diff --git a/tests/distributed/omni_connectors/test_cuda_ipc_connector.py b/tests/distributed/omni_connectors/test_cuda_ipc_connector.py
new file mode 100644
index 000000000..1ce9b3b66
--- /dev/null
+++ b/tests/distributed/omni_connectors/test_cuda_ipc_connector.py
@@ -0,0 +1,101 @@
+"""CudaIpcConnector: leaf walking (CPU) and a spawned producer/consumer round trip (CUDA)."""
+
+from __future__ import annotations
+
+import multiprocessing as mp
+import os
+import uuid
+
+import pytest
+import torch
+
+from tests.helpers.mark import hardware_test
+from vllm_omni.distributed.omni_connectors.connectors import cuda_ipc_connector as cic
+from vllm_omni.distributed.omni_connectors.factory import OmniConnectorFactory
+
+
+@pytest.mark.core_model
+@pytest.mark.cpu
+def test_walk_rewrites_tensor_leaves_only():
+    seen = []
+
+    def leaf(t):
+        seen.append(t)
+        return "L"
+
+    data = {"a": torch.zeros(2), "b": [1, torch.ones(1), (torch.ones(1), "s")], "c": {"d": None}}
+    out = cic._walk(data, leaf)
+    assert out == {"a": "L", "b": [1, "L", ("L", "s")], "c": {"d": None}}
+    assert len(seen) == 3
+    # marker dicts are treated as leaves too
+    marker = {cic.MARKER: 1, "args": b"", "event": b"", "dtype": "torch.float32", "shape": [1]}
+    assert cic._walk({"m": marker}, lambda m: "M") == {"m": "M"}
+    assert cic.is_marker(marker) and not cic.is_marker({"x": 1})
+
+
+@pytest.mark.core_model
+@pytest.mark.cpu
+def test_cpu_leaves_pass_through_shared_memory(tmp_path):
+    assert "CudaIpcConnector" in OmniConnectorFactory._registry
+    conn = cic.CudaIpcConnector({"stage_id": 0, "extra": {"cuda_ipc_enabled": False}})
+    key = f"cic_cpu_{uuid.uuid4().hex[:8]}"
+    payload = {"codes": {"audio": torch.arange(6).reshape(3, 2)}, "meta": {"finished": True}}
+    ok, size, meta = conn.put("stage0", "stage1", key, payload)
+    assert ok and size > 0 and "cuda_ipc" not in (meta or {})
+    got, _ = conn.get("stage0", "stage1", key, meta)
+    assert torch.equal(got["codes"]["audio"], payload["codes"]["audio"]) and got["meta"]["finished"] is True
+    conn.cleanup(key)
+    assert "cuda_ipc" in conn.health()
+    conn.close()
+
+
+def _producer(key: str, q: mp.Queue):
+    conn = cic.CudaIpcConnector({"stage_id": 0})
+    t = torch.randint(0, 2048, (25, 16), device="cuda", dtype=torch.int64)
+    hs = torch.randn(4, 8, device="cuda", dtype=torch.float16)
+    ok, size, meta = conn.put(
+        "stage0", "stage1", key, {"codes": {"audio": t}, "hidden": hs, "meta": {"n": 25}, "cpu": torch.ones(2)}
+    )
+    q.put(("put", ok, size, meta, t.cpu(), hs.cpu()))
+    # wait for consumer ack, then sweep on cleanup
+    deadline = 60
+    import time
+
+    while deadline > 0 and not os.path.exists(cic._ack_path(key)):
+        time.sleep(0.1)
+        deadline -= 0.1
+    conn.cleanup(key)
+    q.put(("done", len(conn._exported)))
+
+
+def _consumer(key: str, meta, q: mp.Queue):
+    conn = cic.CudaIpcConnector({"stage_id": 1})
+    got, size = conn.get("stage0", "stage1", key, meta)
+    q.put(("got", got["codes"]["audio"].cpu(), got["hidden"].cpu(), got["meta"], got["cpu"], conn.health()["rebuilt"]))
+
+
+@pytest.mark.core_model
+@hardware_test(res={"cuda": "L4"}, num_cards=1)
+def test_cuda_ipc_round_trip_between_processes():
+    ctx = mp.get_context("spawn")
+    key = f"cic_gpu_{uuid.uuid4().hex[:8]}"
+    q = ctx.Queue()
+    prod = ctx.Process(target=_producer, args=(key, q))
+    prod.start()
+    tag, ok, size, meta, t_ref, hs_ref = q.get(timeout=120)
+    assert tag == "put" and ok and meta.get("cuda_ipc") == 2
+    cons = ctx.Process(target=_consumer, args=(key, meta, q))
+    cons.start()
+    tag, codes, hidden, m, cpu_t, rebuilt = q.get(timeout=120)
+    assert tag == "got" and rebuilt == 2
+    assert (
+        torch.equal(codes, t_ref)
+        and torch.equal(hidden, hs_ref)
+        and m == {"n": 25}
+        and torch.equal(cpu_t, torch.ones(2))
+    )
+    tag, live = q.get(timeout=120)
+    assert tag == "done" and live == 0
+    prod.join(30)
+    cons.join(30)
+    assert not os.path.exists(cic._ack_path(key))
diff --git a/vllm_omni/deploy/qwen3_tts_cuda_ipc.yaml b/vllm_omni/deploy/qwen3_tts_cuda_ipc.yaml
new file mode 100644
index 000000000..07b459522
--- /dev/null
+++ b/vllm_omni/deploy/qwen3_tts_cuda_ipc.yaml
@@ -0,0 +1,20 @@
+# Qwen3-TTS with the CUDA-IPC connector between talker and code2wav
+# (co-located CUDA stages, same GPU). GPU tensor leaves in the stage payload
+# are shared through CUDA IPC instead of D2H + msgpack; CPU leaves and the
+# control-plane metadata still use shared memory. Measurement gate: adopt in
+# the edge profile only if the per-chunk hop latency improves (see
+# benchmarks/tts/edge_results/stream_latency and WP7 notes).
+base_config: qwen3_tts.yaml
+async_chunk: true
+connectors:
+  connector_of_shared_memory:
+    name: CudaIpcConnector
+    extra:
+      codec_streaming: true
+      connector_get_sleep_s: 0.01
+      connector_get_max_wait_first_chunk: 3000
+      connector_get_max_wait: 300
+      codec_chunk_frames: 25
+      codec_left_context_frames: 72
+      initial_codec_chunk_frames: 1
+      decode_batch_max_size: 4
diff --git a/vllm_omni/distributed/omni_connectors/connectors/cuda_ipc_connector.py b/vllm_omni/distributed/omni_connectors/connectors/cuda_ipc_connector.py
new file mode 100644
index 000000000..73b11701e
--- /dev/null
+++ b/vllm_omni/distributed/omni_connectors/connectors/cuda_ipc_connector.py
@@ -0,0 +1,227 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+"""Device-direct connector for co-located CUDA stages (single host, same GPU).
+
+Tensor leaves that live on a CUDA device are exported through CUDA IPC
+(``torch.multiprocessing.reductions.reduce_tensor``) instead of being copied
+to host memory and msgpack-serialized; everything else (CPU tensors, ids,
+metadata, msgspec structs) still travels through the wrapped
+:class:`SharedMemoryConnector`, so the control-plane contract (``put`` /
+``get`` / ``cleanup``, key-addressed lookup, metadata pass-through) is
+unchanged.
+
+Protocol per exported tensor:
+
+* producer records an inter-process CUDA event on the current stream, packs
+  ``(rebuild args, event handle, dtype, shape)`` into a marker dict and keeps
+  the source tensor alive in ``_exported[put_key]``;
+* consumer rebuilds the tensor, waits on the event, **clones** it into its own
+  allocation and writes an ack file; the producer releases the source on its
+  next ``put``/``cleanup`` sweep (or when the registry exceeds ``max_live``).
+
+Requirements: producer and consumer are different processes on the same GPU
+(``VLLM_WORKER_MULTIPROC_METHOD=spawn``). Tensors allocated inside CUDA-graph
+private pools cannot be exported; they are cloned into a plain allocation first.
+"""
+
+from __future__ import annotations
+
+import copy
+import os
+import pickle
+import time
+from collections import OrderedDict
+from typing import Any
+
+import torch
+from vllm.logger import init_logger
+
+from vllm_omni.distributed.omni_connectors.connectors.base import OmniConnectorBase
+from vllm_omni.distributed.omni_connectors.connectors.shm_connector import SharedMemoryConnector
+
+logger = init_logger(__name__)
+
+MARKER = "__cuda_ipc__"
+ACK_DIR = "/dev/shm"
+
+
+def _ack_path(put_key: str) -> str:
+    return os.path.join(ACK_DIR, f"cuda_ipc_ack_{put_key}")
+
+
+def is_marker(obj: Any) -> bool:
+    return isinstance(obj, dict) and obj.get(MARKER) == 1
+
+
+class _Exported:
+    __slots__ = ("tensor", "event", "t")
+
+    def __init__(self, tensor: torch.Tensor, event: Any):
+        self.tensor, self.event, self.t = tensor, event, time.time()
+
+
+def export_tensor(t: torch.Tensor) -> tuple[dict[str, Any], _Exported]:
+    """Export a CUDA tensor; returns (marker dict, keep-alive record)."""
+    from torch.multiprocessing.reductions import reduce_tensor
+
+    src = t.detach()
+    if not src.is_contiguous():
+        src = src.contiguous()
+    try:
+        _, args = reduce_tensor(src)
+    except RuntimeError:
+        # e.g. CUDA-graph private pool allocation: move to a plain allocation
+        src = src.clone()
+        _, args = reduce_tensor(src)
+    event = torch.cuda.Event(interprocess=True)
+    event.record(torch.cuda.current_stream(src.device))
+    marker = {
+        MARKER: 1,
+        "args": pickle.dumps(args, protocol=pickle.HIGHEST_PROTOCOL),
+        "event": event.ipc_handle(),
+        "device": int(src.device.index or 0),
+        "dtype": str(src.dtype),
+        "shape": list(src.shape),
+    }
+    return marker, _Exported(src, event)
+
+
+def rebuild_tensor(marker: dict[str, Any]) -> torch.Tensor:
+    """Rebuild an exported tensor in the consumer process and own a copy of it."""
+    from torch.multiprocessing.reductions import rebuild_cuda_tensor
+
+    args = pickle.loads(marker["args"])
+    device = torch.device("cuda", int(marker.get("device", 0)))
+    with torch.cuda.device(device):
+        tensor = rebuild_cuda_tensor(*args)
+        event = torch.cuda.Event.from_ipc_handle(device, marker["event"])
+        event.wait(torch.cuda.current_stream(device))
+        owned = tensor.clone()
+    return owned
+
+
+def _walk(obj: Any, fn) -> Any:
+    """Apply ``fn`` to tensor leaves / marker dicts inside dicts, lists, tuples and msgspec structs."""
+    if isinstance(obj, torch.Tensor) or is_marker(obj):
+        return fn(obj)
+    if isinstance(obj, dict):
+        return {k: _walk(v, fn) for k, v in obj.items()}
+    if isinstance(obj, list):
+        return [_walk(v, fn) for v in obj]
+    if isinstance(obj, tuple):
+        return tuple(_walk(v, fn) for v in obj)
+    fields = getattr(obj, "__struct_fields__", None)
+    if fields:
+        out = copy.copy(obj)
+        for name in fields:
+            setattr(out, name, _walk(getattr(obj, name), fn))
+        return out
+    return obj
+
+
+class CudaIpcConnector(OmniConnectorBase):
+    """CUDA-IPC fast path for GPU tensor leaves, shared memory for the rest."""
+
+    supports_raw_data = False
+
+    def __init__(self, config: dict[str, Any]):
+        self.config = config
+        self.stage_id = config.get("stage_id", -1)
+        extra = config.get("extra", {}) if isinstance(config.get("extra"), dict) else {}
+        self.max_live = int(extra.get("cuda_ipc_max_live", 256))
+        self.enabled = bool(extra.get("cuda_ipc_enabled", True)) and torch.cuda.is_available()
+        self._inner = SharedMemoryConnector(config)
+        self._exported: OrderedDict[str, list[_Exported]] = OrderedDict()
+        self._metrics = {"ipc_tensors": 0, "ipc_bytes": 0, "fallback_leaves": 0, "rebuilt": 0}
+
+    # ------------------------------------------------------------- producer
+    def _sweep(self) -> None:
+        for key in list(self._exported.keys()):
+            ack = _ack_path(key)
+            if os.path.exists(ack):
+                try:
+                    os.remove(ack)
+                except OSError:
+                    pass
+                self._exported.pop(key, None)
+        while len(self._exported) > self.max_live:
+            key, _ = self._exported.popitem(last=False)
+            logger.warning("CudaIpcConnector: dropping unacked export %s (registry > %d)", key, self.max_live)
+
+    def put(self, from_stage: str, to_stage: str, put_key: str, data: Any) -> tuple[bool, int, dict[str, Any] | None]:
+        self._sweep()
+        keep: list[_Exported] = []
+
+        def leaf(t: Any) -> Any:
+            if isinstance(t, torch.Tensor) and t.is_cuda and self.enabled and t.numel() > 0:
+                try:
+                    marker, rec = export_tensor(t)
+                except Exception as exc:  # any IPC failure -> host path for this leaf
+                    logger.debug("CudaIpcConnector: export failed (%s); falling back to shm", exc)
+                    self._metrics["fallback_leaves"] += 1
+                    return t
+                keep.append(rec)
+                self._metrics["ipc_tensors"] += 1
+                self._metrics["ipc_bytes"] += t.numel() * t.element_size()
+                return marker
+            return t
+
+        rewritten = _walk(data, leaf)
+        ok, size, meta = self._inner.put(from_stage, to_stage, put_key, rewritten)
+        if ok and keep:
+            self._exported[put_key] = keep
+            meta = dict(meta or {})
+            meta["cuda_ipc"] = len(keep)
+        return ok, size, meta
+
+    # ------------------------------------------------------------- consumer
+    def get(
+        self, from_stage: str, to_stage: str, get_key: str, metadata: dict[str, Any] | None = None
+    ) -> tuple[Any, int] | None:
+        result = self._inner.get(from_stage, to_stage, get_key, metadata)
+        if result is None:
+            return None
+        obj, size = result
+        rebuilt = 0
+
+        def leaf(m: Any) -> Any:
+            nonlocal rebuilt
+            if is_marker(m):
+                rebuilt += 1
+                return rebuild_tensor(m)
+            return m
+
+        obj = _walk(obj, leaf)
+        if rebuilt:
+            self._metrics["rebuilt"] += rebuilt
+            try:
+                with open(_ack_path(get_key), "w") as f:
+                    f.write(str(time.time()))
+            except OSError:
+                pass
+        return obj, size
+
+    # ------------------------------------------------------------- lifecycle
+    def cleanup(self, request_id: str) -> None:
+        for key in list(self._exported.keys()):
+            if (
+                key == request_id
+                or key.startswith(request_id + "_")
+                or key.endswith("_" + request_id)
+                or request_id in key
+            ):
+                self._exported.pop(key, None)
+                try:
+                    os.remove(_ack_path(key))
+                except OSError:
+                    pass
+        self._inner.cleanup(request_id)
+
+    def health(self) -> dict[str, Any]:
+        h = self._inner.health()
+        h.update({"cuda_ipc": self.enabled, "live_exports": len(self._exported), **self._metrics})
+        return h
+
+    def close(self) -> None:
+        self._exported.clear()
+        self._inner.close()
diff --git a/vllm_omni/distributed/omni_connectors/factory.py b/vllm_omni/distributed/omni_connectors/factory.py
index c49da1c6c..4ccbfc3a8 100644
--- a/vllm_omni/distributed/omni_connectors/factory.py
+++ b/vllm_omni/distributed/omni_connectors/factory.py
@@ -126,6 +126,13 @@ def _create_mori_transfer_engine_connector(config: dict[str, Any]) -> OmniConnec
 
 
 # Register connectors
+def _create_cuda_ipc_connector(config: dict[str, Any]) -> OmniConnectorBase:
+    from .connectors.cuda_ipc_connector import CudaIpcConnector
+
+    return CudaIpcConnector(config)
+
+
+OmniConnectorFactory.register_connector("CudaIpcConnector", _create_cuda_ipc_connector)
 OmniConnectorFactory.register_connector("MooncakeStoreConnector", _create_mooncake_store_connector)
 OmniConnectorFactory.register_connector("MooncakeTransferEngineConnector", _create_mooncake_transfer_engine_connector)
 OmniConnectorFactory.register_connector("SharedMemoryConnector", _create_shm_connector)