Skip to content

🧠 两小时从 0 训出你的 64M 大模型。Train your own 64M LLM from scratch in 2 hours.

10.2 KB
Raw Blame
"""
# 如果使用sglang加速,需通过以下命令首先启动(transformers格式)模型:
python -m sglang.launch_server --model-path ./minimind-3 --attention-backend triton --host 0.0.0.0 --port 8998
"""
import os
import sys

__package__ = "trainer"
sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))

import requests
import torch
import torch.distributed as dist
from abc import ABC, abstractmethod
from contextlib import nullcontext
from dataclasses import dataclass
from typing import List, Optional, Tuple
from torch import Tensor
from torch.nn.parallel import DistributedDataParallel
from transformers import AutoTokenizer


# ===== 计算每个 token 的 logprob =====
def compute_per_token_logps(model, input_ids: Tensor, n_keep: int, attention_mask: Optional[Tensor] = None) -> Tensor:
    if n_keep <= 0:
        return input_ids.new_empty((input_ids.size(0), 0), dtype=torch.float32)
    unwrapped = model.module if isinstance(model, DistributedDataParallel) else model
    input_ids = input_ids.detach().clone() if input_ids.is_inference() else input_ids
    logits = unwrapped(input_ids, attention_mask=attention_mask, logits_to_keep=n_keep + 1).logits[:, :-1, :]
    per_token_logps = []
    for logits_row, ids_row in zip(logits, input_ids[:, -n_keep:]):
        ids_row = ids_row.detach().clone() if ids_row.is_inference() else ids_row
        per_token_logps.append(
            torch.gather(logits_row.log_softmax(dim=-1), 1, ids_row.unsqueeze(1)).squeeze(1)
        )
    return torch.stack(per_token_logps)


# ===== Rollout 结果 =====
@dataclass
class RolloutResult:
    output_ids: Tensor
    completion_ids: Tensor
    per_token_logps: Tensor
    completions: List[str]
    prompt_lens: Tensor
    completion_mask: Tensor


# ===== Rollout 引擎抽象基类 =====
class RolloutEngine(ABC):
    tokenizer = None
    
    @abstractmethod
    def rollout(self, prompt_ids: Tensor, attention_mask: Tensor, num_generations: int, max_new_tokens: int, temperature: float = 0.8) -> RolloutResult:
        pass
    
    @abstractmethod
    def update_policy(self, model: torch.nn.Module):
        pass


# ===== PyTorch 原生推理引擎 =====
class TorchRolloutEngine(RolloutEngine):
    def __init__(self, policy_model: torch.nn.Module, tokenizer, device: str = "cuda", autocast_ctx=None):
        self.policy_model = policy_model
        self.tokenizer = tokenizer
        self.device = device
        self.autocast_ctx = autocast_ctx
    
    def rollout(self, prompt_ids: Tensor, attention_mask: Tensor, num_generations: int, max_new_tokens: int, temperature: float = 0.8) -> RolloutResult:
        model = self.policy_model.module if isinstance(self.policy_model, DistributedDataParallel) else self.policy_model
        ctx = self.autocast_ctx if self.autocast_ctx else nullcontext()
        with torch.no_grad(), ctx:
            output_ids = model.generate(
                input_ids=prompt_ids.repeat_interleave(num_generations, dim=0),
                attention_mask=attention_mask.repeat_interleave(num_generations, dim=0),
                max_new_tokens=max_new_tokens,
                do_sample=True,
                temperature=temperature,
                num_return_sequences=1,
                pad_token_id=self.tokenizer.pad_token_id,
                eos_token_id=self.tokenizer.eos_token_id,
            ).clone()  # [B*num_gen, P+R]
            prompt_len = prompt_ids.size(1)
            completion_ids = output_ids[:, prompt_len:]  # [B*num_gen, R]
            full_mask = torch.cat([attention_mask.repeat_interleave(num_generations, dim=0), attention_mask.new_ones(output_ids.size(0), completion_ids.size(1))], dim=1)
            per_token_logps = compute_per_token_logps(self.policy_model, output_ids, completion_ids.size(1), attention_mask=full_mask)
        completions = self.tokenizer.batch_decode(completion_ids, skip_special_tokens=True)
        return RolloutResult(output_ids, completion_ids, per_token_logps, completions,
                             prompt_ids.new_full((output_ids.size(0),), prompt_len),
                             attention_mask.new_ones(output_ids.size(0), completion_ids.size(1)))
    
    def update_policy(self, model: torch.nn.Module):
        self.policy_model = model


# ===== SGLang HTTP API 推理引擎 =====
class SGLangRolloutEngine(RolloutEngine):
    def __init__(self, base_url: str, model_path: str, shared_ckpt_path: str = "./sglang_ckpt", timeout: int = 120):
        self.base_url = base_url.rstrip('/')
        self.shared_ckpt_path = shared_ckpt_path
        self.timeout = timeout
        self.tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
        self.http = requests
    
    def rollout(self, prompt_ids: Tensor, attention_mask: Tensor, num_generations: int, max_new_tokens: int, temperature: float = 0.8) -> RolloutResult:
        # 去除左侧 padding tokens,只保留有效 token
        input_ids_list = []
        for ids, mask in zip(prompt_ids, attention_mask):
            valid_ids = ids[mask.bool()].tolist()
            input_ids_list.append(valid_ids)
        all_input_ids = [ids for ids in input_ids_list for _ in range(num_generations)]
        
        payload = {
            "input_ids": all_input_ids,
            "sampling_params": {
                "temperature": temperature,
                "max_new_tokens": max_new_tokens,
                "stop_token_ids": [self.tokenizer.eos_token_id] if self.tokenizer.eos_token_id else [],
            },
            "return_logprob": True,
        }
        
        resp = self.http.post(f"{self.base_url}/generate", json=payload, timeout=self.timeout)
        resp.raise_for_status()
        
        results = resp.json()
        if not isinstance(results, list):
            results = [results]
        
        all_output_ids, all_completion_ids, all_logprobs = [], [], []
        completions = []
        
        for i, result in enumerate(results):
            meta = result.get("meta_info", {})
            completion_ids = meta.get("output_ids", result.get("output_ids", []))
            raw_logprobs = meta.get("output_token_logprobs", [])
            
            logprobs = []
            for item in raw_logprobs:
                if isinstance(item, (list, tuple)) and len(item) >= 1:
                    logprobs.append(item[0])
                elif isinstance(item, (int, float)):
                    logprobs.append(item)
            
            if len(logprobs) < len(completion_ids):
                logprobs = [0.0] * (len(completion_ids) - len(logprobs)) + logprobs
            elif len(logprobs) > len(completion_ids):
                logprobs = logprobs[-len(completion_ids):] if completion_ids else []
            prompt = all_input_ids[i]
            full_output = prompt + completion_ids
            all_output_ids.append(full_output)
            all_completion_ids.append(completion_ids)
            all_logprobs.append(logprobs)
            completions.append(self.tokenizer.decode(completion_ids, skip_special_tokens=True))
        
        device = prompt_ids.device
        max_comp_len = max(1, max(len(ids) for ids in all_completion_ids))
        max_out_len = max(len(ids) for ids in all_input_ids) + max_comp_len
        
        def pad_to_tensor(seqs, max_len, pad_val=0):
            return torch.tensor([s + [pad_val] * (max_len - len(s)) for s in seqs], device=device)
        
        pad_id = self.tokenizer.pad_token_id
        return RolloutResult(
            output_ids=pad_to_tensor(all_output_ids, max_out_len, pad_val=pad_id),
            completion_ids=pad_to_tensor(all_completion_ids, max_comp_len, pad_val=pad_id),
            per_token_logps=pad_to_tensor(all_logprobs, max_comp_len, pad_val=0.0),
            completions=completions,
            prompt_lens=torch.tensor([len(ids) for ids in all_input_ids], device=device),
            completion_mask=torch.tensor([[1] * len(ids) + [0] * (max_comp_len - len(ids)) for ids in all_completion_ids], device=device),
        )
    
    def update_policy(self, model: torch.nn.Module):
        ok = True
        if not dist.is_initialized() or dist.get_rank() == 0:
            try:
                unwrapped = model.module if isinstance(model, DistributedDataParallel) else model
                unwrapped = getattr(unwrapped, '_orig_mod', unwrapped)
                abs_path = os.path.abspath(self.shared_ckpt_path)
                state_dict = {k: v.detach().half().cpu() for k, v in unwrapped.state_dict().items()}
                unwrapped.save_pretrained(abs_path, state_dict=state_dict, safe_serialization=False)
                self.tokenizer.save_pretrained(abs_path)
                resp = self.http.post(f"{self.base_url}/update_weights_from_disk", json={"model_path": abs_path}, timeout=self.timeout)
                if resp.status_code != 200: print(f"[SGLANG WARNING] update_weights 失败: {resp.status_code}, {resp.text}")
                ok = resp.status_code == 200
            except Exception as e:
                print(f"[SGLANG WARNING] update_weights 异常: {e}"); ok = False
        if dist.is_initialized():
            ok_t = torch.tensor(int(ok), device=next(model.parameters()).device)
            dist.broadcast(ok_t, src=0); dist.barrier(); ok = bool(ok_t.item())
        if not ok: raise RuntimeError("SGLang update_policy failed")
        return ok
    
    def flush_cache(self) -> bool:
        resp = self.http.post(f"{self.base_url}/flush_cache", timeout=30)
        return resp.status_code == 200
    
    def health(self) -> bool:
        try:
            resp = self.http.get(f"{self.base_url}/health", timeout=5)
            return resp.status_code == 200
        except:
            return False


# ===== 工厂函数 =====
def create_rollout_engine(
    engine_type: str = "torch",
    policy_model: torch.nn.Module = None,
    tokenizer = None,
    device: str = "cuda",
    autocast_ctx = None,
    sglang_base_url: str = None,
    sglang_model_path: str = None,
    sglang_shared_path: str = None,
) -> RolloutEngine:
    if engine_type == "torch":
        return TorchRolloutEngine(policy_model, tokenizer, device, autocast_ctx)
    elif engine_type == "sglang":
        return SGLangRolloutEngine(sglang_base_url, sglang_model_path, sglang_shared_path)
    else:
        raise ValueError(f"不支持的引擎类型: {engine_type}")