diff --git a/python/minisgl/core.py b/python/minisgl/core.py
index be4d643e..0a70be81 100644
--- a/python/minisgl/core.py
+++ b/python/minisgl/core.py
@@ -40,6 +40,7 @@ def __post_init__(self) -> None:
         self.device_len = len(self.input_ids)
         self.max_device_len = len(self.input_ids) + self.output_len
         assert 0 <= self.cached_len < self.device_len <= self.max_device_len
+        self._host_buf: torch.Tensor | None = None
 
     @property
     def remain_len(self) -> int:
@@ -54,7 +55,16 @@ def complete_one(self) -> None:
         self.device_len += 1
 
     def append_host(self, next_token: torch.Tensor) -> None:
-        self.input_ids = torch.cat([self.input_ids, next_token])
+        # Write into a lazily pre-allocated buffer (O(1) per token) instead of 
+        # using torch.cat (O(n^2) per request) to prevent GPU stalling on the scheduler thread.
+        # Lazy allocation ensures non-decoding requests pay zero memory overhead.
+        # Previously written positions are read-only to guarantee radix cache safety.
+        if self._host_buf is None:
+             self._host_buf = torch.empty(self.max_device_len, dtype=self.input_ids.dtype)
+             self._host_buf[: len(self.input_ids)] = self.input_ids
+         new_len = len(self.input_ids) + 1
+         self._host_buf[new_len - 1] = next_token
+         self.input_ids = self._host_buf[:new_len]
 
     @property
     def can_decode(self) -> bool:
diff --git a/tests/core/test_req_append.py b/tests/core/test_req_append.py
new file mode 100644
index 00000000..ff5334f3
--- /dev/null
+++ b/tests/core/test_req_append.py
@@ -0,0 +1,82 @@
+from __future__ import annotations
+
+import pytest
+import torch
+from minisgl.core import Req, SamplingParams
+
+
+def _make_req(prompt_len: int = 8, output_len: int = 16, dtype=torch.int32) -> Req:
+    return Req(
+        input_ids=torch.arange(prompt_len, dtype=dtype),
+        table_idx=0,
+        cached_len=0,
+        output_len=output_len,
+        uid=0,
+        sampling_params=SamplingParams(max_tokens=output_len),
+        cache_handle=None,  # type: ignore[arg-type]
+    )
+
+
+def test_append_host_matches_cat_reference():
+    req = _make_req(prompt_len=8, output_len=16)
+    reference = req.input_ids.clone()
+
+    for token in range(100, 116):
+        next_token = torch.tensor([token], dtype=torch.int32)
+        reference = torch.cat([reference, next_token])
+        req.complete_one()
+        req.append_host(next_token)
+        assert req.input_ids.is_cpu
+        assert torch.equal(req.input_ids, reference)
+        assert len(req.input_ids) == req.device_len
+
+
+def test_append_host_preserves_dtype():
+    for dtype in (torch.int32, torch.int64):
+        req = _make_req(dtype=dtype)
+        req.complete_one()
+        req.append_host(torch.tensor([42], dtype=torch.int32))
+        assert req.input_ids.dtype == dtype
+
+
+def test_append_host_never_mutates_written_positions():
+    # The radix cache stores slices of `input_ids` by reference (see
+    # RadixTreeNode.set_key_value), so previously-appended positions must
+    # never change under later appends.
+    req = _make_req(prompt_len=4, output_len=8)
+    req.complete_one()
+    req.append_host(torch.tensor([100], dtype=torch.int32))
+    snapshot_view = req.input_ids[:5]  # what a radix key would hold
+    snapshot_copy = snapshot_view.clone()
+
+    for token in range(101, 108):
+        req.complete_one()
+        req.append_host(torch.tensor([token], dtype=torch.int32))
+
+    assert torch.equal(snapshot_view, snapshot_copy)
+
+
+def test_append_host_overflow_raises():
+    # Appending beyond max_device_len must fail loudly, not silently drop.
+    req = _make_req(prompt_len=2, output_len=1)
+    req.complete_one()
+    req.append_host(torch.tensor([7], dtype=torch.int32))
+    req.complete_one()
+    with pytest.raises(IndexError):
+        req.append_host(torch.tensor([8], dtype=torch.int32))
+
+
+def test_chunked_req_append_still_forbidden():
+    from minisgl.scheduler.prefill import ChunkedReq
+
+    req = ChunkedReq(
+        input_ids=torch.arange(4, dtype=torch.int32),
+        table_idx=0,
+        cached_len=0,
+        output_len=4,
+        uid=0,
+        sampling_params=SamplingParams(),
+        cache_handle=None,  # type: ignore[arg-type]
+    )
+    with pytest.raises(NotImplementedError):
+        req.append_host(torch.tensor([1], dtype=torch.int32))