Fix query RoPE when reusing cached vision keys

Apply query RoPE on every cross-attention step, including cached-vision steps. Rotate new vision keys only once. Preserve the current NPU, idle-wait and text-cleanup code. No weights, configs, processors or tokenizer files changed. CPU fresh/cached regression passed.

modeling_moss_vl.py

@@ -644,6 +644,11 @@ class MossVLTextCrossAttention(nn.Module):
         query_states = query_states.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2)
         query_states = self.q_norm(query_states)
 
+        # Every new query needs RoPE, including reads from the vision KV cache.
+        if query_position_embeddings is not None:
+            cos, sin = query_position_embeddings
+            query_states = apply_rotary_pos_emb_cross_attention(query_states, cos, sin)
+
         if cross_attention_states is not None:
             # Key and Value from vision cross_attention_states
             key_states = self.k_proj(cross_attention_states)
@@ -653,10 +658,7 @@ class MossVLTextCrossAttention(nn.Module):
             key_states = self.k_norm(key_states)
             value_states = value_states.view(batch_size, -1, self.num_key_value_heads, self.head_dim).transpose(1, 2)
 
-            # Apply different RoPE for query (text position) and key (vision position)
-            if query_position_embeddings is not None:
-                cos, sin = query_position_embeddings
-                query_states = apply_rotary_pos_emb_cross_attention(query_states, cos, sin)
+            # Rotate new keys once; cached keys already include vision RoPE.
             
             if vision_position_embeddings is not None:
                 vision_cos, vision_sin = vision_position_embeddings