[BugFix] fix mm rope (#7274)

Author: xiaoxiaohehe001

fastdeploy/model_executor/layers/rotary_embedding.py

@@ -458,12 +458,12 @@ def __call__(self, position_ids, max_len_lst, cumsum_seqlens):
 
         # Build position_ids_3d: [bsz, max_position, 3]
         position_ids_3d = paddle.tile(
-            paddle.arange(self.max_position, dtype="int64").unsqueeze(0).unsqueeze(-1),
+            paddle.arange(self.max_position, dtype="float32").unsqueeze(0).unsqueeze(-1),
             [bsz, 1, 3],
         )
         for i in range(bsz):
             position_ids_cur = position_ids[cumsum_seqlens[i] : cumsum_seqlens[i + 1]]
-            prefix_max_position_ids = paddle.max(position_ids_cur) + 1
+            prefix_max_position_ids = paddle.max(position_ids_cur[..., 0]) + 1
             dec_pos_ids = paddle.tile(
                 paddle.arange(max_len_lst[i], dtype="int64").unsqueeze(-1),
                 [1, 3],
@@ -530,12 +530,12 @@ def __call__(self, position_ids, max_len_lst, cumsum_seqlens):
         bsz = len(cumsum_seqlens) - 1
         # position_ids_3d: [bsz, seq_len, 3]
         position_ids_3d = paddle.tile(
-            paddle.arange(self.max_position, dtype="int64").unsqueeze(0).unsqueeze(-1),
+            paddle.arange(self.max_position, dtype="float32").unsqueeze(0).unsqueeze(-1),
             [bsz, 1, 3],
         )
         for i in range(bsz):
             position_ids_cur = position_ids[cumsum_seqlens[i] : cumsum_seqlens[i + 1]]
-            prefix_max_position_ids = paddle.max(position_ids_cur) + 1
+            prefix_max_position_ids = paddle.max(position_ids_cur[..., 0]) + 1
             dec_pos_ids = paddle.tile(
                 paddle.arange(max_len_lst[i], dtype="int64").unsqueeze(-1),
                 [1, 3],

fastdeploy/worker/gpu_model_runner.py

@@ -1013,7 +1013,7 @@ def insert_tasks_v1(self, req_dicts: List[Request], num_running_requests: int =
 
         if len(rope_3d_position_ids["position_ids_idx"]) > 0 and self.enable_mm:
             packed_position_ids = paddle.to_tensor(
-                np.concatenate(rope_3d_position_ids["position_ids_lst"]), dtype="int64"
+                np.concatenate(rope_3d_position_ids["position_ids_lst"]), dtype="float32"
             )
             rope_3d_lst = self.prepare_rope3d(
                 packed_position_ids,