,
                config.hidden_size,
                org_num_embeddings=config.vocab_size,
                padding_size=(
                    DEFAULT_VOCAB_PADDING_SIZE
                    # We need bigger padding if using lora for kernel
                    # compatibility
                    if not lora_config else
                    lora_config.lora_vocab_padding_size),
                quant_config=quant_config,
                prefix=maybe_prefix(prefix, "lm_head"),
            )
            if config.tie_word_embeddings:
                self.lm_head = self.lm_head.tie_weights(
                    self.model.embed_tokens)

            logit_scale = getattr(config, "logit_scale", 1.0)
            self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                                    config.vocab_size,
                                                    logit_scale)
        else:
            self.lm_head = PPMissingLayer()

        self.sampler = get_sampler()

        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors)
        

    def _init_model(self, vllm_config: VllmConfig, prefix: str = ""):
        return LlamaModel(vllm_config=vllm_config, prefix=prefix)

    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,
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        model_output = self.model(input_ids, positions, kv_caches,
                                  attn_metadata, intermediate_tensors,
                                  inputs_embeds)
        return model_output

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


    def load_weights(self, weights: Iterable[Tuple[str,
                                                   torch.Tensor]]) -> Set[str]:
        loader = AutoWeightsLoader(
            self,
            skip_prefixes=(["lm_head."]
                           if self.config.tie_word_embeddings else None),
        )
        return loader.load_weights(
            self.maybe_remap_mistral(name, loaded_weight)
            for name, loaded_weight in weights)


    # This function is used to remap the mistral format as
    # used by Mistral and Llama <=2
    def maybe_remap_mistral(
        self,
        name: str,
        loaded_weight: torch.Tensor,
    ) -> Tuple[str, torch.Tensor]:

        def permute(w: torch.Tensor, n_heads: int):
            attn_in = self.config.head_dim * n_heads
            attn_out = self.config.hidden_size

            return w.view(n_heads, attn_in // n_heads // 2, 2,
                          attn_out).transpose(1, 2).reshape(attn_in, attn_out)

        mapping = self.mistral_mapping
        modules = name.split(".")

        # rotary embeds should be sliced
        if "wk" in modules:
            loaded_weight = permute(loaded_weight,
                                    self.config.num_key_value_heads)
        elif "wq" in modules:
            loaded_weight = permute(loaded_weight,
                                    self.config.num_attention_heads)

        for item in modules:
            if item in mapping and mapping[item] not in name:
                name = name.replace(item, mapping[item])

        return name, loaded_weight