baichuan.py

# coding=utf-8
# 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 BaiChuan model compatible with HuggingFace weights."""
import math
from typing import Iterable, List, Optional, Tuple, Union

import torch
from torch import nn
from transformers import PretrainedConfig
import os
import re

from vllm.attention import Attention, AttentionMetadata
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, LoRAConfig
from vllm.distributed import (get_pp_group, get_tensor_model_parallel_rank,
                              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 Sampler, SamplerOutput
from vllm.model_executor.layers.vocab_parallel_embedding import (
    ParallelLMHead, VocabParallelEmbedding)
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.model_executor.sampling_metadata import SamplingMetadata
from vllm.sequence import IntermediateTensors

from .interfaces import SupportsLoRA, SupportsPP
from .utils import (is_pp_missing_parameter,
                    make_empty_intermediate_tensors_factory, make_layers)

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



def _get_alibi_slopes(total_num_heads: int) -> torch.Tensor:
    closest_power_of_2 = 2**math.floor(math.log2(total_num_heads))
    base = torch.tensor(
        2**(-(2**-(math.log2(closest_power_of_2) - 3))),
        dtype=torch.float32,
    )
    powers = torch.arange(1, 1 + closest_power_of_2, dtype=torch.int32)
    slopes = torch.pow(base, powers)

    if closest_power_of_2 != total_num_heads:
        extra_base = torch.tensor(
            2**(-(2**-(math.log2(2 * closest_power_of_2) - 3))),
            dtype=torch.float32,
        )
        num_remaining_heads = min(closest_power_of_2,
                                  total_num_heads - closest_power_of_2)
        extra_powers = torch.arange(start=1,
                                    end=1 + 2 * num_remaining_heads,
                                    step=2,
                                    dtype=torch.int32)
        slopes = torch.cat(
            [slopes, torch.pow(extra_base, extra_powers)], dim=0)
    return slopes


class BaiChuanMLP(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
        quant_config: Optional[QuantizationConfig] = None,
    ):
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
            hidden_size, [intermediate_size] * 2,
            bias=False,
            quant_config=quant_config)
        self.down_proj = RowParallelLinear(intermediate_size,
                                           hidden_size,
                                           bias=False,
                                           quant_config=quant_config)
        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 BaiChuanAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(
        self,
        hidden_size: int,
        num_heads: int,
        position_embedding: str,
        rope_theta: float = 10000,
        max_position_embeddings: int = 8192,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
    ):
        super().__init__()
        self.hidden_size = hidden_size
        tensor_model_parallel_world_size = get_tensor_model_parallel_world_size(
        )
        self.total_num_heads = num_heads
        assert self.total_num_heads % tensor_model_parallel_world_size == 0
        self.num_heads = (self.total_num_heads //
                          tensor_model_parallel_world_size)
        self.head_dim = hidden_size // self.total_num_heads
        self.postion_embedding = position_embedding
        self.rope_theta = rope_theta
        self.max_position_embeddings = max_position_embeddings

        # pylint: disable=invalid-name
        self.W_pack = QKVParallelLinear(
            hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_heads,
            bias=False,
            quant_config=quant_config,
        )
        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            hidden_size,
            bias=False,
            quant_config=quant_config,
        )
        # Create the alibi slopes and slice them.
        if self.postion_embedding == "ALIBI":
            tp_rank = get_tensor_model_parallel_rank()
            head_start = tp_rank * self.num_heads
            head_end = (tp_rank + 1) * self.num_heads
            alibi_slopes = _get_alibi_slopes(self.total_num_heads)
            alibi_slopes = alibi_slopes[head_start:head_end].tolist()

            scaling = self.head_dim**-0.5
            self.attn = Attention(self.num_heads,
                                  self.head_dim,
                                  scaling,
                                  alibi_slopes=alibi_slopes,
                                  quant_config=quant_config)
        else:
            self.rotary_emb = get_rope(
                self.head_dim,
                rotary_dim=self.head_dim,
                max_position=self.max_position_embeddings,
                base=self.rope_theta,
            )
            self.scaling = self.head_dim**-0.5
            self.attn = Attention(self.num_heads,
                                  self.head_dim,
                                  self.scaling,
                                  cache_config=cache_config,
                                  quant_config=quant_config)
            
        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.W_pack(hidden_states)
        if os.environ.get('FA_PAD') == '1' and self.quant_method is None:
            qkv = qkv[...,:-32]
        q, k, v = qkv.chunk(chunks=3, dim=-1)
        if self.postion_embedding != "ALIBI":
            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 BaiChuanDecoderLayer(nn.Module):

    def __init__(self,
                 config: PretrainedConfig,
                 position_embedding: str,
                 cache_config: Optional[CacheConfig] = None,
                 quant_config: Optional[QuantizationConfig] = None):
        super().__init__()
        self.hidden_size = config.hidden_size
        rope_theta = getattr(config, "rope_theta", 10000)
        max_position_embeddings = getattr(config, "max_position_embeddings",
                                          8192)
        self.self_attn = BaiChuanAttention(
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            position_embedding=position_embedding,
            rope_theta=rope_theta,
            max_position_embeddings=max_position_embeddings,
            cache_config=cache_config,
            quant_config=quant_config,
        )
        self.mlp = BaiChuanMLP(
            hidden_size=self.hidden_size,
            intermediate_size=config.intermediate_size,
            hidden_act=config.hidden_act,
            quant_config=quant_config,
        )
        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 BaiChuanModel(nn.Module):

    def __init__(self,
                 config: PretrainedConfig,
                 position_embedding: str,
                 cache_config: Optional[CacheConfig] = None,
                 quant_config: Optional[QuantizationConfig] = None,
                 prefix: str = ""):
        super().__init__()
        self.config = config
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        self.embed_tokens = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
        )
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: BaiChuanDecoderLayer(config, position_embedding,
                                                cache_config, quant_config),
            prefix=f"{prefix}.layers",
        )
        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(
                ["hidden_states", "residual"], config.hidden_size))

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
        intermediate_tensors: Optional[IntermediateTensors],
    ) -> Union[torch.Tensor, IntermediateTensors]:
        if get_pp_group().is_first_rank:
            hidden_states = self.embed_tokens(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


class BaiChuanBaseForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
    packed_modules_mapping = {
        "W_pack": ["W_pack"],
        "gate_up_proj": [
            "gate_proj",
            "up_proj",
        ],
    }
    # LoRA specific attributes
    supported_lora_modules = [
        "W_pack",
        "o_proj",
        "gate_up_proj",
        "down_proj",
    ]
    embedding_modules = {}
    embedding_padding_modules = []

    def __init__(
        self,
        config: PretrainedConfig,
        position_embedding: str,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        lora_config: Optional[LoRAConfig] = None,
    ):
        super().__init__()

        self.config = config
        self.lora_config = lora_config

        self.quant_config = quant_config
        self.model = BaiChuanModel(config, position_embedding, cache_config,
                                   quant_config)
        self.lm_head = ParallelLMHead(config.vocab_size,
                                      config.hidden_size,
                                      quant_config=quant_config)
        if self.config.tie_word_embeddings:
            self.lm_head.weight = self.model.embed_tokens.weight
        self.logits_processor = LogitsProcessor(config.vocab_size)
        self.sampler = Sampler()
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors)
        
        self.quant_method = None
        if quant_config is not None:
            self.quant_method=quant_config.get_name()
            self.quant_config=quant_config

        self.use_llama_nn = os.environ.get('LLAMA_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'
        
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
        intermediate_tensors: Optional[IntermediateTensors] = None,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        hidden_states = self.model(input_ids, positions, kv_caches,
                                   attn_metadata, intermediate_tensors)
        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 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]]):
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
        ]
        params_dict = dict(self.named_parameters())
        for name, loaded_weight in weights:
            if "rotary_emb.inv_freq" in name:
                continue
            if name == "lm_head.weight":
                # Unlike Baichuan, Baichuan2 normalizes the head weights.
                # Refer to:
                # https://huggingface.co/baichuan-inc/Baichuan2-7B-Chat/blob/84603cde5ebffb6084e476cfaeceaf0b8b91fe54/modeling_baichuan.py#L508
                # Distinguish between Baichuan and Baichuan2 by checking the
                # vocab size. This is suggested by
                # https://github.com/vllm-project/vllm/pull/1022#discussion_r1325652704
                is_baichuan2 = self.config.vocab_size == 125696
                if is_baichuan2:
                    loaded_weight = torch.nn.functional.normalize(
                        loaded_weight)

            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)
                
        if self.use_llama_nn and self.quant_method is None :
            lay_key_words = [
                "self_attn.W_pack.weight",
                "self_attn.o_proj.weight",
                "mlp.gate_up_proj.weight",
                "mlp.down_proj.weight",
                "lm_head.weight"
            ]
            combined_words = "|".join(lay_key_words)
            
            lay_qkv_words = ["self_attn.W_pack.weight"]   
            qkv_words = "|".join(lay_qkv_words)  
            
            for layername, weight in params_dict.items():
                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)

        if self.quant_method == "awq":
            lay_key_words = [
                "self_attn.W_pack.qweight",
                "self_attn.o_proj.qweight",
                "mlp.gate_up_proj.qweight",
                "mlp.down_proj.qweight"
            ]
            combined_words = "|".join(lay_key_words)
            
            for layername, weight in params_dict.items():
                
                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()

class BaichuanForCausalLM(BaiChuanBaseForCausalLM):
    """Baichuan 13B and Baichuan2 7B/13B.
    NOTE: the class name has a lower case 'c'.
    """

    def __init__(
        self,
        config: PretrainedConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        lora_config: Optional[LoRAConfig] = None,
    ):
        if config.hidden_size == 4096:  # baichuan2 7b
            super().__init__(config, "ROPE", cache_config, quant_config,
                             lora_config)
        else:  # baichuan 13b, baichuan2 13b
            super().__init__(config, "ALIBI", cache_config, quant_config,
                             lora_config)


class BaiChuanForCausalLM(BaiChuanBaseForCausalLM):
    """Baichuan 7B.
    NOTE: the class name has an upper case 'C'.
    """

    def __init__(
        self,
        config: PretrainedConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        lora_config: Optional[LoRAConfig] = None,
    ):
        super().__init__(config, "ROPE", cache_config, quant_config,
                         lora_config)