Merge 743420daac into 89d674b8a5

Co-authored-by: Cursor <cursoragent@cursor.com>

.gitignore

@@ -4,3 +4,9 @@ __pycache__/
 out
 website/
 docs-minimind/
+
+checkpoints/
+dataset/
+
+minimind-3/
+minimind-3-moe/

scripts/.gitignore

@@ -0,0 +1,5 @@
+# .gitignore 忽略下列文件和目录
+
+#比如minimind\scripts\minimind-3和minimind\scripts\minimind-3-moe
+minimind-3/
+minimind-3-moe/

scripts/web_demo.py

@@ -64,6 +64,28 @@ st.markdown("""
             margin: 0 2px !important;  /* 调整这里的 margin 值 */
         }
         
+        /* 单独恢复侧边栏按钮（新建对话）的正常横向长方形外观，避免文字被挤成竖排 */
+        [data-testid="stSidebar"] .stButton > button {
+            border-radius: 8px !important;      /* 恢复稍微圆角的矩形 */
+            width: 100% !important;             /* 撑满侧边栏宽度 */
+            height: auto !important;
+            min-width: 100% !important;         /* 覆盖掉 18px 的限制 */
+            min-height: 40px !important;
+            max-width: 100% !important;
+            max-height: none !important;
+            padding: 8px !important;            /* 恢复内边距 */
+            background-color: #f4f6f8 !important; /* 增加一点按钮底色以免透明看不清 */
+            border: 1px solid #e0e0e0 !important;
+            font-size: 15px !important;
+            color: #333 !important;
+            margin: 10px 0 !important;
+        }
+        [data-testid="stSidebar"] .stButton > button:hover {
+            border-color: #999 !important;
+            background-color: #e4e8f0 !important;
+            color: #000 !important;
+        }
+
     </style>
 """, unsafe_allow_html=True)
 
@@ -83,6 +105,7 @@ LANG_TEXTS = {
         'disclaimer': 'AI 生成内容可能存在错误，请仔细核实',
         'think_tip': '自适应思考，目前多轮对话或Tool Call共存时思考不稳定',
         'tool_select': '工具选择（最多4个）',
+        'new_chat': '新建对话',  # 新增
     },
     'en': {
         'settings': 'Model Settings',
@@ -96,6 +119,7 @@ LANG_TEXTS = {
         'disclaimer': 'AI-generated content may be inaccurate, please verify',
         'think_tip': 'Adaptive thinking; may be unstable with multi-turn or Tool Call',
         'tool_select': 'Tool Selection (max 4)',
+        'new_chat': 'New Chat',  # 新增
     }
 }
 
@@ -237,15 +261,20 @@ def regenerate_answer(index):
 
 
 # 动态扫描模型目录
-script_dir = os.path.dirname(os.path.abspath(__file__))
-MODEL_PATHS = {}
-for d in sorted(os.listdir(script_dir), reverse=True):
-    full_path = os.path.join(script_dir, d)
-    if os.path.isdir(full_path) and not d.startswith('.') and not d.startswith('_'):
-        if any(f.endswith(('.bin', '.safetensors', '.pt')) or os.path.exists(os.path.join(full_path, 'model.safetensors.index.json')) for f in os.listdir(full_path) if os.path.isfile(os.path.join(full_path, f))):
-            MODEL_PATHS[d] = [d, d]
-if not MODEL_PATHS:
-    MODEL_PATHS = {"No models found": ["", "No models"]}
+@st.cache_data(ttl=3600)  # 增加缓存(有效期1小时)，避免每次刷新页面都去高频读取硬盘文件树
+def get_model_paths():
+    script_dir = os.path.dirname(os.path.abspath(__file__))
+    paths = {}
+    for d in sorted(os.listdir(script_dir), reverse=True):
+        full_path = os.path.join(script_dir, d)
+        if os.path.isdir(full_path) and not d.startswith('.') and not d.startswith('_'):
+            if any(f.endswith(('.bin', '.safetensors', '.pt')) or os.path.exists(os.path.join(full_path, 'model.safetensors.index.json')) for f in os.listdir(full_path) if os.path.isfile(os.path.join(full_path, f))):
+                paths[d] =[d, d]
+    if not paths:
+        paths = {"No models found": ["", "No models"]}
+    return paths
+
+MODEL_PATHS = get_model_paths()
 
 # 模型选择
 selected_model = st.sidebar.selectbox('Model', list(MODEL_PATHS.keys()), index=0)
@@ -285,7 +314,27 @@ with st.sidebar.expander(get_text('tools')):
         if checked and len(st.session_state.selected_tools) < 4:
             st.session_state.selected_tools.append(name)
 
-image_url = "https://www.modelscope.cn/api/v1/studio/gongjy/MiniMind/repo?Revision=master&FilePath=images%2Flogo2.png&View=true"
+# 新建对话
+st.sidebar.markdown('<hr style="margin: 12px 0 16px 0;">', unsafe_allow_html=True)
+
+# 新建对话
+if st.sidebar.button(get_text('new_chat'), key="new_chat", use_container_width=True):
+    clear_chat_messages()
+    st.rerun()
+
+# 将外部网络图片改为读取本地图片转 base64 渲染
+import base64
+@st.cache_data
+def get_image_base64(image_path):
+    try:
+        with open(image_path, "rb") as img_file:
+            return f"data:image/png;base64,{base64.b64encode(img_file.read()).decode()}"
+    except:
+        # 如果本地图片不存在，再退回使用网络URL
+        return "https://www.modelscope.cn/api/v1/studio/gongjy/MiniMind/repo?Revision=master&FilePath=images%2Flogo2.png&View=true"
+
+local_logo_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "images", "logo2.png")
+image_url = get_image_base64(local_logo_path)
 
 st.markdown(
     f'<div style="display: flex; flex-direction: column; align-items: center; text-align: center; margin: 0; padding: 0;">'

trainer/train_ppo.py

@@ -12,7 +12,6 @@ import warnings
 import torch
 import torch.distributed as dist
 import torch.nn.functional as F
-from transformers import AutoTokenizer
 from contextlib import nullcontext
 from torch import optim, nn
 from torch.nn.parallel import DistributedDataParallel
@@ -76,7 +75,7 @@ def calculate_rewards(prompts, responses, reward_model):
     return rewards
 
 
-def ppo_train_epoch(epoch, loader, iters, rollout_engine, ref_model, actor_scheduler, critic_scheduler, reward_model, start_step=0, wandb=None, use_sglang=False):
+def ppo_train_epoch(epoch, loader, iters, rollout_engine, ref_model, actor_scheduler, critic_scheduler, reward_model, start_step=0, wandb=None):
     actor_model.train()
     critic_model.train()
     grad_accum_step = 0
@@ -174,10 +173,24 @@ def ppo_train_epoch(epoch, loader, iters, rollout_engine, ref_model, actor_sched
                 with autocast_ctx:
                     res = actor_unwrapped(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)
-
-                mb_resp_logp = F.log_softmax(res.logits[:, :-1], dim=-1).gather(2, labels[inds].unsqueeze(-1)).squeeze(-1).gather(1, logp_pos[inds])
+                    # 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。
+                    mb_resp_logp = F.log_softmax(res.logits[:, :-1], dim=-1).gather(2, labels[inds].unsqueeze(-1)).squeeze(-1).gather(1, logp_pos[inds])
 
                 log_ratio = mb_resp_logp - old_resp_logp[inds]
+
+                # 可开关的诊断：观察首轮首个 minibatch 的 mb 与 old logp 差异。
+                if args.debug_log_ratio and ppo_epoch == 0 and i == 0 and is_main_process():
+                    _lr = log_ratio.detach()
+                    _m = resp_policy_mask[inds].bool()
+                    if _m.any():
+                        _lrv = _lr[_m]
+                        Logger(f"[DBG log_ratio] step={step} max|lr|={_lrv.abs().max().item():.6e} "
+                               f"mean|lr|={_lrv.abs().mean().item():.6e} "
+                               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}")
                 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
@@ -331,6 +344,7 @@ if __name__ == "__main__":
     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）")
     parser.add_argument("--debug_mode", action="store_true", help="是否打印训练调试采样")
     parser.add_argument("--debug_interval", type=int, default=20, help="debug模式下每隔多少step打印一次采样")
+    parser.add_argument("--debug_log_ratio", action="store_true", help="打印首轮首个minibatch的log_ratio差异量级，用于核查ratio≈1是否成立")
     parser.add_argument("--thinking_ratio", type=float, default=0.9, help="按概率开启thinking（0.0~1.0）")
     parser.add_argument("--rollout_engine", type=str, default="torch", choices=["torch", "sglang"], help="rollout引擎类型")
     parser.add_argument("--sglang_base_url", type=str, default="http://localhost:8998", help="SGLang服务器URL")
@@ -427,9 +441,9 @@ if __name__ == "__main__":
         loader = DataLoader(train_ds, batch_sampler=batch_sampler, num_workers=args.num_workers, pin_memory=True)
         if skip > 0: 
             Logger(f'Epoch [{epoch + 1}/{args.epochs}]: 跳过前{start_step}个step，从step {start_step + 1}开始')
-            ppo_train_epoch(epoch, loader, len(loader) + skip, rollout_engine, ref_model, actor_scheduler, critic_scheduler, reward_model, start_step, wandb, use_sglang = (args.rollout_engine == "sglang"))
+            ppo_train_epoch(epoch, loader, len(loader) + skip, rollout_engine, ref_model, actor_scheduler, critic_scheduler, reward_model, start_step, wandb)
         else:
-            ppo_train_epoch(epoch, loader, len(loader), rollout_engine, ref_model, actor_scheduler, critic_scheduler, reward_model, 0, wandb, use_sglang = (args.rollout_engine == "sglang"))
+            ppo_train_epoch(epoch, loader, len(loader), rollout_engine, ref_model, actor_scheduler, critic_scheduler, reward_model, 0, wandb)
     
     # ========== 9. 清理分布进程 ==========
     if dist.is_initialized():