fix: SBLAttention forward() guard for removed projections

- Add runtime check in SBLAttention.forward() to prevent AttributeError
- Update test_sbla.py to use forward_with_qkv() API
- All unit tests passing (test_simple_import.py, test_sbla.py)

models/sbla_attention.py

@@ -288,6 +288,14 @@ class SBLAttention(nn.Module):
         device = hidden_states.device
         
         # ========== 1. Q/K/V 投影 ==========
+        # S-NEW-8: Q/K/V projections moved to FusionAttention (for RoPE)
+        # If you see this error, use FusionAttention or forward_with_qkv() instead
+        if not hasattr(self, 'q_proj') or self.q_proj is None:
+            raise RuntimeError(
+                "SBLAttention.forward() requires Q/K/V projections, but they were "
+                "removed to avoid parameter waste. Use FusionAttention wrapper or "
+                "call sbla.forward_with_qkv(Q, K, V, ...) instead."
+            )
         Q = self.q_proj(hidden_states)  # (batch, seq_len, num_heads * head_dim)
         K = self.k_proj(hidden_states)  # (batch, seq_len, num_kv_heads * head_dim)
         V = self.v_proj(hidden_states)

tests/test_sbla.py

@@ -19,6 +19,15 @@ batch_size, seq_len = 2, 16
 hidden_states = torch.randn(batch_size, seq_len, 64)
 attention_mask = torch.ones(batch_size, seq_len)
 
-output, cache = sbla(hidden_states=hidden_states, attention_mask=attention_mask)
+# Create Q/K/V manually (simulating FusionAttention's role)
+import torch.nn.functional as F
+q_proj = torch.nn.Linear(64, 64)
+k_proj = torch.nn.Linear(64, 64)
+v_proj = torch.nn.Linear(64, 64)
+Q = q_proj(hidden_states).view(batch_size, seq_len, 4, 16).transpose(1, 2)
+K = k_proj(hidden_states).view(batch_size, seq_len, 4, 16).transpose(1, 2)
+V = v_proj(hidden_states).view(batch_size, seq_len, 4, 16).transpose(1, 2)
+
+output, cache = sbla.forward_with_qkv(Q, K, V, attention_mask)
 print(f"OK: shape={output.shape}, no NaN={not torch.isnan(output).any()}, cache={cache}")
 print("[PASS] SBLA Attention working!")