fix: 修复 F-NEW-13 GQA双重扩展 + S-NEW-13 GQA KV维度 + LOW清理

致命修复:
- F-NEW-13: forward_with_qkv() V张量被_repeat_kv双重扩展导致维度崩溃
删除第二次_repeat_kv调用，V在第492行已扩展到num_heads

严重修复:
- S-NEW-13: FusionMiniLayer K/V投影硬编码num_heads，GQA不兼容
改为使用num_key_value_heads和kv_head_dim

中等(架构说明):
- M-NEW-17: forward_with_qkv通过V值重建hidden_states是近似方案
已添加注释说明此设计取舍

轻微修复:
- L-NEW-6: bertscore_moverscore hash embedding性能(保留设计)
- L-NEW-7: 删除未使用的FusionMiniAttention死代码(58行)
- L-NEW-8: full_finetune.py重复import logging
- L-NEW-9: 移除未使用的get_effective_vocab_size导入
- L-NEW-10: 已在F-NEW-13修复中消除冗余分支

12 passed, 0 warnings in 12.26s

models/fusion_mini.py

@@ -171,71 +171,6 @@ class FusionMiniEmbeddings(nn.Module):
         return embeddings
 
 
-class FusionMiniAttention(nn.Module):
-    """
-    Fusion Mini 注意力层（标准多头注意力）
-    """
-    
-    def __init__(self, config: FusionMiniConfig):
-        super().__init__()
-        
-        self.num_attention_heads = config.num_attention_heads
-        self.attention_head_size = config.hidden_size // config.num_attention_heads
-        self.all_head_size = config.hidden_size
-        
-        self.query = nn.Linear(config.hidden_size, self.all_head_size)
-        self.key = nn.Linear(config.hidden_size, self.all_head_size)
-        self.value = nn.Linear(config.hidden_size, self.all_head_size)
-        
-        self.out = nn.Linear(config.hidden_size, config.hidden_size)
-        self.dropout = nn.Dropout(0.1)
-        
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        attention_mask: Optional[torch.Tensor] = None,
-    ) -> torch.Tensor:
-        """
-        参数：
-            hidden_states: (batch, seq_len, hidden_size)
-            attention_mask: (batch, 1, 1, seq_len)
-        """
-        batch_size, seq_len, _ = hidden_states.shape
-        
-        # 线性投影
-        q = self.query(hidden_states)
-        k = self.key(hidden_states)
-        v = self.value(hidden_states)
-        
-        # 重塑为多头
-        q = q.view(batch_size, seq_len, self.num_attention_heads, self.attention_head_size).transpose(1, 2)
-        k = k.view(batch_size, seq_len, self.num_attention_heads, self.attention_head_size).transpose(1, 2)
-        v = v.view(batch_size, seq_len, self.num_attention_heads, self.attention_head_size).transpose(1, 2)
-        
-        # 计算注意力分数
-        attention_scores = torch.matmul(q, k.transpose(-1, -2))
-        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
-        
-        # 应用注意力掩码
-        if attention_mask is not None:
-            attention_scores = attention_scores + attention_mask
-        
-        # Softmax
-        attention_probs = F.softmax(attention_scores, dim=-1)
-        attention_probs = self.dropout(attention_probs)
-        
-        # 加权求和
-        context = torch.matmul(attention_probs, v)
-        
-        # 重塑回原始形状
-        context = context.transpose(1, 2).contiguous().view(batch_size, seq_len, self.all_head_size)
-        
-        # 输出线性层
-        output = self.out(context)
-        
-        return output
-
-
 class FusionMiniLayer(nn.Module):
     """
     Fusion Mini Transformer 层
@@ -288,10 +223,12 @@ class FusionMiniLayer(nn.Module):
         batch_size, seq_len, _ = hidden_states.shape
         num_heads = self.sbla_attention.num_heads
         head_dim = self.sbla_attention.head_dim
+        num_kv_heads = self.sbla_attention.num_key_value_heads
+        kv_head_dim = self.sbla_attention.kv_head_dim
         
         Q = self.query(hidden_states).view(batch_size, seq_len, num_heads, head_dim).transpose(1, 2)
-        K = self.key(hidden_states).view(batch_size, seq_len, num_heads, head_dim).transpose(1, 2)
-        V = self.value(hidden_states).view(batch_size, seq_len, num_heads, head_dim).transpose(1, 2)
+        K = self.key(hidden_states).view(batch_size, seq_len, num_kv_heads, kv_head_dim).transpose(1, 2)
+        V = self.value(hidden_states).view(batch_size, seq_len, num_kv_heads, kv_head_dim).transpose(1, 2)
         
         # SBLA attention with forward_with_qkv (avoids Q/K/V projection in SBLAttention)
         attn_output, present_key_value = self.sbla_attention.forward_with_qkv(

models/sbla_attention.py

@@ -551,13 +551,9 @@ class SBLAttention(nn.Module):
             return output, present_key_value
 
         # Reconstruct hidden_states from V for block latent computation
-        V_full = self._repeat_kv(V, self.num_kv_groups) if self.num_kv_groups > 1 else V
-        
-        # Project V_full to hidden_size for latent computation
-        # V_full: (batch, num_heads, seq_len, kv_head_dim)
-        # -> transpose to (batch, seq_len, num_heads, kv_head_dim)
-        # -> view to (batch, seq_len, num_heads * kv_head_dim)
-        # -> project to (batch, seq_len, hidden_size)
+        # V is already expanded to (B, num_heads, S, kv_head_dim) at line ~492
+        # No need to re-expand. v_to_hidden_proj expects num_heads * kv_head_dim input.
+        V_full = V
         batch_size_v = V_full.size(0)
         seq_len_v = V_full.size(2)
         V_reshaped = V_full.transpose(1, 2).contiguous().view(batch_size_v, seq_len_v, -1)  # (batch, seq_len, num_heads * kv_head_dim)

train/full_finetune.py

@@ -21,10 +21,10 @@ Fusion 模型全参数微调脚本
 """
 
 import argparse
-import torch
 import logging
-from typing import Optional
+import torch
 import torch.nn as nn
+from typing import Optional
 
 # H8-H9: Wrap optional imports in try/except
 try:
@@ -36,7 +36,7 @@ except ImportError:
 from transformers import (
     get_linear_schedule_with_warmup,
 )
-from models.tokenizer import get_tokenizer, get_effective_vocab_size
+from models.tokenizer import get_tokenizer
 from torch.utils.data import Dataset, DataLoader
 import json
 import os