Merge 1a18e651f6 into 7a9137d2e9

[perf] skip the per-forward RoPE buffer broadcast in the RL trainers

scripts/eval_toolcall.py

@@ -12,7 +12,7 @@ from datetime import datetime
 from transformers import AutoTokenizer, AutoModelForCausalLM, TextStreamer
 from openai import OpenAI
 from model.model_minimind import MiniMindConfig, MiniMindForCausalLM
-from trainer.trainer_utils import setup_seed, get_model_params
+from trainer.trainer_utils import setup_seed, get_model_params, safe_math_eval
 warnings.filterwarnings('ignore')
 
 TOOLS = [
@@ -27,7 +27,7 @@ TOOLS = [
 ]
 
 MOCK_RESULTS = {
-    "calculate_math": lambda args: {"result": str(eval(str(args.get("expression", "0")).replace("^", "**").replace("×", "*").replace("÷", "/").replace("−", "-").replace("²", "**2").replace("³", "**3").replace("（", "(").replace("）", ")")))},
+    "calculate_math": lambda args: {"result": str(safe_math_eval(args.get("expression", "0")))},
     "get_current_time": lambda args: {"datetime": datetime.now().strftime("%Y-%m-%d %H:%M:%S"), "timezone": args.get("timezone", "Asia/Shanghai")},
     "random_number": lambda args: {"result": random.randint(int(args.get("min", 0)), int(args.get("max", 100)))},
     "text_length": lambda args: {"characters": len(args.get("text", "")), "words": len(args.get("text", "").split())},

scripts/web_demo.py

@@ -2,6 +2,7 @@ import random
 import re
 import json
 import os
+import sys
 from threading import Thread
 
 import torch
@@ -9,6 +10,9 @@ import numpy as np
 import streamlit as st
 from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
 
+sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
+from trainer.trainer_utils import safe_math_eval
+
 st.set_page_config(page_title="MiniMind", initial_sidebar_state="collapsed")
 
 st.markdown("""
@@ -125,7 +129,7 @@ def execute_tool(tool_name, args):
     import datetime
     try:
         if tool_name == 'calculate_math':
-            return {"result": eval(args.get('expression', '0'))}
+            return {"result": safe_math_eval(args.get('expression', '0'))}
         elif tool_name == 'get_current_time':
             tz = args.get('timezone', 'Asia/Shanghai')
             return {"result": datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S')}

trainer/train_agent.py

@@ -10,7 +10,6 @@ import gc
 import json
 import math
 import random
-import signal
 import argparse
 import warnings
 import torch
@@ -24,7 +23,7 @@ from torch.optim.lr_scheduler import CosineAnnealingLR
 from transformers import AutoTokenizer
 from model.model_minimind import MiniMindConfig, MiniMindForCausalLM
 from dataset.lm_dataset import AgentRLDataset
-from trainer.trainer_utils import Logger, is_main_process, lm_checkpoint, init_distributed_mode, setup_seed, SkipBatchSampler, init_model, LMForRewardModel
+from trainer.trainer_utils import Logger, is_main_process, lm_checkpoint, init_distributed_mode, setup_seed, SkipBatchSampler, init_model, LMForRewardModel, safe_math_eval
 from trainer.rollout_engine import create_rollout_engine, compute_per_token_logps
 
 warnings.filterwarnings('ignore')
@@ -55,7 +54,7 @@ UNIT_DATA = {"km_miles": 0.621371, "miles_km": 1.60934, "kg_pounds": 2.20462, "p
 
 # ======== 模拟执行 ========
 MOCK_RESULTS = {
-    "calculate_math": lambda args: {"result": str(eval(str(args.get("expression", "0")).replace("^", "**").replace("×", "*").replace("÷", "/").replace("−", "-").replace("（", "(").replace("）", ")"), {"__builtins__": {}, "math": math}))},
+    "calculate_math": lambda args: {"result": str(safe_math_eval(args.get("expression", "0")))},
     "unit_converter": lambda args: {"result": round(float(args.get("value", 0)) * UNIT_DATA.get(f"{args.get('from_unit', '').lower()}_{args.get('to_unit', '').lower()}", 1), 4)},
     "get_current_weather": lambda args: (lambda w: {"city": args.get("location"), "temperature": w[0], "humidity": "65%", "condition": w[1]})(WEATHER_DATA.get(args.get("location"), ("22°C", "晴"))),
     "get_current_time": lambda args: {"datetime": TIME_DATA.get(args.get("timezone", "Asia/Shanghai"), "2025-03-07 14:30:00"), "timezone": args.get("timezone", "Asia/Shanghai")},
@@ -85,14 +84,9 @@ def execute_tool(name, args):
     fn = MOCK_RESULTS.get(name)
     if not fn: return None
     try:
-        signal.signal(signal.SIGALRM, lambda *_: (_ for _ in ()).throw(TimeoutError()))
-        signal.alarm(1)
         return fn(args)
-    except:
+    except Exception:
         return None
-    finally:
-        try: signal.alarm(0)
-        except: pass
 
 # ======== 多轮 Rollout ========
 def rollout_single(rollout_engine, tokenizer, messages, tools, max_turns=3, max_new_tokens=256, thinking_ratio=0.5, device="cuda"):
@@ -278,9 +272,10 @@ def rl_train_epoch(epoch, loader, iters, rollout_engine, ref_model, reward_model
 
         rewards = calculate_rewards(prompts, completions, gt_batch, tools_batch, args.num_generations, reward_model, device=args.device, turn_outputs_batch=turn_outputs_batch, unfinished_batch=unfinished_batch)
 
-        model_unwrapped = model.module if isinstance(model, DistributedDataParallel) else model
         with autocast_ctx:
-            res = model_unwrapped(input_ids, attention_mask=full_mask)
+            # 反向传播必须经过 DDP 包装后的模块：直接调用 .module 会跳过
+            # DDP 的 prepare_for_backward，梯度不会 all-reduce，各卡静默发散。
+            res = model(input_ids, attention_mask=full_mask)
             aux_loss = res.aux_loss if lm_config.use_moe else torch.tensor(0.0, device=args.device)
             logits = res.logits[:, :-1, :]
             per_token_logps = F.log_softmax(logits, dim=-1).gather(2, input_ids[:, 1:].unsqueeze(-1)).squeeze(-1)
@@ -478,7 +473,8 @@ if __name__ == "__main__":
         Logger('torch.compile enabled')
         rollout_engine.update_policy(model)
     if dist.is_initialized():
-        model = DistributedDataParallel(model, device_ids=[local_rank])
+        # 同 train_ppo：RoPE buffer 各 rank 一致，每步广播纯属浪费
+        model = DistributedDataParallel(model, device_ids=[local_rank], broadcast_buffers=False)
     rollout_engine.update_policy(model)
 
     for epoch in range(start_epoch, args.epochs):

trainer/train_dpo.py

@@ -50,6 +50,19 @@ def dpo_loss(ref_log_probs, policy_log_probs, mask, beta):
     return loss.mean()
 
 
+def sft_ce_loss(logits, labels, mask):
+    """SFT正则项：对chosen回答部分计算标准语言建模交叉熵（按掩码位置取平均），
+    与SFT阶段的损失口径一致，用于防止偏好优化中遗忘语言能力。"""
+    vocab_size = logits.size(-1)
+    ce = F.cross_entropy(
+        logits.reshape(-1, vocab_size),
+        labels.reshape(-1),
+        reduction='none'
+    )
+    mask_flat = mask.reshape(-1).float()
+    return (ce * mask_flat).sum() / mask_flat.sum().clamp(min=1.0)
+
+
 def train_epoch(epoch, loader, iters, ref_model, lm_config, start_step=0, wandb=None, beta=0.1):
     start_time = time.time()
     last_step = start_step
@@ -81,7 +94,13 @@ def train_epoch(epoch, loader, iters, ref_model, lm_config, start_step=0, wandb=
             policy_log_probs = logits_to_log_probs(logits, y)
             
             dpo_loss_val = dpo_loss(ref_log_probs, policy_log_probs, mask, beta=beta)
-            loss = dpo_loss_val + outputs.aux_loss
+            # SFT/CE正则项：只对chosen回答部分计算交叉熵，权重由 --ce_alpha 控制
+            chosen_num = x.size(0) // 2
+            if args.ce_alpha > 0:
+                ce_loss_val = sft_ce_loss(logits[:chosen_num], y[:chosen_num], mask[:chosen_num])
+            else:
+                ce_loss_val = logits.new_zeros(())
+            loss = dpo_loss_val + args.ce_alpha * ce_loss_val + outputs.aux_loss
             loss = loss / args.accumulation_steps
 
         scaler.scale(loss).backward()
@@ -97,13 +116,14 @@ def train_epoch(epoch, loader, iters, ref_model, lm_config, start_step=0, wandb=
             spend_time = time.time() - start_time
             current_loss = loss.item() * args.accumulation_steps
             current_dpo_loss = dpo_loss_val.item()
+            current_ce_loss = ce_loss_val.item()
             current_aux_loss = outputs.aux_loss.item()
             current_lr = optimizer.param_groups[-1]['lr']
             eta_min = spend_time / max(step - start_step, 1) * (iters - step) // 60
             
-            Logger(f'Epoch:[{epoch + 1}/{args.epochs}]({step}/{iters}), loss: {current_loss:.4f}, dpo_loss: {current_dpo_loss:.4f}, aux_loss: {current_aux_loss:.4f}, learning_rate: {current_lr:.8f}, epoch_time: {eta_min:.3f}min')
+            Logger(f'Epoch:[{epoch + 1}/{args.epochs}]({step}/{iters}), loss: {current_loss:.4f}, dpo_loss: {current_dpo_loss:.4f}, ce_loss: {current_ce_loss:.4f}, aux_loss: {current_aux_loss:.4f}, learning_rate: {current_lr:.8f}, epoch_time: {eta_min:.3f}min')
             
-            if wandb: wandb.log({"loss": current_loss, "dpo_loss": current_dpo_loss, "aux_loss": current_aux_loss, "learning_rate": current_lr, "epoch_time": eta_min})
+            if wandb: wandb.log({"loss": current_loss, "dpo_loss": current_dpo_loss, "ce_loss": current_ce_loss, "aux_loss": current_aux_loss, "learning_rate": current_lr, "epoch_time": eta_min})
 
         if (step % args.save_interval == 0 or step == iters) and is_main_process():
             model.eval()
@@ -118,7 +138,7 @@ def train_epoch(epoch, loader, iters, ref_model, lm_config, start_step=0, wandb=
             del state_dict
 
         del x_chosen, x_rejected, y_chosen, y_rejected, mask_chosen, mask_rejected, x, y, mask
-        del ref_outputs, ref_logits, ref_log_probs, outputs, logits, policy_log_probs, loss
+        del ref_outputs, ref_logits, ref_log_probs, outputs, logits, policy_log_probs, ce_loss_val, loss
 
     if last_step > start_step and last_step % args.accumulation_steps != 0:
         scaler.unscale_(optimizer)
@@ -150,6 +170,7 @@ if __name__ == "__main__":
     parser.add_argument('--from_weight', default='full_sft', type=str, help="基于哪个权重训练")
     parser.add_argument('--from_resume', default=0, type=int, choices=[0, 1], help="是否自动检测&续训（0=否，1=是）")
     parser.add_argument('--beta', default=0.15, type=float, help="DPO中的beta参数")
+    parser.add_argument('--ce_alpha', default=0.1, type=float, help="SFT/CE正则项权重（0表示关闭，建议0.05~0.5；防止DPO训练中遗忘语言能力）")
     parser.add_argument("--use_wandb", action="store_true", help="是否使用wandb")
     parser.add_argument("--wandb_project", type=str, default="MiniMind-DPO", help="wandb项目名")
     parser.add_argument("--use_compile", default=0, type=int, choices=[0, 1], help="是否使用torch.compile加速（0=否，1=是）")

trainer/train_grpo.py

@@ -95,9 +95,10 @@ def grpo_train_epoch(epoch, loader, iters, rollout_engine, ref_model, reward_mod
 
         rewards = calculate_rewards(prompts, completions, reward_model).to(args.device)  # [B*num_gen]
 
-        model_unwrapped = model.module if isinstance(model, DistributedDataParallel) else model
         with autocast_ctx:
-            res = model_unwrapped(outputs, attention_mask=full_mask)
+            # 反向传播必须经过 DDP 包装后的模块：直接调用 .module 会跳过
+            # DDP 的 prepare_for_backward，梯度不会 all-reduce，各卡静默发散。
+            res = model(outputs, attention_mask=full_mask)
             aux_loss = res.aux_loss if lm_config.use_moe else torch.tensor(0.0, device=args.device)
             per_token_logps = F.log_softmax(res.logits[:, :-1, :], dim=-1).gather(2, outputs[:, 1:].unsqueeze(-1)).squeeze(-1).gather(1, logp_pos)
 
@@ -313,7 +314,8 @@ if __name__ == "__main__":
         Logger('torch.compile enabled')
         rollout_engine.update_policy(model)
     if dist.is_initialized():
-        model = DistributedDataParallel(model, device_ids=[local_rank])
+        # 同 train_ppo：RoPE buffer 各 rank 一致，每步广播纯属浪费
+        model = DistributedDataParallel(model, device_ids=[local_rank], broadcast_buffers=False)
     rollout_engine.update_policy(model)
     
     # ========== 8. 开始训练 ==========

trainer/train_ppo.py

@@ -36,7 +36,7 @@ def rep_penalty(text, n=3, cap=0.5):
 class CriticModel(MiniMindForCausalLM):
     def __init__(self, params):
         super().__init__(params)
-        # 替换lm_head为输出单一价值的线性层
+        # lm_head 不参与 forward，仅靠 tie_word_embeddings 与 embed_tokens 共享权重，解绑后 DDP 会报未使用参数
         self.value_head = nn.Linear(params.hidden_size, 1)
 
     def forward(self, input_ids=None, attention_mask=None, **kwargs):
@@ -159,8 +159,6 @@ def ppo_train_epoch(epoch, loader, iters, rollout_engine, ref_model, actor_sched
         clipfrac_sum = 0.0
         aux_loss_sum = 0.0
         log_count = 0
-        actor_unwrapped = actor_model.module if isinstance(actor_model, DistributedDataParallel) else actor_model
-        critic_unwrapped = critic_model.module if isinstance(critic_model, DistributedDataParallel) else critic_model
         for ppo_epoch in range(args.ppo_update_iters):
             if stop_ppo:
                 break
@@ -168,11 +166,13 @@ def ppo_train_epoch(epoch, loader, iters, rollout_engine, ref_model, actor_sched
             for i in range(0, B, mb_size):
                 inds = b_inds[i:i + mb_size]
                 
-                mb_values_seq = critic_unwrapped(input_ids=gen_out[inds], attention_mask=full_mask[inds])
+                # 反向传播必须经过 DDP 包装后的模块：直接调用 .module 会跳过
+                # DDP 的 prepare_for_backward，梯度不会 all-reduce，各卡静默发散。
+                mb_values_seq = critic_model(input_ids=gen_out[inds], attention_mask=full_mask[inds])
                 mb_resp_values = mb_values_seq.gather(1, logp_pos[inds])
 
                 with autocast_ctx:
-                    res = actor_unwrapped(input_ids=gen_out[inds], attention_mask=full_mask[inds])
+                    res = actor_model(input_ids=gen_out[inds], attention_mask=full_mask[inds])
                     aux_loss = res.aux_loss if lm_config.use_moe else torch.tensor(0.0, device=args.device)
                     # 在 autocast 内计算 log_softmax，避免直接对 fp16/bf16 logits
                     # 计算造成额外数值偏差。
@@ -191,7 +191,7 @@ def ppo_train_epoch(epoch, loader, iters, rollout_engine, ref_model, actor_sched
                                f"ratio_max={torch.exp(_lrv).max().item():.6f} "
                                f"ratio_min={torch.exp(_lrv).min().item():.6f} "
                                f"dropout={getattr(lm_config, 'dropout', None)} "
-                               f"training={actor_unwrapped.training}")
+                               f"training={actor_model.training}")
                 approx_kl = (0.5 * (log_ratio ** 2) * resp_policy_mask[inds]).sum() / resp_policy_mask[inds].sum().clamp(min=1)
                 
                 # 同步各卡的 approx_kl，防止某卡 break 而其它卡继续导致 DDP 死锁
@@ -429,8 +429,9 @@ if __name__ == "__main__":
         Logger('torch.compile enabled')
         rollout_engine.update_policy(actor_model)
     if dist.is_initialized():
-        actor_model = DistributedDataParallel(actor_model, device_ids=[local_rank])
+        # freqs_cos/freqs_sin 各 rank 由 config 确定性算出，默认每步广播一次纯属浪费
-        critic_model = DistributedDataParallel(critic_model, device_ids=[local_rank])
+        actor_model = DistributedDataParallel(actor_model, device_ids=[local_rank], broadcast_buffers=False)
+        critic_model = DistributedDataParallel(critic_model, device_ids=[local_rank], broadcast_buffers=False)
     rollout_engine.update_policy(actor_model)
     
     # ========== 8. 开始训练 ==========

trainer/trainer_utils.py

@@ -5,6 +5,8 @@ import os
 import sys
 __package__ = "trainer"
 sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
+import ast
+import operator
 import random
 import math
 import numpy as np
@@ -175,3 +177,33 @@ class LMForRewardModel:
         ]
         score = self.model.get_score(self.tokenizer, eval_messages)
         return max(min(score, 3.0), -3.0)
+
+
+# ===== 数学表达式安全求值：替代 eval，只放行算术运算与 math 白名单（长度上限 512） =====
+def safe_math_eval(expression):
+    """对模型生成的数学表达式求值：支持 + - * / // % **、math 白名单函数与 pi/e/tau 常量。"""
+    def pow_guard(base, exp):  # 幂运算统一走这里，拦截 9**9**9 / 10**99999 这类把进程算死的输入
+        if abs(exp) > 1e4 or (abs(base) > 1 and abs(exp) * math.log10(abs(base)) > 100): raise ValueError('幂运算结果过大')
+        return base ** exp
+    def resolve(node):  # 同时兼容 sqrt(4) 与 math.sqrt(4) 两种写法
+        if isinstance(node, ast.Name): return node.id
+        if isinstance(node, ast.Attribute) and getattr(node.value, 'id', '') == 'math': return node.attr
+    def walk(node):
+        if isinstance(node, ast.Constant) and type(node.value) in (int, float): return node.value
+        name = resolve(node)
+        if name in consts: return consts[name]
+        if isinstance(node, ast.UnaryOp): return unary_ops[type(node.op)](walk(node.operand))
+        if isinstance(node, ast.BinOp): return bin_ops[type(node.op)](walk(node.left), walk(node.right))
+        if isinstance(node, ast.Call) and not node.keywords: return funcs[resolve(node.func)](*map(walk, node.args))
+        raise ValueError(f'不支持的表达式语法: {type(node).__name__}')
+    consts = {'pi': math.pi, 'e': math.e, 'tau': math.tau}
+    funcs = {'pow': pow_guard, **{n: getattr(math, n) for n in 'sqrt exp log log2 log10 sin cos tan asin acos atan atan2 sinh cosh tanh floor ceil trunc fabs fmod hypot gcd degrees radians'.split()}}
+    bin_ops = {ast.Add: operator.add, ast.Sub: operator.sub, ast.Mult: operator.mul, ast.Div: operator.truediv, ast.FloorDiv: operator.floordiv, ast.Mod: operator.mod, ast.Pow: pow_guard}
+    unary_ops = {ast.UAdd: operator.pos, ast.USub: operator.neg}
+    chars = str.maketrans({'^': '**', '×': '*', '÷': '/', '−': '-', '²': '**2', '³': '**3', '（': '(', '）': ')'})
+    expr = str(expression).translate(chars).strip()
+    if not expr or len(expr) > 512: raise ValueError('表达式为空或过长')
+    try:
+        return walk(ast.parse(expr, mode='eval').body)
+    except (KeyError, SyntaxError, TypeError):
+        raise ValueError('不支持的表达式语法') from None