[Bugfix][Model] Fix Mllama SDPA illegal memory access for batched multi-image (#9626)

Signed-off-by: mgoin <michael@neuralmagic.com>

vllm/model_executor/models/mllama.py

@@ -795,17 +795,19 @@ class MllamaTextCrossAttention(nn.Module):
         kv_len = k.shape[0]
         q = q.transpose(0, 1).view(self.num_local_key_value_heads,
                                    self.num_key_value_groups, q_len,
-                                   self.head_dim)
+                                   self.head_dim).contiguous()
         k = k.transpose(0,
                         1)[:,
                            None, :, :].expand(self.num_local_key_value_heads,
                                               self.num_key_value_groups,
-                                              kv_len, self.head_dim)
+                                              kv_len,
+                                              self.head_dim).contiguous()
         v = v.transpose(0,
                         1)[:,
                            None, :, :].expand(self.num_local_key_value_heads,
                                               self.num_key_value_groups,
-                                              kv_len, self.head_dim)
+                                              kv_len,
+                                              self.head_dim).contiguous()
         attention_mask = attention_mask.view(1, 1, q_len, kv_len)
         output = F.scaled_dot_product_attention(q,
                                                 k,