{"prompt": "Task: Find out what are the key topics in the document? output \"topic 1\", \"topic 2\", ... , \"topic n\".\n\nThe United States has withdrawn from the Paris Climate Agreement.\n\n", "completion": "Topic 1<|endoftext|>"}
{"prompt": "Given a list of numbers, find out if they are prime or not. you can answer \"prime\" or \"not prime\".\n\nInput: List: 1, 4, 6, 8, 9\n\nOutput:", "completion": " not prime<|endoftext|>"}
{"prompt": "Task: What are some of your favorite websites, and why do you visit them often?\n\n", "completion": "- Stack Overflow - I like to learn new things, and this is a great place to find answers to questions.<|endoftext|>"}
{"prompt": "Should i take this class or not?\nThe professor of this class is not good at all. He doesn't teach well and he is always late for class.\n\n", "completion": "No<|endoftext|>"}
{"prompt": "Given an array of integers, find two integers that sum up to a specific target integer.\n\n[1, 2, 3, 4]\nTarget: 7\n\n", "completion": "5 and 2<|endoftext|>"}
{"prompt": "Find out how many times does the word appear in the document. output the number of times.\nThe quick brown fox jumps over the lazy dog.\n\n", "completion": "3<|endoftext|>"}
{"prompt": "Rank these users according to their social influence.\nUser: The new movie is great. I would recommend it to anyone who likes action movies.\n\n", "completion": "2<|endoftext|>"}

    val df_test = spark.read.parquet("src/main/resource/data/parquet/rm-static-test.parquet")
    val df_train = spark.read.parquet("src/main/resource/data/parquet/rm-static-train.parquet")
    df_test.show()
    df_test.coalesce(1).write
      .option("header", "true")
      .csv("test.csv")
    df_train.coalesce(1).write
      .option("header", "true")
      .csv("train.csv")

[{
        "instruction": "请讲解如何缓解上班族病的症状。",
        "input": "",
        "output": "一种有效的缓解上班族病的症状方法是出去散步，每天晚上可以花几个小时去散步，减少坐姿固定的时间，放松肩痛、腰痛和背部综合症的发作。另外，可以试着利用午休时间或者其他空余时间锻炼一下，比如慢跑，打太极拳等，帮助舒缓，运动释放时也可以练习深呼吸，这能帮助消除压力，更有利于解除病症。",
        "id": 0
        "instruction": "描述一下优质的客户关系管理模式。",
        "input": "",
        "output": "优质的客户关系管理模式是指，通过建立客户之间的关系，建立并提供优质的服务，将客户的满意度转变为重复购买概念。这种模式强调人性化服务，以及知识和技能的结合，建立关系是一种长期的过程，而且可以建立起客户的忠诚度和口碑好评。该模式还可以培养客户之间的信任关系，增强客户感受到优质服务的同时，建立起长期客户风险防范机制，以及客户满意度机制，使企业拥有稳定、可持续和良好的客户关系管理能力。",
        "id": 1
        "instruction": "如果你要在你的城市建造一座新图书馆，你会考虑哪些因素，例如地点、设计和设施？",
        "input": "",
        "output": "如果要在城市建立一座新图书馆，首先需要考虑其地理位置，以及图书馆所在社区的需求状况，这样才能保证图书馆的功能能够有效地发挥出来。此外，需要考虑的还有当地特色，如地形、气候、文化等，以及周边的交通状况，可以考虑把新图书馆放在通勤方便的位置。最后是设计和设施的考量，应该根据预期使用用户的需求，把简洁的外观设计和时尚的设施结合在一起，以及图书馆有利的环境，来提高职能，建造现代化的图书馆。",
        "id": 2

def convert_to_lora_recursively(module: nn.Module, lora_rank: int) -> None:
    for name, child in module.named_children():
        if isinstance(child, nn.Linear):
            setattr(module, name, lora_linear_wrapper(child, lora_rank))
        else:
            convert_to_lora_recursively(child, lora_rank)

torchrun --standalone --nproc_per_node=4 train_sft.py \
    --pretrain "/path/to/LLaMa-7B/" \
    --model 'llama' \
    --strategy colossalai_zero2 \
    --log_interval 10 \
    --save_path  /path/to/Coati-7B \
    --dataset "yizhongw/self_instruct" \
    --batch_size 4 \
    --accimulation_steps 8 \
    --lr 2e-5 \
    --max_datasets_size 512 \
    --max_epochs 1 \

class LogSigLoss(nn.Module):
    Pairwise Loss for Reward Model
    Details: https://arxiv.org/abs/2203.02155
    def forward(self, chosen_reward: torch.Tensor, reject_reward: torch.Tensor) -> torch.Tensor:
        probs = torch.sigmoid(chosen_reward - reject_reward)
        log_probs = torch.log(probs)
        loss = -log_probs.mean()
        return loss

class RewardModel(LoRAModule):
    def forward(self, sequences: torch.LongTensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
        outputs = self.model(sequences, attention_mask=attention_mask)
        last_hidden_states = outputs['last_hidden_state']
        values = self.value_head(last_hidden_states)[:, :-1]
        value = values.mean(dim=1).squeeze(1)    # ensure shape is (B)
        return value

torchrun --standalone --nproc_per_node=4 train_reward_model.py
    --pretrain "/path/to/Coati-7B" \
    --model 'llama' \
    --dataset 'Dahoas/rm-static'
    --strategy colossalai_zero2 \
    --loss_fn 'log_exp'\
    --save_path '/path/to/rmstatic.pt' \

class Actor(LoRAModule):
    def forward(self,
                sequences: torch.LongTensor,
                num_actions: int,
                attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
        """Returns action log probs
        output = self.model(sequences, attention_mask=attention_mask)
        logits = output['logits']
        log_probs = log_probs_from_logits(logits[:, :-1, :], sequences[:, 1:])
        return log_probs[:, -num_actions:]
class Critic(LoRAModule):
    def forward(self,
                sequences: torch.LongTensor,
                action_mask: Optional[torch.Tensor] = None,
                attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
        outputs = self.model(sequences, attention_mask=attention_mask)
        last_hidden_states = outputs['last_hidden_state']
        values = self.value_head(last_hidden_states).squeeze(-1)
        if action_mask is not None and self.use_action_mask:
            num_actions = action_mask.size(1)
            prompt_mask = attention_mask[:, :-num_actions]
            values = values[:, :-num_actions]
            value = masked_mean(values, prompt_mask, dim=1)
            return value
        values = values[:, :-1]
        value = values.mean(dim=1)
        return value

torchrun --standalone --nproc_per_node=4 train_prompts.py \
         --pretrain "/path/to/Coati-7B" \
         --model 'llama' \
         --strategy colossalai_zero2 \
         --prompt_path /path/to/your/instinwild_ch.json \
         --pretrain_dataset /path/to/gpt3_finetuning_data.jsonl \
         --rm_pretrain "/path/to/Coati-7B" \
         --rm_path "/path/to/rmstatic.pt"

from transformers import LlamaForCausalLM
USE_8BIT = True # use 8-bit quantization; otherwise, use fp16
model = LlamaForCausalLM.from_pretrained(
            "pretrained/path",
            load_in_8bit=USE_8BIT,
            torch_dtype=torch.float16,
            device_map="auto",
if not USE_8BIT:
    model.half()  # use fp16
model.eval()
python server.py /path/to/pretrained — quant 4bit — gptq_checkpoint /path/to/coati-7b-4bit-128g.pt — gptq_group_size 128

CUDA_VISIBLE_DEVICES=0 python llama.py /path/to/pretrained/llama-7b c4 --wbits 4 --groupsize 128 --save llama7b-4bit.pt

from typing import Optional, Tuple, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
from ..generation import generate
from ..lora import LoRAModule
from ..utils import log_probs_from_logits
class Actor(LoRAModule):
    Actor model base class.
    Args:
        model (nn.Module): Actor Model.
        lora_rank (int): LoRA rank.
        lora_train_bias (str): LoRA bias training mode.
    def __init__(self, model: nn.Module, lora_rank: int = 0, lora_train_bias: str = 'none') -> None:
        super().__init__(lora_rank=lora_rank, lora_train_bias=lora_train_bias)
        self.model = model
        self.convert_to_lora()
    @torch.no_grad()
    def generate(
        self,
        input_ids: torch.Tensor,
        return_action_mask: bool = True,
        **kwargs
    ) -> Union[Tuple[torch.LongTensor, torch.LongTensor], Tuple[torch.LongTensor, torch.LongTensor, torch.BoolTensor]]:
        sequences = generate(self.model, input_ids, **kwargs)
        attention_mask = None
        pad_token_id = kwargs.get('pad_token_id', None)
        if pad_token_id is not None:
            attention_mask = sequences.not_equal(pad_token_id).to(dtype=torch.long, device=sequences.device)
        if not return_action_mask:
            return sequences, attention_mask, None
        input_len = input_ids.size(1)
        eos_token_id = kwargs.get('eos_token_id', None)
        if eos_token_id is None:
            action_mask = torch.ones_like(sequences, dtype=torch.bool)
        else:
            # left padding may be applied, only mask action
            action_mask = (sequences[:, input_len:] == eos_token_id).cumsum(dim=-1) == 0
            action_mask = F.pad(action_mask, (1 + input_len, -1), value=True)    # include eos token and input
        action_mask[:, :input_len] = False
        action_mask = action_mask[:, 1:]
        return sequences, attention_mask, action_mask[:, -(sequences.size(1) - input_len):]
    def forward(self,
                sequences: torch.LongTensor,
                num_actions: int,
                attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
        """Returns action log probs
        output = self.model(sequences, attention_mask=attention_mask)
        logits = output['logits']
        log_probs = log_probs_from_logits(logits[:, :-1, :], sequences[:, 1:])
        return log_probs[:, -num_actions:]
    def get_base_model(self):
        return self.model
class LlamaActor(Actor):
    Llama Actor model.
    Args:
        pretrained (str): Pretrained model name or path.
        config (LlamaConfig): Model config.
        checkpoint (bool): Enable gradient checkpointing.
        lora_rank (int): LoRA rank.
        lora_train_bias (str): LoRA bias training mode.
    def __init__(self,
                 pretrained: Optional[str] = None,
                 config: Optional[LlamaConfig] = None,
                 checkpoint: bool = False,
                 lora_rank: int = 0,
                 lora_train_bias: str = 'none') -> None:
        if pretrained is not None:
            model = LlamaForCausalLM.from_pretrained(pretrained)
        elif config is not None:
            model = LlamaForCausalLM(config)
        else:
            model = LlamaForCausalLM(LlamaConfig())
        if checkpoint:
            model.gradient_checkpointing_enable()
        super().__init__(model, lora_rank, lora_train_bias)

from typing import Optional
import torch
import torch.nn as nn
from ..lora import LoRAModule
from ..utils import masked_mean
class Critic(LoRAModule):
    Critic model base class.
    Args:
        model (nn.Module): Critic model.
        value_head (nn.Module): Value head to get value.
        lora_rank (int): LoRA rank.
        lora_train_bias (str): LoRA bias training mode.
    def __init__(
        self,
        model: nn.Module,
        value_head: nn.Module,
        lora_rank: int = 0,
        lora_train_bias: str = 'none',
        use_action_mask: bool = False,
    ) -> None:
        super().__init__(lora_rank=lora_rank, lora_train_bias=lora_train_bias)
        self.model = model
        self.value_head = value_head
        self.use_action_mask = use_action_mask
        self.convert_to_lora()
    def forward(self,
                sequences: torch.LongTensor,
                action_mask: Optional[torch.Tensor] = None,
                attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
        outputs = self.model(sequences, attention_mask=attention_mask)
        last_hidden_states = outputs['last_hidden_state']
        values = self.value_head(last_hidden_states).squeeze(-1)
        if action_mask is not None and self.use_action_mask:
            num_actions = action_mask.size(1)
            prompt_mask = attention_mask[:, :-num_actions]
            values = values[:, :-num_actions]
            value = masked_mean(values, prompt_mask, dim=1)
            return value
        values = values[:, :-1]
        value = values.mean(dim=1)
        return value
class LlamaCritic(Critic):
    Llama Critic model.
    Args:
        pretrained (str): Pretrained model name or path.
        config (LlamaConfig): Model config.
        checkpoint (bool): Enable gradient checkpointing.
        lora_rank (int): LoRA rank.
        lora_train_bias (str): LoRA bias training mode.
    def __init__(self,
                 pretrained: Optional[str] = None,
                 config: Optional[LlamaConfig] = None,
                 checkpoint: bool = False,
                 lora_rank: int = 0,
                 lora_train_bias: str = 'none',
                 **kwargs) -> None:
        if pretrained is not None:
            model = LlamaForCausalLM.from_pretrained(pretrained)
        elif config is not None:
            model = LlamaForCausalLM(config)
        else:
            model = LlamaForCausalLM(LlamaConfig())
        if checkpoint:
            model.gradient_checkpointing_enable()
        value_head = nn.Linear(model.config.hidden_size, 1)
        super().__init__(model, value_head, lora_rank, lora_train_bias, **kwargs)

from typing import Optional
import torch
import torch.nn as nn
from ..lora import LoRAModule
class RewardModel(LoRAModule):
    Reward model base class.
    Args:
        model (nn.Module): Reward model.
        value_head (nn.Module): Value head to get reward score.
        lora_rank (int): LoRA rank.
        lora_train_bias (str): LoRA bias training mode.
    def __init__(self,
                 model: nn.Module,
                 value_head: Optional[nn.Module] = None,
                 lora_rank: int = 0,
                 lora_train_bias: str = 'none') -> None:
        super().__init__(lora_rank=lora_rank, lora_train_bias=lora_train_bias)
        self.model = model
        self.convert_to_lora()
        if value_head is not None:
            if value_head.out_features != 1:
                raise ValueError("The value head of reward model's output dim should be 1!")
            self.value_head = value_head
        else:
            self.value_head = nn.Linear(model.config.n_embd, 1)
    def forward(self, sequences: torch.LongTensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
        outputs = self.model(sequences, attention_mask=attention_mask)
        last_hidden_states = outputs['last_hidden_state']
        values = self.value_head(last_hidden_states)[:, :-1]
        value = values.mean(dim=1).squeeze(1)    # ensure shape is (B)
        return value
class LlamaRM(RewardModel):
    Llama Reward model.
    Args:
        pretrained (str): Pretrained model name or path.
        config (LlamaConfig): Model config.
        checkpoint (bool): Enable gradient checkpointing.
        lora_rank (int): LoRA rank.
        lora_train_bias (str): LoRA bias training mode.
    def __init__(self,
                 pretrained: Optional[str] = None,
                 config: Optional[LlamaConfig] = None,
                 checkpoint: bool = False,
                 lora_rank: int = 0,
                 lora_train_bias: str = 'none') -> None:
        if pretrained is not None:
            model = LlamaModel.from_pretrained(pretrained)
        elif config is not None:
            model = LlamaModel(config)
        else:
            model = LlamaModel(LlamaConfig())
        if checkpoint:
            model.gradient_checkpointing_enable()
        value_head = nn.Linear(model.config.hidden_size, 1)
        value_head.weight.data.normal_(mean=0.0, std=1 / (model.config.hidden_size + 1))
        super().__init__(model, value_head, lora_rank, lora_train_bias)
# This software may be used and distributed according to the terms of the GNU General Public License version 3.
from typing import Optional, Tuple
from dataclasses import dataclass
import math
import torch
from torch import nn
import torch.nn.functional as F
import fairscale.nn.model_parallel.initialize as fs_init
from fairscale.nn.model_parallel.layers import (
    ParallelEmbedding,
    RowParallelLinear,
    ColumnParallelLinear,
@dataclass
class ModelArgs:
    dim: int = 512
    n_layers: int = 8
    n_heads: int = 8
    vocab_size: int = -1  # defined later by tokenizer
    multiple_of: int = 256  # make SwiGLU hidden layer size multiple of large power of 2
    norm_eps: float = 1e-5
    max_batch_size: int = 32
    max_seq_len: int = 2048
class RMSNorm(torch.nn.Module):
    def __init__(self, dim: int, eps: float = 1e-6):
        super().__init__()
        self.eps = eps
        self.weight = nn.Parameter(torch.ones(dim))
    def _norm(self, x):
        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
    def forward(self, x):
        output = self._norm(x.float()).type_as(x)
        return output * self.weight
def precompute_freqs_cis(dim: int, end: int, theta: float = 10000.0):
    freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
    t = torch.arange(end, device=freqs.device)  # type: ignore
    freqs = torch.outer(t, freqs).float()  # type: ignore
    freqs_cis = torch.polar(torch.ones_like(freqs), freqs)  # complex64
    return freqs_cis
def reshape_for_broadcast(freqs_cis: torch.Tensor, x: torch.Tensor):
    ndim = x.ndim
    assert 0 <= 1 < ndim
    assert freqs_cis.shape == (x.shape[1], x.shape[-1])
    shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)]
    return freqs_cis.view(*shape)
def apply_rotary_emb(
    xq: torch.Tensor,
    xk: torch.Tensor,
    freqs_cis: torch.Tensor,
) -> Tuple[torch.Tensor, torch.Tensor]:
    xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))
    xk_ = torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2))
    freqs_cis = reshape_for_broadcast(freqs_cis, xq_)
    xq_out = torch.view_as_real(xq_ * freqs_cis).flatten(3)
    xk_out = torch.view_as_real(xk_ * freqs_cis).flatten(3)
    return xq_out.type_as(xq), xk_out.type_as(xk)
class Attention(nn.Module):
    def __init__(self, args: ModelArgs):
        super().__init__()
        self.n_local_heads = args.n_heads // fs_init.get_model_parallel_world_size()
        self.head_dim = args.dim // args.n_heads
        self.wq = ColumnParallelLinear(
            args.dim,
            args.n_heads * self.head_dim,
            bias=False,
            gather_output=False,
            init_method=lambda x: x,
        self.wk = ColumnParallelLinear(
            args.dim,
            args.n_heads * self.head_dim,
            bias=False,
            gather_output=False,
            init_method=lambda x: x,
        self.wv = ColumnParallelLinear(
            args.dim,
            args.n_heads * self.head_dim,
            bias=False,
            gather_output=False,
            init_method=lambda x: x,
        self.wo = RowParallelLinear(
            args.n_heads * self.head_dim,
            args.dim,
            bias=False,
            input_is_parallel=True,
            init_method=lambda x: x,
        self.cache_k = torch.zeros(
            (args.max_batch_size, args.max_seq_len, self.n_local_heads, self.head_dim)
        ).cuda()
        self.cache_v = torch.zeros(
            (args.max_batch_size, args.max_seq_len, self.n_local_heads, self.head_dim)
        ).cuda()
    def forward(self, x: torch.Tensor, start_pos: int, freqs_cis: torch.Tensor, mask: Optional[torch.Tensor]):
        bsz, seqlen, _ = x.shape
        xq, xk, xv = self.wq(x), self.wk(x), self.wv(x)
        xq = xq.view(bsz, seqlen, self.n_local_heads, self.head_dim)
        xk = xk.view(bsz, seqlen, self.n_local_heads, self.head_dim)
        xv = xv.view(bsz, seqlen, self.n_local_heads, self.head_dim)
        xq, xk = apply_rotary_emb(xq, xk, freqs_cis=freqs_cis)
        self.cache_k = self.cache_k.to(xq)
        self.cache_v = self.cache_v.to(xq)
        self.cache_k[:bsz, start_pos : start_pos + seqlen] = xk
        self.cache_v[:bsz, start_pos : start_pos + seqlen] = xv
        keys = self.cache_k[:bsz, : start_pos + seqlen]
        values = self.cache_v[:bsz, : start_pos + seqlen]
        xq = xq.transpose(1, 2)
        keys = keys.transpose(1, 2)
        values = values.transpose(1, 2)
        #计算未归一化的注意力分数
        scores = torch.matmul(xq, keys.transpose(2, 3)) / math.sqrt(self.head_dim)
        if mask is not None:
            scores = scores + mask  # (bs, n_local_heads, slen, cache_len + slen)
        #计算归一化的注意力分数
        scores = F.softmax(scores.float(), dim=-1).type_as(xq)
        #将V使用注意力分数加权获得输出表示
        output = torch.matmul(scores, values)  # (bs, n_local_heads, slen, head_dim)
        output = output.transpose(
            1, 2
        ).contiguous().view(bsz, seqlen, -1)
        return self.wo(output)
class FeedForward(nn.Module):
    def __init__(
        self,
        dim: int,
        hidden_dim: int,
        multiple_of: int,
        super().__init__()
        hidden_dim = int(2 * hidden_dim / 3)
        hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
        self.w1 = ColumnParallelLinear(
            dim, hidden_dim, bias=False, gather_output=False, init_method=lambda x: x
        self.w2 = RowParallelLinear(
            hidden_dim, dim, bias=False, input_is_parallel=True, init_method=lambda x: x
        self.w3 = ColumnParallelLinear(
            dim, hidden_dim, bias=False, gather_output=False, init_method=lambda x: x
    def forward(self, x):
        return self.w2(F.silu(self.w1(x)) * self.w3(x))
class TransformerBlock(nn.Module):
    def __init__(self, layer_id: int, args: ModelArgs):
        super().__init__()
        self.n_heads = args.n_heads
        self.dim = args.dim
        self.head_dim = args.dim // args.n_heads
        self.attention = Attention(args)
        self.feed_forward = FeedForward(
            dim=args.dim, hidden_dim=4 * args.dim, multiple_of=args.multiple_of
        self.layer_id = layer_id
        self.attention_norm = RMSNorm(args.dim, eps=args.norm_eps)
        self.ffn_norm = RMSNorm(args.dim, eps=args.norm_eps)
    def forward(self, x: torch.Tensor, start_pos: int, freqs_cis: torch.Tensor, mask: Optional[torch.Tensor]):
        h = x + self.attention.forward(self.attention_norm(x), start_pos, freqs_cis, mask)
        out = h + self.feed_forward.forward(self.ffn_norm(h))
        return out
class Transformer(nn.Module):
    def __init__(self, params: ModelArgs):
        super().__init__()
        self.params = params
        self.vocab_size = params.vocab_size
        self.n_layers = params.n_layers
        self.tok_embeddings = ParallelEmbedding(
            params.vocab_size, params.dim, init_method=lambda x: x
        self.layers = torch.nn.ModuleList()
        for layer_id in range(params.n_layers):
            self.layers.append(TransformerBlock(layer_id, params))
        self.norm = RMSNorm(params.dim, eps=params.norm_eps)
        self.output = ColumnParallelLinear(
            params.dim, params.vocab_size, bias=False, init_method=lambda x: x
        self.freqs_cis = precompute_freqs_cis(
            self.params.dim // self.params.n_heads, self.params.max_seq_len * 2
    @torch.inference_mode()
    def forward(self, tokens: torch.Tensor, start_pos: int):
        _bsz, seqlen = tokens.shape
        h = self.tok_embeddings(tokens)
        self.freqs_cis = self.freqs_cis.to(h.device)
        freqs_cis = self.freqs_cis[start_pos : start_pos + seqlen]
        mask = None
        if seqlen > 1:
            mask = torch.full((1, 1, seqlen, seqlen), float("-inf"), device=tokens.device)
            mask = torch.triu(mask, diagonal=start_pos + 1).type_as(h)
        for layer in self.layers:
            h = layer(h, start_pos, freqs_cis, mask)
        h = self.norm(h)
        output = self.output(h[:, -1, :])  # only compute last logits
        return output.float()

from typing import Optional, Tuple, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
from ..generation import generate
from .actor import Actor
class LM(Actor):
    Language model base class.
    Args:
        model (nn.Module): Language Model.
        lora_rank (int): LoRA rank.
        lora_train_bias (str): LoRA bias training mode.
    def __init__(self, model: nn.Module, lora_rank: int = 0, lora_train_bias: str = 'none') -> None:
        super().__init__(model=model, lora_rank=lora_rank, lora_train_bias=lora_train_bias)
    def forward(self, sequences: torch.LongTensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
        """Returns output log probs
        output = self.model(sequences, attention_mask=attention_mask)
        logits = output['logits']
        log_probs = F.log_softmax(logits, dim=-1)
        return log_probs

from typing import Any, Callable, Dict, List, Optional
import torch
import torch.nn as nn
from coati.experience_maker import Experience, NaiveExperienceMaker
from coati.models.base import Actor, Critic
from coati.models.generation_utils import update_model_kwargs_fn
from coati.models.loss import PolicyLoss, ValueLoss
from coati.replay_buffer import NaiveReplayBuffer
from torch.optim import Optimizer
from transformers.tokenization_utils_base import PreTrainedTokenizerBase
from .base import Trainer
from .callbacks import Callback
from .strategies import Strategy
class PPOTrainer(Trainer):
        Trainer for PPO algorithm.
    Args:
        strategy (Strategy): the strategy to use for training
        actor (Actor): the actor model in ppo algorithm
        critic (Critic): the critic model in ppo algorithm
        reward_model (nn.Module): the reward model in rlhf algorithm to make reward of sentences
        initial_model (Actor): the initial model in rlhf algorithm to generate reference logits to limit the update of actor
        actor_optim (Optimizer): the optimizer to use for actor model
        critic_optim (Optimizer): the optimizer to use for critic model
        kl_coef (float, defaults to 0.1): the coefficient of kl divergence loss
        train_batch_size (int, defaults to 8): the batch size to use for training
        buffer_limit (int, defaults to 0): the max_size limitaiton of replay buffer
        buffer_cpu_offload (bool, defaults to True): whether to offload replay buffer to cpu
        eps_clip (float, defaults to 0.2): the clip coefficient of policy loss
        value_clip (float, defaults to 0.4): the clip coefficient of value loss
        experience_batch_size (int, defaults to 8): the batch size to use for experience generation
        max_epochs (int, defaults to 1): the number of epochs of training process
        tokenier (Callable, optional): the tokenizer to use for tokenizing the input
        sample_replay_buffer (bool, defaults to False): whether to sample from replay buffer
        dataloader_pin_memory (bool, defaults to True): whether to pin memory for data loader
        callbacks (List[Callback], defaults to []): the callbacks to call during training process
        generate_kwargs (dict, optional): the kwargs to use while model generating
    def __init__(self,
                 strategy: Strategy,
                 actor: Actor,
                 critic: Critic,
                 reward_model: nn.Module,
                 initial_model: Actor,
                 actor_optim: Optimizer,
                 critic_optim: Optimizer,
                 kl_coef: float = 0.1,
                 ptx_coef: float = 0.9,
                 train_batch_size: int = 8,
                 buffer_limit: int = 0,
                 buffer_cpu_offload: bool = True,
                 eps_clip: float = 0.2,
                 value_clip: float = 0.4,
                 experience_batch_size: int = 8,
                 max_epochs: int = 1,
                 tokenizer: Optional[Callable[[Any], dict]] = None,
                 sample_replay_buffer: bool = False,
                 dataloader_pin_memory: bool = True,
                 callbacks: List[Callback] = [],
                 **generate_kwargs) -> None:
        experience_maker = NaiveExperienceMaker(actor, critic, reward_model, initial_model, kl_coef)
        replay_buffer = NaiveReplayBuffer(train_batch_size, buffer_limit, buffer_cpu_offload)
        generate_kwargs = _set_default_generate_kwargs(strategy, generate_kwargs, actor)
        super().__init__(strategy, experience_maker, replay_buffer, experience_batch_size, max_epochs, tokenizer,
                         sample_replay_buffer, dataloader_pin_memory, callbacks, **generate_kwargs)
        self.actor = actor
        self.critic = critic
        self.actor_loss_fn = PolicyLoss(eps_clip)
        self.critic_loss_fn = ValueLoss(value_clip)
        self.ptx_loss_fn = nn.CrossEntropyLoss(ignore_index=-100)
        self.ptx_coef = ptx_coef
        self.actor_optim = actor_optim
        self.critic_optim = critic_optim
    def training_step(self, experience: Experience) -> Dict[str, float]:
        self.actor.train()
        self.critic.train()
        # policy loss
        num_actions = experience.action_mask.size(1)
        action_log_probs = self.actor(experience.sequences, num_actions, attention_mask=experience.attention_mask)
        actor_loss = self.actor_loss_fn(action_log_probs,
                                        experience.action_log_probs,
                                        experience.advantages,
                                        action_mask=experience.action_mask)
        # ptx loss
        if self.ptx_coef != 0:
            ptx = next(iter(self.pretrain_dataloader))['input_ids'].to(torch.cuda.current_device())
            label = next(iter(self.pretrain_dataloader))['labels'].to(torch.cuda.current_device())[:, 1:]
            attention_mask = next(iter(self.pretrain_dataloader))['attention_mask'].to(torch.cuda.current_device())
            ptx_log_probs = self.actor.get_base_model()(ptx, attention_mask=attention_mask)['logits'][..., :-1, :]
            ptx_loss = self.ptx_loss_fn(ptx_log_probs.view(-1, ptx_log_probs.size(-1)), label.view(-1))
            actor_loss = ptx_loss * self.ptx_coef + actor_loss * (1 - self.ptx_coef)
        self.strategy.backward(actor_loss, self.actor, self.actor_optim)
        self.strategy.optimizer_step(self.actor_optim)
        self.actor_optim.zero_grad()
        # value loss
        values = self.critic(experience.sequences,
                             action_mask=experience.action_mask,
                             attention_mask=experience.attention_mask)
        critic_loss = self.critic_loss_fn(values,
                                          experience.values,
                                          experience.reward,
                                          action_mask=experience.action_mask)
        self.strategy.backward(critic_loss, self.critic, self.critic_optim)
        self.strategy.optimizer_step(self.critic_optim)
        self.critic_optim.zero_grad()
        return {'reward': experience.reward.mean().item()}
def _set_default_generate_kwargs(strategy: Strategy, generate_kwargs: dict, actor: Actor) -> None:
    origin_model = strategy._unwrap_actor(actor)
    new_kwargs = {**generate_kwargs}
    # use huggingface models method directly
    if 'prepare_inputs_fn' not in generate_kwargs and hasattr(origin_model, 'prepare_inputs_for_generation'):
        new_kwargs['prepare_inputs_fn'] = origin_model.prepare_inputs_for_generation
    if 'update_model_kwargs_fn' not in generate_kwargs:
        new_kwargs['update_model_kwargs_fn'] = update_model_kwargs_fn
    return new_kwargs
def save_model(self, path: str, only_rank0: bool = False, tokenizer: Optional[PreTrainedTokenizerBase] = None) -> None:
    self.strategy.save_model(model=self.actor, path=path, only_rank0=only_rank0, tokenizer=tokenizer)