diff --git a/python/sglang/srt/layers/rotary_embedding.py b/python/sglang/srt/layers/rotary_embedding.py
index 56516b41b..480579e01 100644
--- a/python/sglang/srt/layers/rotary_embedding.py
+++ b/python/sglang/srt/layers/rotary_embedding.py
@@ -16,6 +16,7 @@ from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import (
     cpu_has_amx_support,
     get_bool_env_var,
+    get_compiler_backend,
     is_cpu,
     is_cuda,
     is_hip,
@@ -2748,6 +2749,7 @@ def rotate_half(x):
     return torch.cat((-x2, x1), dim=-1)
 
 
+@torch.compile(dynamic=True, backend=get_compiler_backend())
 def apply_rotary_pos_emb_native(
     q: torch.Tensor,
     k: torch.Tensor,
diff --git a/python/sglang/srt/model_executor/forward_batch_info.py b/python/sglang/srt/model_executor/forward_batch_info.py
index 94fb8a545..4402d0928 100644
--- a/python/sglang/srt/model_executor/forward_batch_info.py
+++ b/python/sglang/srt/model_executor/forward_batch_info.py
@@ -668,15 +668,8 @@ class ForwardBatch:
         self,
         mm_input: MultimodalInputs,
         seq_len: int,
-        device: torch.device,
     ) -> torch.Tensor:
-        if mm_input.mrope_position_delta.device.type != device:
-            # transfer mrope_position_delta to device when the first running,
-            # avoiding successvie host-to-device data transfer
-            mm_input.mrope_position_delta = mm_input.mrope_position_delta.to(
-                device, non_blocking=True
-            )
-
+        # doing below compute on cpu to avoid frequent small kernels
         mrope_position_deltas = mm_input.mrope_position_delta.flatten()
         mrope_positions = (
             (mrope_position_deltas + seq_len - 1).unsqueeze(0).repeat(3, 1)
@@ -701,11 +694,10 @@ class ForwardBatch:
                         (3, 1),
                         self.seq_lens_cpu[batch_idx] - 1,
                         dtype=torch.int64,
-                        device=model_runner.device,
                     )
                 else:
                     mrope_positions = self._expand_mrope_from_input(
-                        mm_input, self.seq_lens_cpu[batch_idx], model_runner.device
+                        mm_input, self.seq_lens_cpu[batch_idx]
                     )
                     mrope_positions_list[batch_idx] = mrope_positions
             elif self.forward_mode.is_extend():
@@ -737,14 +729,14 @@ class ForwardBatch:
                     ]
                     if mrope_positions.numel() == 0:
                         mrope_positions = self._expand_mrope_from_input(
-                            mm_input, self.seq_lens[batch_idx], model_runner.device
+                            mm_input, self.seq_lens_cpu[batch_idx]
                         )
                 mrope_positions_list[batch_idx] = mrope_positions
 
         self.mrope_positions = torch.cat(
-            [pos.to(device=model_runner.device) for pos in mrope_positions_list],
+            [pos for pos in mrope_positions_list],
             dim=1,
-        ).to(dtype=torch.int64, device=model_runner.device)
+        ).to(dtype=torch.int64, device=model_runner.device, non_blocking=True)
 
     def get_max_chunk_capacity(self):
         # Maximum number of tokens in each chunk
diff --git a/python/sglang/srt/models/qwen2_5_vl.py b/python/sglang/srt/models/qwen2_5_vl.py
index d2753fbf1..aafde99bb 100644
--- a/python/sglang/srt/models/qwen2_5_vl.py
+++ b/python/sglang/srt/models/qwen2_5_vl.py
@@ -47,6 +47,7 @@ from sglang.srt.distributed import (
 )
 from sglang.srt.distributed.parallel_state import get_pp_group
 from sglang.srt.environ import envs
+from sglang.srt.layers.activation import SiluAndMul
 from sglang.srt.layers.attention.vision import VisionAttention
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.linear import (
@@ -114,12 +115,21 @@ class Qwen2_5_VLMLP(nn.Module):
             tp_size=self.tp_size,
             tp_rank=self.tp_rank,
         )
-        self.act = ACT2FN[hidden_act]
+        self.hidden_act = hidden_act
+        if self.hidden_act == "silu":
+            self.act = SiluAndMul()
+        else:
+            base_act = ACT2FN[self.hidden_act]
+
+            def _act_fn(x: torch.Tensor) -> torch.Tensor:
+                gate, up = x.chunk(2, dim=-1)
+                return base_act(gate) * up
+
+            self.act = _act_fn
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         gate_up, _ = self.gate_up_proj(x)
-        gate, up = gate_up.chunk(2, dim=-1)
-        x = self.act(gate) * up
+        x = self.act(gate_up)
         x_down, _ = self.down_proj(x)
         return x_down