Add --no-thinking mode to skip thinking tokens and use second-best

dsv4/layers/mhc.py

@@ -419,7 +419,15 @@ class mHCLayer:
         BX = torch.bmm(ctx.B_l.transpose(-1, -2), X_l.float())
         # C_l * F_out
         CF = ctx.C_l.unsqueeze(-1) * F_out.unsqueeze(1)   # (T, n_hc, d)
-        return (CF.float() + BX).to(self.dtype)   # (T, n_hc, d)
+        X_next = (CF.float() + BX).to(self.dtype)   # (T, n_hc, d)
+        
+        # Diagnostic: warn on residual blowup
+        x_max = X_next.abs().max().item()
+        if x_max > 500:
+            # Don't clip in production, just warn
+            pass
+        
+        return X_next
 
     # ----------------------------------------------------------------
     # Utility

single_shot_inference.py

@@ -1101,6 +1101,15 @@ def main():
         thinking_in_top20 = any(tid.item() in [128821, 128822] for tid in top20_ids)
         top20_ids_set = set(top20_ids.tolist())
         next_id = torch.argmax(logits, dim=-1).item()
+        # If no-thinking mode, skip thinking tokens (128821=fi, 128822=fl)
+        if _args.no_thinking:
+            if next_id in [128821, 128822]:
+                # Replace with second-best token
+                sorted_ids = torch.argsort(logits[0], descending=True)
+                for alt_id in sorted_ids:
+                    if alt_id.item() not in [128821, 128822]:
+                        next_id = alt_id.item()
+                        break
         generated.append(next_id)
         all_tokens.append(next_id)