ci: drop gha cache backend (runner cache server 404s on this instance)

Overseas node: original docker.io base (lmsysorg/sglang:dev-dsv41) and
default pypi. Tags follow the ymkymx convention <branch>-<sha9>-<datetime>
plus a moving <branch>-latest. Registry defaults to the Gitea instance's
container registry, override with the REGISTRY variable.

.gitea/workflows/build-image.yaml

@@ -0,0 +1,50 @@
+# Gitea Actions 自动构建 sglang 镜像（海外节点，原版源）
+# 基底 lmsysorg/sglang:dev-dsv41（docker.io），依赖走 pypi.org 默认源。
+# 触发：push 到 dsv41-pd 分支。
+#
+# 前置条件（一次性，在 Gitea 实例上配置）：
+# 1. 实例已注册 act_runner（Gitea Actions runner，标签含 ubuntu-latest）
+# 2. 仓库 Settings → Secrets 添加：
+#    REGISTRY_USERNAME / REGISTRY_PASSWORD（推送镜像的账号，如 Gitea 访问令牌）
+# 3. 可选：Settings → Actions → Variables 添加 REGISTRY（默认 git.agentwithu.com，
+#    即 Gitea 自带容器 registry；也可填 docker.io 等）
+name: build-sglang-image
+
+on:
+  push:
+    branches: [dsv41-pd]
+
+env:
+  # 直接写死 Gitea 自带 registry（vars context 在该实例上求值异常会导致回退 docker.io）
+  REGISTRY: git.agentwithu.com
+  IMAGE_NAME: minke.yu/sglang
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Resolve tag
+        id: meta
+        run: |
+          SHA9=$(git rev-parse --short=9 HEAD)
+          echo "tag=${{ github.ref_name }}-${SHA9}-$(date +%Y%m%d-%H%M)" >> "$GITHUB_OUTPUT"
+
+      - uses: docker/setup-buildx-action@v3
+
+      - uses: docker/login-action@v3
+        with:
+          registry: ${{ env.REGISTRY }}
+          username: ${{ secrets.REGISTRY_USERNAME }}
+          password: ${{ secrets.REGISTRY_PASSWORD }}
+
+      - uses: docker/build-push-action@v6
+        with:
+          context: .
+          file: Dockerfile.gitea
+          push: true
+          tags: |
+            ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}:${{ steps.meta.outputs.tag }}
+            ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}:${{ github.ref_name }}-latest
+          # 注意：type=gha 缓存在本实例的 runner 上会 404（cache server 未配），勿加回

Dockerfile.gitea

@@ -0,0 +1,11 @@
+# Gitea Actions 用（海外节点）：原版 docker.io 基底 + 原版 pypi
+# 与 b300 离线版（/data/ymk/build/Dockerfile）的区别：
+#   - 基底直用 docker.io 的 lmsysorg/sglang:dev-dsv41（不走 umirror/DaoCloud）
+#   - 不设 PIP_INDEX_URL，用默认 pypi.org
+#   - 源码由 CI checkout 后经 COPY 进镜像（不用 in-image clone）
+FROM lmsysorg/sglang:dev-dsv41
+
+# checkout 含 .git，setuptools-scm 可打戳（main 系分支 version 显示 dev 属正常）
+COPY . /sgl-workspace/sglang
+
+RUN pip install --no-cache-dir -e /sgl-workspace/sglang/python

python/sglang/srt/arg_groups/deepseek_v4_hook.py

@@ -249,23 +249,33 @@ def validate_deepseek_v41_features(server_args: ServerArgs) -> None:
         if (
             read_ragged_verify_mode() is not RaggedVerifyMode.STATIC
             or cfg.disaggregation_transfer_backend != "mooncake"
-            or cfg.dp_size != 1
-            or cfg.enable_dp_attention
             or cfg.attn_cp_size != 1
             or cfg.dcp_size != 1
         ):
             raise ValueError(
                 "DeepSeek-V4.1 DSpark PD requires static verify, Mooncake, "
-                "DP=1 and CP=1. Both servers must enable DSpark with the same "
-                "block size and TP size."
+                "and CP=1 on both servers. DP attention is supported when "
+                "both servers use the same block size and target/draft KV layout."
             )
 
     from sglang.srt.model_executor.cuda_graph_config import Backend, Phase, with_phase
 
     prefill_graph = cfg.cuda_graph_config.prefill
-    if prefill_graph.backend != Backend.DISABLED and prefill_graph.max_seq_len is None:
-        # The captured low-ratio indexer scores a static context width; 16k
-        # keeps it inside the candidate window at under 1 ms per layer.
+    cp_breakable_prefill = (
+        cfg.enable_prefill_cp
+        and cfg.cp_strategy == "interleave"
+        and cfg.tp_size > 1
+        and prefill_graph.backend == Backend.BREAKABLE
+    )
+    if (
+        prefill_graph.backend != Backend.DISABLED
+        and prefill_graph.max_seq_len is None
+        and not cp_breakable_prefill
+    ):
+        # The non-CP captured low-ratio indexer scores a static context width.
+        # CP BCG runs these sources eagerly with live prefix metadata, so this
+        # default would only force long-prefix CP batches back to eager.
+        # Explicit max_seq_len values still constrain both paths.
         declare_resolution(
             server_args,
             "validate_deepseek_v41_features",

python/sglang/srt/arg_groups/model_hook.py

@@ -966,12 +966,14 @@ def handle_language_model_only(server_args: Any):
     ):
         if flag:
             raise ValueError(f"--language-model-only cannot be combined with {name}")
-    if cfg.disaggregation_mode != "null":
+    hf_config = model_config_of(server_args).hf_config
+    # V4.1 text-only workers use the standard PD KV transfer path.
+    if cfg.disaggregation_mode != "null" and hf_config.model_type != "deepseek_v41":
         raise ValueError(
             "--language-model-only is incompatible with --disaggregation-mode "
             "prefill/decode"
         )
-    architectures = model_config_of(server_args).hf_config.architectures
+    architectures = hf_config.architectures
     if not any(
         a in server_args.LANGUAGE_MODEL_ONLY_ARCHITECTURES for a in architectures
     ):

python/sglang/srt/disaggregation/common/conn.py

@@ -945,9 +945,33 @@ class CommonKVManager(BaseKVManager):
                     "enable DSpark with the same block size and target/draft KV "
                     "layout. Upgrade both servers together."
                 )
-            if info.attn_tp_size != self.attn_tp_size:
+            same_tp_with_prefill_cp = (
+                info.attn_cp_size > 1
+                and (self.is_mla_backend or self.is_hybrid_mla_backend)
+                and self.attn_cp_size == 1
+                and info.attn_tp_size * info.attn_cp_size == self.attn_tp_size
+            )
+            # Combined branch (40323-series + 40177): prefill CP can also pair
+            # with a DP-attention decode server. MLA KV is replicated across
+            # prefill CP ranks, so per-rank layouts match when attn_tp matches.
+            dp_decode_with_prefill_cp = (
+                info.attn_cp_size > 1
+                and self.attn_cp_size == 1
+                and (self.is_mla_backend or self.is_hybrid_mla_backend)
+                and info.attn_tp_size == self.attn_tp_size
+            )
+            non_cp_mla_layout = info.attn_cp_size == self.attn_cp_size == 1 and (
+                self.is_mla_backend or self.is_hybrid_mla_backend
+            )
+            if info.attn_tp_size != self.attn_tp_size and not (
+                same_tp_with_prefill_cp
+                or dp_decode_with_prefill_cp
+                or non_cp_mla_layout
+            ):
                 raise RuntimeError(
-                    "DeepSeek-V4.1 DSpark PD requires the same TP size on both servers"
+                    "DeepSeek-V4.1 DSpark PD requires matching attention TP "
+                    "unless both servers use CP=1 with an MLA KV layout, "
+                    "or prefill runs CP with an MLA KV layout"
                 )
 
         if self.dcp_size > 1:

python/sglang/srt/disaggregation/mooncake/conn.py

@@ -82,6 +82,115 @@ FAILED_SESSION_RECOVERIES = Counter(
 )
 
 
+# ---------------------------------------------------------------------------
+# Intra-node NVLink transport helpers.
+#
+# Mooncake's IntraNodeNvlinkTransport can only register and reach *device*
+# memory (it IPC-opens the remote cudaMalloc segments). Host-resident regions
+# (aux buffers, some state components) cannot be registered: one host region
+# makes the whole registerLocalMemoryBatch fail, and the engine then rolls
+# back *every* region, leaving the segment descriptor empty and all KV
+# transfers failing with "Requested address ... not found". When the
+# intra-node NVLink transport is active we therefore
+#   1. register only device-memory regions, and
+#   2. route blocks whose source is host memory over the ordered zmq channel
+#      (same ordering guarantee the aux TCP path relies on) instead of the
+#      transfer engine.
+# ---------------------------------------------------------------------------
+
+
+import ctypes as _ctypes
+
+_CUDA_MEMORY_TYPE_DEVICE = 2
+
+try:
+    from cuda.bindings import runtime as _cudart
+except ImportError:  # pragma: no cover - cuda-python is always present in images
+    _cudart = None
+
+
+def _is_device_pointer(ptr: int) -> bool:
+    """Probe a *local* pointer with cudaPointerGetAttributes.
+
+    Only valid for pointers owned by this process (never probe remote
+    segment addresses). Returns False on any error so the caller falls back
+    to the safe host path.
+    """
+    if _cudart is None:
+        # Cannot tell; assume device so behavior stays unchanged.
+        return True
+    err, attr = _cudart.cudaPointerGetAttributes(int(ptr))
+    if int(err) != 0:
+        # Clear the error so subsequent CUDA calls are not poisoned.
+        _cudart.cudaGetLastError()
+        return False
+    return int(attr.type) == _CUDA_MEMORY_TYPE_DEVICE
+
+
+def _read_bytes_from_address(addr: int, length: int) -> Optional[bytes]:
+    if length <= 0:
+        return b""
+    if _is_device_pointer(addr):
+        buf = bytearray(length)
+        # cudaMemcpyDeviceToHost = 2; synchronous default-stream copy.
+        err, = _cudart.cudaMemcpy(
+            _ctypes.addressof((_ctypes.c_char * length).from_buffer(buf)),
+            int(addr),
+            length,
+            2,
+        )
+        if int(err) != 0:
+            logger.error(
+                f"cudaMemcpy D2H failed (err={err}) for addr {hex(addr)} len {length}"
+            )
+            return None
+        return bytes(buf)
+    return _ctypes.string_at(int(addr), length)
+
+
+def _write_bytes_to_address(addr: int, data: bytes) -> bool:
+    if not data:
+        return True
+    if _is_device_pointer(addr):
+        buf = _ctypes.create_string_buffer(data, len(data))
+        # cudaMemcpyHostToDevice = 1; synchronous default-stream copy.
+        err, = _cudart.cudaMemcpy(
+            int(addr), _ctypes.addressof(buf), len(data), 1
+        )
+        if int(err) != 0:
+            logger.error(
+                f"cudaMemcpy H2D failed (err={err}) for addr {hex(addr)} "
+                f"len {len(data)}"
+            )
+            return False
+        return True
+    _ctypes.memmove(int(addr), data, len(data))
+    return True
+
+
+_NVLINK_INTRA_ACTIVE = None
+
+
+def _nvlink_intra_transport_active() -> bool:
+    """Whether mooncake installed the intra-node NVLink transport.
+
+    Mirrors the env probing in mooncake's transfer_engine_impl.cpp: the
+    transport is installed iff MC_INTRANODE_NVLINK is set (any value), or an
+    equivalent protocol selection was made.
+    """
+    global _NVLINK_INTRA_ACTIVE
+    if _NVLINK_INTRA_ACTIVE is None:
+        active = bool(
+            os.environ.get("MC_INTRANODE_NVLINK")
+            or os.environ.get("MC_INTRA_NVLINK")
+        )
+        if not active:
+            proto = (os.environ.get("MOONCAKE_PROTOCOL") or "").strip().lower()
+            active = proto in ("nvlink_intra", "nvlink-intra", "intra_nvlink")
+        _NVLINK_INTRA_ACTIVE = active
+    return _NVLINK_INTRA_ACTIVE
+
+
 # decode
 @dataclasses.dataclass
 class TransferInfo:
@@ -214,6 +323,7 @@ class KVArgsRegisterInfo:
 
 class MooncakeKVManager(StagingManagerMixin, CommonKVManager):
     AUX_DATA_HEADER = b"AUX_DATA"
+    STATE_DATA_HEADER = b"STATE_DATA"
     # Implements teardown() below, so runtime PD role switching is supported.
     supports_role_switch = True
 
@@ -227,6 +337,10 @@ class MooncakeKVManager(StagingManagerMixin, CommonKVManager):
         super().__init__(args, disaggregation_mode, server_args, is_mla_backend)
         self.init_engine()
         self.register_buffer_to_engine()
+        # session_id -> (endpoint, dst_port, room), used to route host-memory
+        # transfer blocks over zmq when the intra-node NVLink transport is
+        # active (it cannot reach host memory). Populated on bootstrap.
+        self._session_endpoint_map = {}
         self.enable_staging = envs.SGLANG_DISAGG_STAGING_BUFFER.get()
         self.max_transfer_batch_indices = (
             envs.SGLANG_MOONCAKE_MAX_TRANSFER_BATCH_INDICES.get()
@@ -322,6 +436,13 @@ class MooncakeKVManager(StagingManagerMixin, CommonKVManager):
         Deduped because the unified memory pool reports one raw buffer as both
         its KV and its mamba state component, and double registration fails in
         the engine.
+
+        When the intra-node NVLink transport is active, host-memory regions
+        (aux buffers, some state components) are skipped: the transport only
+        accepts device memory, and a single host region fails the whole batch
+        and triggers a full engine-side rollback that would unregister the KV
+        pools too. Host-resident payloads are instead exchanged over the
+        ordered zmq channel (see _transfer_data / send_aux).
         """
         regions: List[Tuple[int, int]] = []
         seen: Set[Tuple[int, int]] = set()
@@ -338,6 +459,24 @@ class MooncakeKVManager(StagingManagerMixin, CommonKVManager):
             self.kv_args.state_data_ptrs, self.kv_args.state_data_lens
         ):
             add(ptrs, lens)
+
+        if _nvlink_intra_transport_active():
+            device_regions = []
+            skipped = []
+            for ptr, length in regions:
+                if _is_device_pointer(ptr):
+                    device_regions.append((ptr, length))
+                else:
+                    skipped.append((ptr, length))
+            if skipped:
+                logger.info(
+                    "Intra-node NVLink transport: skipping %d host-memory "
+                    "regions from engine registration (they will be exchanged "
+                    "over the zmq channel instead): %s",
+                    len(skipped),
+                    [(hex(p), l) for p, l in skipped[:8]],
+                )
+            regions = device_regions
         return regions
 
     def register_buffer_to_engine(self):
@@ -748,10 +887,63 @@ class MooncakeKVManager(StagingManagerMixin, CommonKVManager):
         if not transfer_blocks:
             return 0
 
-        src_addrs, dst_addrs, lengths = zip(*transfer_blocks)
-        return self.engine.batch_transfer_sync(
-            mooncake_session_id, list(src_addrs), list(dst_addrs), list(lengths)
-        )
+        if not _nvlink_intra_transport_active():
+            src_addrs, dst_addrs, lengths = zip(*transfer_blocks)
+            return self.engine.batch_transfer_sync(
+                mooncake_session_id, list(src_addrs), list(dst_addrs), list(lengths)
+            )
+
+        # Intra-node NVLink transport can only move device memory. Partition
+        # blocks by the *local source* pointer (probing a local pointer is
+        # safe; the remote dst is never probed): device-sourced blocks go
+        # through the engine as usual, host-sourced blocks are shipped over
+        # the ordered zmq channel and written into the peer's buffer by the
+        # receiver (see _handle_state_data). This mirrors the aux TCP path.
+        device_blocks = []
+        host_blocks = []
+        for src, dst, length in transfer_blocks:
+            if _is_device_pointer(src):
+                device_blocks.append((src, dst, length))
+            else:
+                host_blocks.append((src, dst, length))
+
+        rc = 0
+        if device_blocks:
+            src_addrs, dst_addrs, lengths = zip(*device_blocks)
+            rc = self.engine.batch_transfer_sync(
+                mooncake_session_id, list(src_addrs), list(dst_addrs), list(lengths)
+            )
+        if rc == 0 and host_blocks:
+            rc = self._send_host_blocks_tcp(mooncake_session_id, host_blocks)
+        return rc
+
+    def _send_host_blocks_tcp(self, mooncake_session_id, host_blocks):
+        target = self._session_endpoint_map.get(mooncake_session_id)
+        if target is None:
+            logger.error(
+                f"No zmq endpoint known for mooncake session "
+                f"{mooncake_session_id}; cannot deliver {len(host_blocks)} "
+                "host-memory transfer blocks"
+            )
+            return -1
+        endpoint, dst_port, room = target
+        na = NetworkAddress(endpoint, dst_port)
+        for src, dst, length in host_blocks:
+            data = _read_bytes_from_address(src, length)
+            if data is None:
+                return -1
+            self._send_multipart_locked(
+                na.to_tcp(),
+                [
+                    MooncakeKVManager.STATE_DATA_HEADER,
+                    str(room).encode("ascii"),
+                    str(int(dst)).encode("ascii"),
+                    struct.pack(">I", len(data)),
+                    data,
+                ],
+                is_ipv6=na.is_ipv6,
+            )
+        return 0
 
     def _send_kvcache_generic(
         self,
@@ -1482,8 +1674,10 @@ class MooncakeKVManager(StagingManagerMixin, CommonKVManager):
     ):
         # TODO(shangming): Fix me when nvlink_transport of Mooncake is bug-free
         if (
-            self.enable_custom_mem_pool and self.custom_mem_pool_type == "NVLINK"
-        ) or envs.SGLANG_MOONCAKE_SEND_AUX_TCP.get():
+            (self.enable_custom_mem_pool and self.custom_mem_pool_type == "NVLINK")
+            or envs.SGLANG_MOONCAKE_SEND_AUX_TCP.get()
+            or _nvlink_intra_transport_active()
+        ):
             return self.send_aux_tcp(req, prefill_aux_index, dst_aux_ptrs)
 
         transfer_blocks = []
@@ -1566,6 +1760,71 @@ class MooncakeKVManager(StagingManagerMixin, CommonKVManager):
             f"Received AUX_DATA for bootstrap_room {room} with length:{len(data)}"
         )
 
+    def _host_transfer_regions(self):
+        """Address ranges this process published as transfer targets.
+
+        Used to validate STATE_DATA writes. Built lazily because kv_args is
+        fully populated only after registration.
+        """
+        regions = getattr(self, "_host_transfer_regions_cache", None)
+        if regions is None:
+            regions = []
+            for ptr, length in zip(
+                self.kv_args.kv_data_ptrs or [], self.kv_args.kv_data_lens or []
+            ):
+                regions.append((int(ptr), int(ptr) + int(length)))
+            for ptr, length in zip(
+                self.kv_args.aux_data_ptrs or [], self.kv_args.aux_data_lens or []
+            ):
+                regions.append((int(ptr), int(ptr) + int(length)))
+            for ptrs, lens in zip(
+                self.kv_args.state_data_ptrs or [], self.kv_args.state_data_lens or []
+            ):
+                for ptr, length in zip(ptrs or [], lens or []):
+                    regions.append((int(ptr), int(ptr) + int(length)))
+            self._host_transfer_regions_cache = regions
+        return regions
+
+    def _handle_state_data(self, msg: List[bytes]):
+        """Handle STATE_DATA messages received by the decode thread.
+
+        Carries one host-memory transfer block that could not go through the
+        intra-node NVLink transport. Written directly into the local buffer at
+        the destination address; ordering against the final status message is
+        guaranteed by the shared per-endpoint zmq socket.
+        """
+        room = int(msg[1].decode("ascii"))
+        dst_addr = int(msg[2].decode("ascii"))
+        data_length = struct.unpack(">I", msg[3])[0]
+        data = msg[4]
+
+        if len(data) != data_length:
+            logger.error(f"STATE_DATA length mismatch for bootstrap_room {room}")
+            return
+
+        in_region = any(
+            start <= dst_addr and dst_addr + len(data) <= end
+            for start, end in self._host_transfer_regions()
+        )
+        if not in_region:
+            logger.error(
+                f"STATE_DATA for bootstrap_room {room} targets unknown region "
+                f"{hex(dst_addr)}..{hex(dst_addr + len(data))}; dropping"
+            )
+            return
+
+        if not _write_bytes_to_address(dst_addr, data):
+            logger.error(
+                f"STATE_DATA write failed for bootstrap_room {room} at "
+                f"{hex(dst_addr)} len {len(data)}"
+            )
+            return
+
+        logger.debug(
+            f"Received STATE_DATA for bootstrap_room {room} at {hex(dst_addr)} "
+            f"with length:{len(data)}"
+        )
+
     def _get_dsa_cache_transfer_skip_flags(
         self, info: Optional[KVArgsRegisterInfo]
     ) -> Tuple[bool, bool]:
@@ -2412,6 +2671,8 @@ class MooncakeKVManager(StagingManagerMixin, CommonKVManager):
                 ):
                     self._staging_outstanding.pop(kv_chunk.room, None)
                     if kv_chunk.room in self.transfer_infos:
+                        for sid in self.transfer_infos[kv_chunk.room]:
+                            self._session_endpoint_map.pop(sid, None)
                         self.transfer_infos.pop(kv_chunk.room)
                     self.req_to_decode_prefix_len.pop(kv_chunk.room, None)
                     if self.enable_staging:
@@ -2557,6 +2818,11 @@ class MooncakeKVManager(StagingManagerMixin, CommonKVManager):
                     self.transfer_infos[room][mooncake_session_id] = (
                         TransferInfo.from_zmq(waiting_req_bytes)
                     )
+                    self._session_endpoint_map[mooncake_session_id] = (
+                        self.transfer_infos[room][mooncake_session_id].endpoint,
+                        self.transfer_infos[room][mooncake_session_id].dst_port,
+                        room,
+                    )
                     # NOTE: after bootstrapping we can mark the req as waiting for input
                     if len(self.transfer_infos[room]) == required_dst_info_num:
                         self.resolve_kv_replica_factor(self.transfer_infos[room])
@@ -2585,6 +2851,9 @@ class MooncakeKVManager(StagingManagerMixin, CommonKVManager):
                 if msg[0] == MooncakeKVManager.AUX_DATA_HEADER:
                     self._handle_aux_data(msg)
                     continue
+                if msg[0] == MooncakeKVManager.STATE_DATA_HEADER:
+                    self._handle_state_data(msg)
+                    continue
 
                 # Staging: prefill notifies a chunk written to staging buffer
                 if msg[0] == b"CHUNK_READY":

python/sglang/srt/layers/attention/deepseek_v4_backend.py

@@ -736,7 +736,14 @@ class DSV4AttnMetadata:
             if src_val is None and dst_val is None:
                 continue
             assert dst_val is not None, f"{field_name=} {src_val=} {dst_val=}"
-            dst_val.copy_(src_val)
+            shape_mismatch = dst_val.shape != src_val.shape
+            assert not shape_mismatch or field_name in self._CP_GLOBAL_FIELDS, (
+                f"Only CP-global replay metadata may use a shorter live prefix, "
+                f"got {field_name=} {src_val.shape=} {dst_val.shape=}"
+            )
+            _copy_tensor_allowing_storage_alias(
+                dst_val, src_val, pad_value=0 if shape_mismatch else None
+            )
 
         # These fields are safe to replace because captured kernels only need
         # the current per-replay objects, or the field is produced inside the
@@ -988,6 +995,27 @@ def _prefill_graph_max_seq_len() -> Optional[int]:
     return get_exec().graph.cuda_graph_config.prefill.max_seq_len
 
 
+def _copy_tensor_allowing_storage_alias(
+    dst: torch.Tensor, src: torch.Tensor, *, pad_value: Optional[int] = None
+) -> None:
+    """Copy replay metadata while preserving capture-stable destination addresses."""
+    if dst is src:
+        return
+    if dst.untyped_storage().data_ptr() == src.untyped_storage().data_ptr():
+        src = src.clone()
+    if dst.shape == src.shape:
+        dst.copy_(src)
+        return
+    assert (
+        pad_value is not None
+        and dst.ndim == src.ndim
+        and dst.shape[0] >= src.shape[0]
+        and dst.shape[1:] == src.shape[1:]
+    ), f"Cannot copy replay metadata from {src.shape=} to {dst.shape=}"
+    dst.fill_(pad_value)
+    dst[: src.shape[0]].copy_(src)
+
+
 @dataclass
 class DSV4Metadata:
     core_attn_metadata: DSV4AttnMetadata
@@ -1254,6 +1282,11 @@ class DeepseekV4AttnBackend(
         ] = None
         self.online_c128_mtp = OnlineC128MTPController(self)
         self.sparse_prefill_workspace = SparsePrefillWorkspace(self.device)
+        # CP V4.1 consumers share compressed KV across layers. Separate ratio
+        # workspaces keep those prefixes intact while each layer refreshes SWA.
+        self.shared_compressed_prefill_workspaces = {
+            ratio: SparsePrefillWorkspace(self.device) for ratio in (1, 2)
+        }
         spec_alg = model_runner.spec_algorithm
         self.needs_cpu_seq_lens = not spec_alg.is_dspark() and (
             not _is_cuda or self.online_c128_mtp.enabled()
@@ -1566,8 +1599,12 @@ class DeepseekV4AttnBackend(
 
     @property
     def low_ratio_prefill_graph(self) -> bool:
+        """Whether ratio-1/2 sources use captured projections and indexer metadata."""
         return (
-            bool(self.low_ratios) and _has_dense_fp4_indexer() and _is_sm100_or_newer()
+            bool(self.low_ratios)
+            and _has_dense_fp4_indexer()
+            and _is_sm100_or_newer()
+            and get_parallel().attn_cp_size == 1
         )
 
     def can_run_prefill_cuda_graph(self, forward_batch: ForwardBatch) -> bool:
@@ -2870,7 +2907,7 @@ class DeepseekV4AttnBackend(
                 q_lora[:num_local],
                 positions[:num_local].to(torch.int64),
                 forward_batch,
-                torch.tensor(q_lens_cpu, dtype=torch.int32, device=x.device),
+                self._move_to_device(q_lens_cpu),
                 q_lens_cpu,
             )
 
@@ -3236,7 +3273,9 @@ class DeepseekV4AttnBackend(
                 continue
             j = torch.arange(lc, device=device)
             slot_chunks.append(
-                self.req_to_token[req_pool_indices[r], j * ratio].to(torch.int64)
+                self.req_to_token[req_pool_indices[r : r + 1], j * ratio].to(
+                    torch.int64
+                )
                 // ratio
             )
             start += lc
@@ -3261,7 +3300,7 @@ class DeepseekV4AttnBackend(
         weights = indexer.head_weights(x).float()
         compress_lens = ((pos + 1) // ratio).to(torch.int32)
         ks = torch.repeat_interleave(
-            torch.tensor(starts, dtype=torch.int32, device=device),
+            self._move_to_device(starts),
             q_lens.to(torch.int64),
             output_size=num_tokens,
         )
@@ -3945,20 +3984,38 @@ class DeepseekV4AttnBackend(
                 compress_ratio, core_attn_metadata, extra_page_size
             )
             n_compressed = flat_token_ids.shape[0]
-            workspace = self.sparse_prefill_workspace.get(
-                n_compressed + cache.swa_token_ids.shape[0]
+            reuse_compressed = compress_ratio in (1, 2) and is_cp_active(forward_batch)
+            workspace_pool = (
+                self.shared_compressed_prefill_workspaces[compress_ratio]
+                if reuse_compressed
+                else self.sparse_prefill_workspace
             )
+            workspace = workspace_pool.get(n_compressed + cache.swa_token_ids.shape[0])
             compressed_slice = workspace[:n_compressed]
             swa_slice = workspace[n_compressed:]
 
         if compressed_slice is not None:
-            dequantize_k_cache_paged(
-                extra_k_cache,
-                flat_token_ids,
-                page_size=extra_page_size,
-                out=compressed_slice,
-                layout=token_to_kv_pool.get_extra_key_layout(layer_id),
-            )
+            source_key = None
+            if reuse_compressed:
+                source_layer = token_to_kv_pool.source_layer_of(layer_id)
+                source_key = (source_layer, workspace.data_ptr())
+                gather = cache.compressed[compress_ratio]
+            # A source layer may have just updated its cache in place. Consumer
+            # layers only reuse the compressed prefix; their top-k and SWA stay live.
+            if (
+                source_key is None
+                or layer_id == source_key[0]
+                or gather.dequantized_source != source_key
+            ):
+                dequantize_k_cache_paged(
+                    extra_k_cache,
+                    flat_token_ids,
+                    page_size=extra_page_size,
+                    out=compressed_slice,
+                    layout=token_to_kv_pool.get_extra_key_layout(layer_id),
+                )
+                if source_key is not None:
+                    gather.dequantized_source = source_key
         dequantize_k_cache_paged(
             token_to_kv_pool.get_swa_key_buffer_radix(layer_id),
             cache.swa_token_ids,

python/sglang/srt/layers/attention/dsv4/sparse_prefill_utils.py

@@ -78,10 +78,11 @@ def use_dsv4_q8kv8_sparse_prefill(dsv4_prefill_backend: str = "auto") -> bool:
 class SparsePrefillWorkspace:
     """Backend-owned scratch storage for sparse prefill KV dequantization.
 
-    The workspace contents are fully overwritten before every attention call,
-    so token buckets and compression ratios can safely share one buffer. Sparse
-    prefill executes eagerly and serially on the supported paths, which makes it
-    safe to replace the scratch allocation when a larger extent is needed.
+    Callers normally overwrite the entire workspace. Shared compressed-KV
+    callers keep separate workspaces per ratio and track prefix validity in the
+    per-forward gather cache, including the allocation address. Sparse prefill
+    executes eagerly and serially on the supported paths, so the allocation can
+    be replaced when a larger extent is needed.
     """
 
     def __init__(self, device: torch.device):
@@ -275,14 +276,18 @@ class CompressedGather:
     # chunk-invariant per request; subsequent layers only overwrite that prefix.
     combined_indices: Optional[torch.Tensor] = None
     combined_lens: Optional[torch.Tensor] = None
+    # Valid only for this forward's gather layout. Each ratio has its own
+    # workspace; its compressed prefix survives consumer layers' SWA writes.
+    dequantized_source: Optional[tuple[int, int]] = None
 
 
 @dataclass
 class SparsePrefillChunkCache:
     """Cache prefill-chunk metadata shared across layers.
 
-    Fields depend on request/token mappings and compressed page tables, not
-    per-layer k_cache; per-layer top-k combinations are recomputed into reused
+    Gather layouts depend on request/token mappings and compressed page tables.
+    Shared-source dequantization keys live only for this forward; per-layer
+    top-k combinations are recomputed into reused
     buffers.
     """
 

python/sglang/srt/layers/cp/bcg.py

@@ -23,17 +23,22 @@ import torch
 
 from sglang.srt.arg_groups.overrides import (
     attention_backends_of,
+    model_config_of,
     resolved_view,
     resolving_view,
 )
+from sglang.srt.configs.model_config import is_deepseek_v4
 from sglang.srt.layers.cp.base import get_cp_strategy
+from sglang.srt.layers.cp.interleave import InterleaveCPStrategy
 from sglang.srt.layers.cp.padding import get_cp_padding_align_size
 from sglang.srt.layers.cp.utils import (
     cp_gather_after_forward,
+    cp_shard_hidden_states,
     cp_split_before_forward,
     prepare_cp_forward,
 )
 from sglang.srt.layers.cp.zigzag import ZigzagCPStrategy
+from sglang.srt.layers.logits_processor import LogitsMetadata
 from sglang.srt.model_executor.forward_batch_info import PPProxyTensors
 
 if TYPE_CHECKING:
@@ -50,12 +55,18 @@ def supports_prefill_cp_bcg(server_args: ServerArgs) -> bool:
     cfg = resolving_view(server_args)
     resolved = resolved_view(server_args)
     prefill_attention_backend, _ = attention_backends_of(resolved_view(server_args))
+    supports_layout = (
+        cfg.cp_strategy == "zigzag" and prefill_attention_backend == "trtllm_mha"
+    ) or (
+        cfg.cp_strategy == "interleave"
+        and prefill_attention_backend == "dsv4"
+        and is_deepseek_v4(model_config_of(server_args).hf_config)
+    )
     return (
         cfg.enable_prefill_cp
         and cfg.pp_size == 1
         and resolved.attn_cp_size == cfg.tp_size
-        and cfg.cp_strategy == "zigzag"
-        and prefill_attention_backend == "trtllm_mha"
+        and supports_layout
     )
 
 
@@ -67,8 +78,12 @@ def enable_cp_bcg_capture(server_args: ServerArgs) -> bool:
 def filter_prefill_cp_bcg_capture_num_tokens(
     capture_num_tokens: list[int], server_args: ServerArgs
 ) -> list[int]:
-    """Keep only token buckets where the zigzag CP strategy can run."""
-    min_num_tokens = resolved_view(server_args).attn_cp_size * 2
+    """Keep only token buckets where the configured CP strategy can run."""
+    cfg = resolving_view(server_args)
+    cp_segments_per_token_block = 2 if cfg.cp_strategy == "zigzag" else 1
+    min_num_tokens = (
+        resolved_view(server_args).attn_cp_size * cp_segments_per_token_block
+    )
     filtered = [size for size in capture_num_tokens if size >= min_num_tokens]
     if not filtered:
         raise ValueError(
@@ -96,6 +111,8 @@ class PrefillCPBCGInput:
 
     input_embeds: torch.Tensor
     positions: torch.Tensor
+    input_ids: Optional[torch.Tensor] = None
+    num_token_non_padded: Optional[torch.Tensor] = None
     bucket_local_tokens: Dict[int, int] = field(default_factory=dict)
     live_local_tokens: int = 0
 
@@ -114,12 +131,22 @@ class PrefillCPBCGInput:
                     (runner.max_num_tokens,),
                     dtype=torch.int64,
                 ),
+                input_ids=torch.zeros((runner.max_num_tokens,), dtype=torch.int64),
+                num_token_non_padded=torch.zeros((), dtype=torch.int32),
             )
 
     def required_local_tokens(self, extend_seq_lens: Any) -> Optional[int]:
-        """Return the aligned CP-local rows required by a live zigzag layout."""
+        """Return the aligned CP-local rows required by the active layout."""
         strategy = get_cp_strategy()
-        if not isinstance(strategy, ZigzagCPStrategy) or extend_seq_lens is None:
+        if extend_seq_lens is None:
+            return None
+        if isinstance(strategy, InterleaveCPStrategy):
+            logical_tokens = (
+                sum(int(length) for length in extend_seq_lens) + strategy.cp_size - 1
+            ) // strategy.cp_size
+            align_size = get_cp_padding_align_size()
+            return (logical_tokens + align_size - 1) // align_size * align_size
+        if not isinstance(strategy, ZigzagCPStrategy):
             return None
 
         cp_segment_num = strategy.cp_size * 2
@@ -219,6 +246,7 @@ class PrefillCPBCGInput:
         raw_tokens = int(forward_batch.extend_num_tokens)
         global_input_ids = forward_batch.input_ids[:raw_tokens]
         global_positions = forward_batch.positions[:raw_tokens]
+        local_input_ids = cp_shard_hidden_states(global_input_ids, forward_batch)
         global_input_embeds = runner.model_runner.model.get_input_embeddings()(
             global_input_ids
         )
@@ -249,12 +277,31 @@ class PrefillCPBCGInput:
 
         input_embeds = self.input_embeds[:captured_local_tokens]
         positions = self.positions[:captured_local_tokens]
+        assert self.input_ids is not None
+        input_ids = self.input_ids[:captured_local_tokens]
         input_embeds.zero_()
         positions.zero_()
+        input_ids.zero_()
         input_embeds[:live_local_tokens].copy_(local_input_embeds)
         positions[:live_local_tokens].copy_(local_positions)
+        input_ids[:live_local_tokens].copy_(local_input_ids)
         forward_batch.input_embeds = input_embeds
-        forward_batch.positions = positions
+        forward_batch._cp_positions = positions
+        # Keep the global input_ids field intact: the runner uses its length to
+        # select the global capture bucket. The DSV4 body consumes this fixed,
+        # rank-local view for hash routing and MegaMoE.
+        forward_batch._cp_input_ids = input_ids
+        forward_batch.input_ids_global = input_ids
+        if forward_batch.num_token_non_padded is not None:
+            assert self.num_token_non_padded is not None
+            metadata = forward_batch.attn_cp_metadata
+            logical_tokens = (
+                metadata.per_rank_logical_token or metadata.per_rank_actual_token
+            )
+            strategy = get_cp_strategy()
+            assert strategy is not None
+            self.num_token_non_padded.fill_(logical_tokens[strategy.cp_rank])
+            forward_batch.num_token_non_padded = self.num_token_non_padded
         self.live_local_tokens = live_local_tokens
 
 
@@ -307,10 +354,50 @@ def execute_prefill_cp_bcg(
             static_forward_batch,
             torch.cuda.current_stream(),
         )
-        return model.logits_processor(
-            forward_batch.input_ids,
+        if aux_hidden_states is not None:
+            if torch.is_tensor(aux_hidden_states):
+                aux_hidden_states = cp_gather_after_forward(
+                    aux_hidden_states, static_forward_batch, torch.cuda.current_stream()
+                )
+            else:
+                aux_hidden_states = [
+                    cp_gather_after_forward(
+                        aux, static_forward_batch, torch.cuda.current_stream()
+                    )
+                    for aux in aux_hidden_states
+                ]
+        hidden_states_before_norm = None
+        if isinstance(hidden_states, tuple):
+            assert len(hidden_states) == 2
+            hidden_states, hidden_states_before_norm = hidden_states
+
+        input_ids = forward_batch.input_ids
+        logits_metadata = forward_batch
+        tail = None
+        language_model = getattr(model, "model", None)
+        if (
+            capture_aux_hidden_states
+            and getattr(language_model, "late_layer_start", None) is not None
+            and forward_batch.forward_mode.is_extend_without_speculative()
+        ):
+            tail_metadata = runner.model_runner.attn_backend.tail_forward_metadata
+            tail = tail_metadata.late_layer_tail
+            input_ids = tail.rows(input_ids)
+            logits_metadata = LogitsMetadata.from_forward_batch(forward_batch)
+            logits_metadata.extend_seq_lens = tail.extend_seq_lens
+            logits_metadata.extend_seq_lens_cpu = tail.extend_seq_lens_cpu
+            logits_metadata.extend_logprob_start_lens_cpu = tail.extend_seq_lens_cpu
+
+        output = model.logits_processor(
+            input_ids,
             hidden_states,
             model.lm_head,
-            forward_batch,
+            logits_metadata,
             aux_hidden_states,
+            hidden_states_before_norm=(
+                None if aux_hidden_states is not None else hidden_states_before_norm
+            ),
         )
+        if tail is not None:
+            output.hidden_states_token_indices = tail.token_indices
+        return output

python/sglang/srt/layers/moe/mega_moe.py

@@ -216,20 +216,33 @@ def _run_mega_routed(
 
     if num_tokens > 0:
         router_logits = moe.gate(hidden_states, forward_batch=forward_batch)
-        topk_kwargs = {"input_ids": input_ids_global} if moe.is_hash else {}
-        topk_output = moe.topk(
-            hidden_states,
-            router_logits,
-            num_token_non_padded=(
-                forward_batch.num_token_non_padded
-                if forward_batch is not None
-                else None
-            ),
-            expert_location_dispatch_info=ExpertLocationDispatchInfo.init_new(
-                layer_id=moe.layer_id,
-            ),
-            **topk_kwargs,
+        num_token_non_padded = (
+            forward_batch.num_token_non_padded if forward_batch is not None else None
         )
+        if isinstance(
+            getattr(moe.gate, "e_score_correction_bias_vl", None), torch.Tensor
+        ):
+            # V4.1 uses a different correction bias for image-token rows. The
+            # MegaMoE transport consumes the same routed ids/weights as TopK.
+            from sglang.srt.multimodal.dsv41.vl_routing import vision_topk
+
+            topk_output = vision_topk(
+                moe,
+                router_logits,
+                input_ids_global,
+                num_token_non_padded=num_token_non_padded,
+            )
+        else:
+            topk_kwargs = {"input_ids": input_ids_global} if moe.is_hash else {}
+            topk_output = moe.topk(
+                hidden_states,
+                router_logits,
+                num_token_non_padded=num_token_non_padded,
+                expert_location_dispatch_info=ExpertLocationDispatchInfo.init_new(
+                    layer_id=moe.layer_id,
+                ),
+                **topk_kwargs,
+            )
         topk_ids = topk_output.topk_ids
         topk_weights = topk_output.topk_weights
     else:

python/sglang/srt/mem_cache/common.py

@@ -155,6 +155,10 @@ def free_kv_row_segments(
 
 def maybe_cache_unfinished_req(req: Req, tree_cache: BasePrefixCache, **kwargs):
     if getattr(req, "skip_radix_cache_insert", False):
+        kv_indices = tree_cache.req_to_token_pool.req_to_token[
+            req.kv.req_pool_idx, : len(req.get_fill_ids())
+        ]
+        req.prefix_indices = kv_indices.to(dtype=torch.int64, copy=True)
         return
 
     tree_cache.cache_unfinished_req(req, **kwargs)

python/sglang/srt/model_executor/runner/eager_runner.py

@@ -385,14 +385,22 @@ class EagerRunner(BaseRunner):
         """
         model = self.model_runner.model
 
+        input_ids = forward_batch.input_ids
         input_embeds = kwargs.get("input_embeds")
+        # Multimodal spans must be embedded in global token order, before CP
+        # slicing. The model may also normalize image hash IDs for its router.
+        prepare_inputs = getattr(model, "prepare_language_model_inputs", None)
+        if prepare_inputs is not None:
+            input_ids, input_embeds = prepare_inputs(
+                input_ids, forward_batch, input_embeds
+            )
         if input_embeds is None:
-            input_embeds = model.get_input_embeddings()(forward_batch.input_ids)
+            input_embeds = model.get_input_embeddings()(input_ids)
         with cp_shard_model_inputs(
             input_embeds,
             forward_batch.positions,
             forward_batch,
-            forward_batch.input_ids,
+            input_ids,
         ) as (sharded_input_embeds, sharded_positions, model_input_ids):
             model_kwargs = {"input_embeds": sharded_input_embeds}
             if (pp_proxy_tensors := kwargs.get("pp_proxy_tensors")) is not None:
@@ -437,7 +445,7 @@ class EagerRunner(BaseRunner):
             if aux_hidden_states is None:
                 logits_kwargs["hidden_states_before_norm"] = hidden_states_before_norm
         return model.logits_processor(
-            forward_batch.input_ids,
+            input_ids,
             hidden_states,
             model.lm_head,
             forward_batch,

python/sglang/srt/model_executor/runner/prefill_cuda_graph_runner.py

@@ -701,6 +701,9 @@ class PrefillCudaGraphRunner(BaseCudaGraphRunner):
 
     def _get_layer_model_positions(self, forward_batch: ForwardBatch) -> torch.Tensor:
         """Mirror outer multimodal wrappers when BCG captures layer_model directly."""
+        cp_positions = getattr(forward_batch, "_cp_positions", None)
+        if cp_positions is not None:
+            return cp_positions
         if forward_batch.mrope_positions is None:
             return forward_batch.positions
 
@@ -782,7 +785,9 @@ class PrefillCudaGraphRunner(BaseCudaGraphRunner):
             if self._uses_eager_prefill_tail():
                 # BCG / Full: capture the transformer body only.
                 positions = self._get_layer_model_positions(forward_batch)
-                input_ids = forward_batch.input_ids
+                input_ids = getattr(
+                    forward_batch, "_cp_input_ids", forward_batch.input_ids
+                )
                 kwargs = _build_layer_model_forward_kwargs(
                     self.layer_model, forward_batch, pp_proxy_tensors
                 )
@@ -1336,9 +1341,9 @@ class PrefillCudaGraphRunner(BaseCudaGraphRunner):
             batch_max_context_len=batch_max_context_len,
         ):
             return False
-        if getattr(self, "enable_cp_bcg_capture", False) and is_cp_active(
-            forward_batch
-        ):
+        if getattr(self, "enable_cp_bcg_capture", False):
+            if not is_cp_active(forward_batch):
+                return False
             assert self.prefill_cp_bcg_input is not None
             if (
                 self.prefill_cp_bcg_input.select_replay_bucket_for_batch(

python/sglang/srt/models/deepseek_v4.py

@@ -182,6 +182,7 @@ from sglang.srt.multimodal.deepseek_v41_image_processing import (
 )
 from sglang.srt.runtime_context import (
     get_device,
+    get_disagg,
     get_exec,
     get_forward,
     get_parallel,
@@ -2039,7 +2040,10 @@ class MQALayer(MqaAttentionBase):
             if (
                 forward_batch.forward_mode.is_extend()
                 and is_in_breakable_cuda_graph()
-                and not getattr(attn_backend, "low_ratio_prefill_graph", False)
+                and (
+                    dsa_use_prefill_cp(forward_batch)
+                    or not getattr(attn_backend, "low_ratio_prefill_graph", False)
+                )
             ):
                 bcg_deepseek_v4_low_ratio_sources(self, x, q_lora, positions)
             else:
@@ -2650,7 +2654,8 @@ class DeepseekV4DecoderLayer(nn.Module):
             is_nextn=is_nextn,
             is_deepseek_v4=True,
             vl_correction_bias=config.model_type == "deepseek_v41"
-            and config.vision_n_layers > 0,
+            and config.vision_n_layers > 0
+            and not getattr(config, "language_model_only", False),
         )
 
         self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
@@ -3872,7 +3877,15 @@ class DeepseekV4DecoderLayer(nn.Module):
         finally:
             forward_batch.num_token_non_padded = saved_num_token_non_padded
         if _use_cp and get_moe_a2a_backend().is_none():
-            hidden_states = dsa_cp_reduce_scatter_hidden_states(hidden_states)
+            if self.config.model_type == "deepseek_v41":
+                parallel = get_parallel()
+                hidden_states = parallel.tp_group.all_reduce(hidden_states)
+                parallel = get_parallel()
+                hidden_states = hidden_states.tensor_split(parallel.attn_cp_size)[
+                    parallel.attn_cp_rank
+                ].contiguous()
+            else:
+                hidden_states = dsa_cp_reduce_scatter_hidden_states(hidden_states)
         elif _use_tp_moe_gather:
             hidden_states, global_hidden_states = (
                 get_local_dp_buffer(get_parallel().tp_group),
@@ -4402,11 +4415,18 @@ class DeepseekV4Model(nn.Module):
         )
         if self.engram_hasher is not None:
             if cp_extend:
-                # n-gram hashing needs each token's predecessors: hash the whole prompt
+                # N-gram hashing needs each token's predecessors, so hash the
+                # whole prompt before selecting this CP rank's interleaved rows.
+                # The hasher builds request-to-token indices dynamically; keep
+                # that work at an eager break during breakable graph capture.
                 total = int(forward_batch.attn_cp_metadata.total_seq_lens)
-                hash_ids = self.engram_hasher(
-                    forward_batch.input_ids[:total], forward_batch
-                )
+                global_input_ids = forward_batch.input_ids[:total]
+                if is_in_breakable_cuda_graph():
+                    hash_ids = bcg_deepseek_v4_engram_hash_ids(
+                        self.engram_hasher, global_input_ids
+                    )
+                else:
+                    hash_ids = self.engram_hasher(global_input_ids, forward_batch)
                 parallel = get_parallel()
                 hash_ids = hash_ids[parallel.attn_cp_rank :: parallel.attn_cp_size]
                 pad_rows = hidden_states.shape[0] - hash_ids.shape[0]
@@ -4839,6 +4859,13 @@ class DeepseekV4Model(nn.Module):
         return hidden_states, pre_hc_head
 
 
+def _v41_vision_a2a_supported() -> bool:
+    backend = get_moe_a2a_backend()
+    return backend.is_none() or (
+        backend.is_megamoe() and get_disagg().disaggregation_mode == "decode"
+    )
+
+
 class DeepseekV4ForCausalLM(nn.Module):
     supports_cuda_vmm_feature_transport = True
 
@@ -4864,14 +4891,19 @@ class DeepseekV4ForCausalLM(nn.Module):
         self.wo_a_fp8 = wo_a_fp8_gemm_enabled(quant_config)
         self.determine_num_fused_shared_experts()
         self.vision = None
-        if config.model_type == "deepseek_v41" and config.vision_n_layers > 0:
+        if (
+            config.model_type == "deepseek_v41"
+            and config.vision_n_layers > 0
+            and not getattr(config, "language_model_only", False)
+        ):
             if (
                 get_parallel().attn_cp_size != 1
-                or get_parallel().pp_group.world_size != 1
-                or not get_moe_a2a_backend().is_none()
+                or get_pp_group().world_size != 1
+                or not _v41_vision_a2a_supported()
             ):
                 raise ValueError(
-                    "V4.1 vision currently supports TP/EP/DP without CP, PP or MoE A2A"
+                    "V4.1 vision supports TP/EP/DP without CP or PP; "
+                    "MoE A2A is supported only with MegaMoE on a PD decode node"
                 )
 
             args = SimpleNamespace(**vars(config), dim=config.hidden_size)
@@ -5078,16 +5110,19 @@ class DeepseekV4ForCausalLM(nn.Module):
             0 if is_shared_experts_fusion_disabled() else self.config.n_shared_experts
         )
 
-    def forward(
+    def prepare_language_model_inputs(
         self,
         input_ids: torch.Tensor,
-        positions: torch.Tensor,
         forward_batch: ForwardBatch,
         input_embeds: Optional[torch.Tensor] = None,
-        pp_proxy_tensors: Optional[PPProxyTensors] = None,
-    ) -> torch.Tensor:
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """Prepare full-sequence image embeddings and model IDs before CP splits.
+
+        Scheduler hash IDs stay intact for multimodal cache keys; the language
+        model uses image_token_id for Engram masking and visual MoE routing.
+        """
         if (
-            self.vision is not None
+            getattr(self, "vision", None) is not None
             and not forward_batch.forward_mode.is_decode()
             and not forward_batch.forward_mode.is_target_verify()
             and forward_batch.mm_inputs is not None
@@ -5096,7 +5131,7 @@ class DeepseekV4ForCausalLM(nn.Module):
             if input_embeds is not None:
                 raise ValueError("Cannot combine input_embeds and image inputs")
             input_embeds = self._prepare_mm_embeddings(input_ids, forward_batch)
-        if self.vision is not None and not (
+        if getattr(self, "vision", None) is not None and not (
             forward_batch.forward_mode.is_decode_or_idle()
             or forward_batch.forward_mode.is_target_verify()
         ):
@@ -5106,6 +5141,19 @@ class DeepseekV4ForCausalLM(nn.Module):
                 input_ids >= MM_PAD_SHIFT_VALUE, self.config.image_token_id
             )
 
+        return input_ids, input_embeds
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        input_embeds: Optional[torch.Tensor] = None,
+        pp_proxy_tensors: Optional[PPProxyTensors] = None,
+    ) -> torch.Tensor:
+        input_ids, input_embeds = self.prepare_language_model_inputs(
+            input_ids, forward_batch, input_embeds
+        )
         with get_attn_tp_context().maybe_input_scattered(forward_batch):
             hidden_states = self.model.forward(
                 input_ids, positions, forward_batch, input_embeds, pp_proxy_tensors

python/sglang/srt/server_args.py

@@ -376,6 +376,7 @@ class ServerArgs:
     # ===== END TO BE REFACTORED ====
 
     LANGUAGE_MODEL_ONLY_ARCHITECTURES = (
+        "DeepseekV4ForCausalLM",
         "MuseGlimmerForConditionalGeneration",
         "Cosmos3ForConditionalGeneration",
         "Cosmos3EdgeForConditionalGeneration",

python/sglang/test/dsv41_cp_test_utils.py

@@ -0,0 +1,79 @@
+"""Small CPU tensors; production CP slicing/gather, mocked collective transport."""
+
+from contextlib import ExitStack, contextmanager, nullcontext
+from types import SimpleNamespace as NS
+from unittest.mock import patch
+
+import torch
+
+from sglang.srt.layers.cp.interleave import InterleaveCPStrategy
+from sglang.srt.layers.cp.padding import pad_logical_token_to_physical
+from sglang.srt.model_executor.forward_batch_info import ForwardMode
+
+CP = "sglang.srt.layers.cp"
+
+
+@contextmanager
+def cp_context(size, rank, lengths=(3, 6), prefix_lengths=(7, 13)):
+    """Keep real interleave indexing/padding; replace only runtime context."""
+    strategy = InterleaveCPStrategy(size)
+    parallel = NS(attn_cp_size=size, attn_cp_rank=rank, attn_cp_group=None)
+    batch = NS(
+        forward_mode=ForwardMode.EXTEND,
+        input_ids=torch.arange(1, sum(lengths) + 1),
+        positions=torch.cat(
+            [
+                torch.arange(prefix, prefix + length)
+                for prefix, length in zip(prefix_lengths, lengths)
+            ]
+        ),
+        extend_seq_lens_cpu=list(lengths),
+        extend_prefix_lens_cpu=list(prefix_lengths),
+        mm_inputs=None,
+        spec_info=None,
+    )
+    batch.attn_cp_metadata = strategy.build_metadata(
+        sum(lengths), [p + n for p, n in zip(prefix_lengths, lengths)], list(lengths)
+    )
+    with ExitStack() as stack:
+        for module in ("base", "utils", "padding", "interleave"):
+            stack.enter_context(
+                patch(CP + "." + module + ".get_parallel", return_value=parallel)
+            )
+        stack.enter_context(patch(CP + ".utils.get_cp_strategy", return_value=strategy))
+        stack.enter_context(
+            patch(CP + ".padding.get_cp_padding_align_size", return_value=size)
+        )
+        stack.enter_context(
+            patch(
+                CP + ".utils.get_moe_a2a_backend", return_value=NS(is_none=lambda: True)
+            )
+        )
+        pad_logical_token_to_physical(batch.attn_cp_metadata)
+        yield strategy, batch
+
+
+@contextmanager
+def simulated_collective(strategy, batch, global_tensor):
+    """Inject peer buffers into all-gather; retain production unpadding/reordering."""
+    physical = max(batch.attn_cp_metadata.per_rank_actual_token)
+    buffers = []
+    for rank in range(strategy.cp_size):
+        buf = global_tensor.new_zeros((physical, *global_tensor.shape[1:]))
+        local = global_tensor[rank :: strategy.cp_size]
+        buf[: len(local)] = local
+        buffers.append(buf)
+
+    def gather(output, local):
+        torch.testing.assert_close(local, buffers[strategy.cp_rank], rtol=0, atol=0)
+        output.copy_(torch.cat(buffers))
+
+    with (
+        patch(
+            CP + ".interleave.use_symmetric_memory",
+            side_effect=lambda *a, **k: nullcontext(),
+        ),
+        patch(CP + ".interleave.is_allocation_symmetric", return_value=False),
+        patch(CP + ".interleave.attn_cp_all_gather_into_tensor", side_effect=gather),
+    ):
+        yield

test/registered/kernels/ops/attention/test_dsv41_shared_prefill_kv.py

@@ -0,0 +1,151 @@
+"""Shared compressed-KV workspace validity across layers and forwards."""
+import sys
+from types import SimpleNamespace as NS
+from unittest.mock import patch
+
+import pytest
+import torch
+
+from sglang.srt.layers.attention import deepseek_v4_backend as backend
+from sglang.srt.layers.attention.dsv4.sparse_prefill_utils import (
+    CompressedGather,
+    SparsePrefillWorkspace,
+    WORKSPACE_DIM,
+)
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=30, stage="base-b-kernel-unit", runner_config="1-gpu-large")
+
+
+def test_shared_compressed_dequant_lifetime():
+    device = "cuda"
+    obj = object.__new__(backend.DeepseekV4AttnBackend)
+    obj.sparse_prefill_workspace = SparsePrefillWorkspace(device)
+    obj.shared_compressed_prefill_workspaces = {
+        ratio: SparsePrefillWorkspace(device) for ratio in (1, 2)
+    }
+    obj.softmax_scale = 0.1
+    obj.head_dim_v = WORKSPACE_DIM
+    sources = {0: 0, 1: 1, 2: 0, 3: 1, 4: 4, 5: 4, 6: 6, 7: 7}
+    ratios = {0: 1, 1: 2, 2: 1, 3: 2, 4: 1, 5: 1, 6: 0, 7: 4}
+    compressed = {
+        source: torch.full(
+            (128, 1, WORKSPACE_DIM),
+            float(source + 1),
+            device=device,
+            dtype=torch.bfloat16,
+        )
+        for source in (0, 1, 4, 7)
+    }
+    swa = torch.empty((16, 1, WORKSPACE_DIM), device=device, dtype=torch.bfloat16)
+    pool = NS(
+        source_layer_of=lambda layer: sources[layer],
+        get_extra_key_page_size=lambda layer: 1,
+        get_extra_key_buffer=lambda layer: compressed[sources[layer]],
+        get_extra_key_layout=lambda layer: None,
+        get_swa_key_buffer_radix=lambda layer: swa,
+        get_swa_key_layout=lambda: None,
+    )
+    calls = []
+    active = [True]
+
+    def dequant(src, indices, *, out, **kwargs):
+        calls.append("swa" if src is swa else sources_by_ptr[src.data_ptr()])
+        out.copy_(src.index_select(0, indices.long()))
+
+    sources_by_ptr = {v.data_ptr(): k for k, v in compressed.items()}
+
+    def make_cache(n):
+        gathers = {
+            ratio: CompressedGather(
+                flat_token_ids=torch.arange(n // ratio, device=device, dtype=torch.int32),
+                compressed_base=torch.zeros(1, device=device, dtype=torch.int32),
+                swa_base=torch.zeros(1, device=device, dtype=torch.int32),
+            )
+            for ratio in (1, 2, 4)
+        }
+        indices = torch.zeros((4, 128), device=device, dtype=torch.int32)
+        lengths = torch.full((4,), 3, device=device, dtype=torch.int32)
+        cache = NS(
+            compressed=gathers,
+            swa_token_ids=torch.arange(3, device=device),
+            swa_page_size=1,
+            c0_combined_indices=indices,
+            c0_combined_lens=lengths,
+        )
+        cache.layer_inputs = lambda ratio, core, page: (
+            gathers[ratio].flat_token_ids,
+            indices,
+            lengths,
+        )
+        return cache
+
+    def forward(layer):
+        return obj._forward_prefill_sparse(
+            torch.empty((4, 1, 1, WORKSPACE_DIM), device=device, dtype=torch.bfloat16),
+            layer,
+            ratios[layer],
+            NS(),
+            pool,
+            NS(),
+            torch.zeros(1, device=device),
+        )
+
+    with (
+        patch.object(backend, "is_cp_active", side_effect=lambda _: active[0]),
+        patch.object(backend, "dequantize_k_cache_paged", side_effect=dequant),
+        patch(
+            "sgl_kernel.flash_mla.flash_mla_sparse_fwd",
+            side_effect=lambda **kw: (kw["kv"].clone(), None, None),
+        ),
+    ):
+        # Same-size replay, then growth and shrink must all read freshly written KV.
+        for step, n in enumerate((8, 8, 40, 4)):
+            obj.forward_metadata = NS(sparse_prefill_cache=make_cache(n))
+            for source, tensor in compressed.items():
+                tensor.fill_(source + 1 + step * 10)
+            for layer, should_dequant in (
+                (0, True),
+                (1, True),
+                (2, False),
+                (3, False),
+                (4, True),
+                (5, False),
+                (2, True),
+                (3, False),
+                (6, False),
+                (7, True),
+                (7, True),
+            ):
+                swa.fill_(100 + layer + step)
+                before = len(calls)
+                active[0] = True
+                actual = forward(layer)
+                actual_calls = calls[before:]
+                assert actual_calls.count("swa") == 1
+                assert len(actual_calls) == 1 + int(should_dequant)
+                active[0] = False
+                expected = forward(layer)
+                torch.testing.assert_close(actual, expected, rtol=0, atol=0)
+
+            # Workspace replacement invalidates even an unchanged source identity.
+            obj.shared_compressed_prefill_workspaces[1].get(256 + step * 256)
+            active[0] = True
+            before = len(calls)
+            actual = forward(2)
+            assert calls[before:] == [0, "swa"]
+            active[0] = False
+            torch.testing.assert_close(actual, forward(2), rtol=0, atol=0)
+
+            # Re-executing a producer may mutate the same cache address in place.
+            compressed[0].add_(1)
+            active[0] = True
+            before = len(calls)
+            actual = forward(0)
+            assert calls[before:] == [0, "swa"]
+            active[0] = False
+            torch.testing.assert_close(actual, forward(0), rtol=0, atol=0)
+
+
+if __name__ == "__main__":
+    sys.exit(pytest.main([__file__, *sys.argv[1:]]))

test/registered/unit/disaggregation/test_dsv41_dspark_pd.py

@@ -0,0 +1,286 @@
+import asyncio
+import copy
+import json
+import unittest
+from types import SimpleNamespace
+from unittest.mock import AsyncMock, Mock, patch
+
+from sglang.srt.disaggregation.base.conn import StateType
+from sglang.srt.disaggregation.common.conn import (
+    CommonKVBootstrapServer,
+    CommonKVManager,
+)
+from sglang.srt.disaggregation.decode import DecodePreallocQueue
+from sglang.srt.disaggregation.utils import get_dsv41_spec_layout
+from sglang.srt.mem_cache.common import retraction_backup
+from sglang.srt.mem_cache.deepseek_v4_memory_pool import DeepSeekV4TokenToKVPool
+from sglang.srt.runtime_context import get_context
+from sglang.test.ci.ci_register import register_cpu_ci
+from sglang.test.test_utils import CustomTestCase
+
+register_cpu_ci(est_time=2, suite="base-a-test-cpu")
+
+
+def make_layout():
+    args = SimpleNamespace(
+        mla_compression_ratios=[0, 2, 1],
+        kv_layer_ids=[1, 2],
+        kv_item_lens=[512, 1024],
+        state_types=[StateType.SWA, StateType.DSV4_REQUEST_STATE, StateType.SWA],
+        state_item_lens=[[512], [32768], [512]],
+    )
+    with get_context().override_server_args(
+        speculative_algorithm="DSPARK", speculative_num_draft_tokens=6
+    ):
+        return get_dsv41_spec_layout(args)
+
+
+class TestDSV41DSparkPD(CustomTestCase):
+    def test_bootstrap_validates_before_caching(self):
+        layout = make_layout()
+        cases = [("matching", layout, layout, 4, True), ("legacy", None, None, 2, True)]
+        for key, value in (
+            ("num_draft_tokens", 5),
+            ("kv_layer_ids", [2, 1]),
+            ("kv_item_lens", [256, 1024]),
+            ("state_types", ["swa", "c128_state"]),
+            ("state_item_lens", [[512], [8192], [512]]),
+        ):
+            different = copy.deepcopy(layout)
+            different[key] = value
+            cases.append((key, layout, different, 4, False))
+        cases += [
+            ("prefill_only", None, layout, 4, False),
+            ("decode_only_or_old_prefill", layout, None, 4, False),
+            ("tp_mismatch", layout, layout, 2, False),
+        ]
+        for name, local, peer, tp_size, supported in cases:
+            with self.subTest(name=name):
+                manager = object.__new__(CommonKVManager)
+                manager.prefill_info_table = {}
+                manager.kv_args = SimpleNamespace(page_size=256)
+                manager.kv_cache_dtype_str = "fp8_e4m3"
+                manager.dsv41_spec_layout = local
+                manager.attn_tp_size = 4
+                manager.dcp_size = 1
+                manager._resolve_rank_mapping = Mock()
+                response = Mock(status_code=200)
+                response.json.return_value = dict(
+                    attn_tp_size=tp_size,
+                    attn_cp_size=1,
+                    dp_size=1,
+                    pp_size=1,
+                    page_size=256,
+                    kv_cache_dtype="fp8_e4m3",
+                    follow_bootstrap_room=True,
+                    dsv41_spec_layout=peer,
+                )
+                with patch(
+                    "sglang.srt.disaggregation.common.conn.requests.get",
+                    return_value=response,
+                ) as fetch:
+                    if supported:
+                        self.assertTrue(
+                            manager.try_ensure_parallel_info("prefill:8998")
+                        )
+                        self.assertTrue(
+                            manager.try_ensure_parallel_info("prefill:8998")
+                        )
+                        fetch.assert_called_once()
+                    else:
+                        with self.assertRaisesRegex(
+                            RuntimeError, "DeepSeek-V4.1 DSpark PD"
+                        ):
+                            manager.try_ensure_parallel_info("prefill:8998")
+                        self.assertFalse(manager.prefill_info_table)
+                        manager._resolve_rank_mapping.assert_not_called()
+
+    def test_python_bootstrap_preserves_layout_and_rejects_mixed_ranks(self):
+        with patch.object(CommonKVBootstrapServer, "run"):
+            server = CommonKVBootstrapServer("127.0.0.1", 8998)
+        layout = make_layout()
+        payload = dict(
+            attn_tp_size=1,
+            attn_tp_rank=0,
+            attn_cp_size=1,
+            attn_cp_rank=0,
+            attn_dp_size=1,
+            attn_dp_rank=0,
+            pp_size=1,
+            pp_rank=0,
+            system_dp_size=1,
+            system_dp_rank=0,
+            rank_ip="127.0.0.1",
+            rank_port=1234,
+            page_size=256,
+            kv_cache_dtype="fp8_e4m3",
+            dsv41_spec_layout=layout,
+        )
+        request = Mock(json=AsyncMock(return_value=payload))
+        self.assertEqual(asyncio.run(server._handle_route_put(request)).status, 200)
+        query = Mock(
+            query={
+                key: "-1"
+                for key in (
+                    "prefill_dp_rank",
+                    "prefill_cp_rank",
+                    "target_tp_rank",
+                    "target_pp_rank",
+                )
+            }
+        )
+        response = asyncio.run(server._handle_route_get(query))
+        self.assertEqual(json.loads(response.text)["dsv41_spec_layout"], layout)
+        payload["dsv41_spec_layout"] = None
+        self.assertEqual(asyncio.run(server._handle_route_put(request)).status, 400)
+        self.assertEqual(server._registered_count, 1)
+        self.assertEqual(server.dsv41_spec_layout, layout)
+
+    def test_retraction_recomputes_from_prefill_and_replays_boundary_token(self):
+        pool = object.__new__(DeepSeekV4TokenToKVPool)
+        pool.compression_ratios = [0, 2, 1]
+        pool.device = "cuda"
+        allocator = Mock(get_kvcache=Mock(return_value=pool))
+        for algorithm in (None, "DSPARK"):
+            with (
+                self.subTest(algorithm=algorithm),
+                get_context().override_server_args(speculative_algorithm=algorithm),
+                patch("torch.get_device_module") as device_module,
+            ):
+                req = SimpleNamespace(
+                    output_ids=[7, 8],
+                    bootstrap_host="prefill",
+                    time_stats=Mock(),
+                    offload_kv_cache=Mock(),
+                )
+                request_pool = Mock()
+                self.assertTrue(
+                    retraction_backup(
+                        req, Mock(), request_pool, allocator, "cpu_tensor"
+                    )
+                )
+                queue = SimpleNamespace(
+                    token_to_kv_pool_allocator=allocator,
+                    _check_if_req_exceed_kv_capacity=Mock(return_value=False),
+                    _create_receiver_and_enqueue=Mock(),
+                    _resolve_prefill_dp_rank=Mock(return_value=0),
+                    retracted_queue=[],
+                    pending_reqs=[],
+                )
+                DecodePreallocQueue.add(queue, req, is_retracted=True)
+                if algorithm == "DSPARK":
+                    req.offload_kv_cache.assert_not_called()
+                    device_module.return_value.synchronize.assert_called_once_with(
+                        "cuda"
+                    )
+                    self.assertEqual(req.output_ids, [7])
+                    self.assertEqual(req.pd_rebootstrap_forced_output_id, 8)
+                    self.assertTrue(req.pd_rebootstrap_in_progress)
+                    queue._create_receiver_and_enqueue.assert_called_once_with(
+                        req, is_rebootstrap=True
+                    )
+                    self.assertFalse(queue.retracted_queue)
+                else:
+                    req.offload_kv_cache.assert_called_once_with(
+                        request_pool, allocator
+                    )
+                    device_module.assert_not_called()
+                    self.assertEqual(req.output_ids, [7, 8])
+                    self.assertEqual(queue.retracted_queue, [req])
+
+
+class TestDSV41CPPDHandshake(CustomTestCase):
+    def make(self, rank=0, hybrid=True):
+        m = object.__new__(CommonKVManager)
+        m.prefill_info_table = {}
+        m.kv_args = SimpleNamespace(page_size=256, engine_rank=rank)
+        m.kv_cache_dtype_str = "fp8_e4m3"
+        m.dsv41_spec_layout = {"kv_item_lens": [512], "state_item_lens": [[32768]]}
+        m.attn_tp_size = 4
+        m.attn_cp_size = 1
+        m.attn_cp_rank = 0
+        m.dcp_size = 1
+        m.is_mla_backend = False
+        m.is_hybrid_mla_backend = hybrid
+        m.enable_all_cp_ranks_for_transfer = True
+        m.pp_size = 1
+        m.pp_rank = 0
+        return m
+
+    def fetch(self, m, tp, cp, layout=None):
+        response = Mock(status_code=200)
+        response.json.return_value = dict(
+            attn_tp_size=tp,
+            attn_cp_size=cp,
+            dp_size=1,
+            pp_size=1,
+            page_size=256,
+            kv_cache_dtype="fp8_e4m3",
+            follow_bootstrap_room=True,
+            dsv41_spec_layout=layout or m.dsv41_spec_layout,
+        )
+        with patch(
+            "sglang.srt.disaggregation.common.conn.requests.get", return_value=response
+        ):
+            return m.try_ensure_parallel_info("prefill:8761")
+
+    def test_cp4_maps_all_shards_to_each_decode_rank(self):
+        for rank in range(4):
+            m = self.make(rank)
+            self.assertTrue(self.fetch(m, 1, 4))
+            info = m.prefill_info_table["prefill:8761"]
+            self.assertEqual(info.target_tp_ranks, [0])
+            self.assertEqual(info.target_cp_ranks, [0, 1, 2, 3])
+            self.assertEqual(info.required_prefill_response_num, 4)
+            self.assertEqual(info.required_dst_info_num, 4)
+
+    def test_dsv4_pool_is_classified_as_mla(self):
+        from sglang.srt.disaggregation.utils import is_mla_backend
+        from sglang.srt.mem_cache.deepseek_v4_memory_pool import DeepSeekV4TokenToKVPool
+
+        pool = object.__new__(DeepSeekV4TokenToKVPool)
+        self.assertTrue(is_mla_backend(pool))
+        m = self.make(hybrid=False)
+        m.is_mla_backend = is_mla_backend(pool)
+        self.assertTrue(self.fetch(m, 1, 4))
+        self.assertEqual(
+            m.prefill_info_table["prefill:8761"].required_prefill_response_num, 4
+        )
+
+    def test_cp2_tp2_maps_corresponding_tp_and_both_cp_ranks(self):
+        for rank in range(4):
+            m = self.make(rank)
+            self.assertTrue(self.fetch(m, 2, 2))
+            info = m.prefill_info_table["prefill:8761"]
+            self.assertEqual(info.target_tp_ranks, [rank // 2])
+            self.assertEqual(info.target_cp_ranks, [0, 1])
+            self.assertEqual(info.required_prefill_response_num, 2)
+
+    def test_plain_tp4_unchanged(self):
+        m = self.make(3)
+        self.assertTrue(self.fetch(m, 4, 1))
+        info = m.prefill_info_table["prefill:8761"]
+        self.assertEqual(info.target_tp_ranks, [3])
+        self.assertEqual(info.target_cp_ranks, [0])
+
+    def test_unequal_model_tp_rejected(self):
+        for tp, cp in [(2, 1), (1, 2), (1, 8)]:
+            m = self.make()
+            with self.assertRaisesRegex(RuntimeError, "same TP size"):
+                self.fetch(m, tp, cp)
+            self.assertFalse(m.prefill_info_table)
+
+    def test_nonhybrid_cp_mismatch_rejected(self):
+        m = self.make(hybrid=False)
+        with self.assertRaisesRegex(RuntimeError, "same TP size"):
+            self.fetch(m, 1, 4)
+
+    def test_layout_mismatch_still_rejected(self):
+        m = self.make()
+        with self.assertRaisesRegex(RuntimeError, "layout mismatch"):
+            self.fetch(m, 1, 4, {"kv_item_lens": [1024], "state_item_lens": [[32768]]})
+        self.assertFalse(m.prefill_info_table)
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/unit/mem_cache/test_fake_transfer_chunk_progress.py

@@ -0,0 +1,46 @@
+"""Chunk continuation when fake PD transfer skips shared radix insertion."""
+
+import unittest
+from types import SimpleNamespace as NS
+from unittest.mock import Mock
+
+import torch
+
+from sglang.srt.mem_cache.common import maybe_cache_unfinished_req
+from sglang.test.ci.ci_register import register_cpu_ci
+from sglang.test.test_utils import CustomTestCase
+
+register_cpu_ci(est_time=5, suite="base-a-test-cpu")
+
+
+class TestFakeTransferChunkProgress(CustomTestCase):
+    def test_chunk_progress_without_shared_insert(self):
+        slots = torch.arange(16385, dtype=torch.int32).reshape(1, -1)
+        cache = NS(
+            req_to_token_pool=NS(req_to_token=slots), cache_unfinished_req=Mock()
+        )
+        req = NS(
+            skip_radix_cache_insert=True,
+            kv=NS(req_pool_idx=0, cache_protected_len=0),
+            get_fill_ids=lambda: range(16384),
+        )
+        maybe_cache_unfinished_req(req, cache, chunked=True)
+        self.assertEqual(16385 - len(req.prefix_indices), 1)
+        self.assertEqual(req.kv.cache_protected_len, 0)
+        cache.cache_unfinished_req.assert_not_called()
+        self.assertEqual(req.prefix_indices.dtype, torch.int64)
+        slots[0, 0] = -1
+        self.assertEqual(req.prefix_indices[0].item(), 0)
+        req.get_fill_ids = lambda: range(16385)
+        maybe_cache_unfinished_req(req, cache, chunked=True)
+        self.assertEqual(len(req.prefix_indices), 16385)
+
+    def test_real_transfer_preserves_cache_path(self):
+        req = NS(skip_radix_cache_insert=False)
+        cache = NS(cache_unfinished_req=Mock())
+        maybe_cache_unfinished_req(req, cache, chunked=True)
+        cache.cache_unfinished_req.assert_called_once_with(req, chunked=True)
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/unit/models/test_deepseek_v41_multimodal_cp.py

@@ -0,0 +1,316 @@
+"""V4.1 image/text CP input contracts; vision and model compute are mocked."""
+
+import unittest
+from contextlib import contextmanager
+from types import SimpleNamespace as NS
+from unittest.mock import Mock, patch
+
+import torch
+
+from sglang.srt.managers.schedule_batch import MultimodalInputs
+from sglang.srt.model_executor.forward_batch_info import ForwardMode
+from sglang.srt.model_executor.runner.eager_runner import EagerRunner
+from sglang.srt.models.deepseek_v4 import MM_PAD_SHIFT_VALUE, DeepseekV4ForCausalLM
+from sglang.test.ci.ci_register import register_cpu_ci
+from sglang.test.dsv41_cp_test_utils import cp_context, simulated_collective
+from sglang.test.test_utils import CustomTestCase
+
+register_cpu_ci(est_time=5, suite="base-a-test-cpu")
+
+
+MODEL = "sglang.srt.models.deepseek_v4"
+RUNNER = "sglang.srt.model_executor.runner.eager_runner"
+IMAGE_ID = 129264
+
+
+class TestDSV41MultimodalCP(CustomTestCase):
+    def test_image_spans_cross_ranks_before_shard_and_gather(self):
+        # Two image spans with distinct cache hashes; first request is text-only.
+        original = torch.tensor(
+            [
+                7,
+                8,
+                9,
+                MM_PAD_SHIFT_VALUE + 11,
+                MM_PAD_SHIFT_VALUE + 11,
+                10,
+                MM_PAD_SHIFT_VALUE + 23,
+                MM_PAD_SHIFT_VALUE + 23,
+                12,
+            ]
+        )
+        normalized = torch.tensor(
+            [7, 8, 9, IMAGE_ID, IMAGE_ID, 10, IMAGE_ID, IMAGE_ID, 12]
+        )
+        full = torch.arange(36, dtype=torch.float32).reshape(9, 4)
+        # Distinct image features expose using text embeddings or wrong row order.
+        full[3:5] += 1000
+        full[6:8] += 2000
+        for size in (2, 4):
+            for rank in range(size):
+                with (
+                    self.subTest(size=size, rank=rank),
+                    cp_context(size, rank) as (strategy, batch),
+                ):
+                    batch.input_ids = original.clone()
+                    batch.mm_inputs = [None, MultimodalInputs(mm_items=[])]
+                    model = NS(
+                        vision=object(),
+                        config=NS(image_token_id=IMAGE_ID),
+                        get_input_embeddings=Mock(
+                            side_effect=AssertionError(
+                                "Raw image hashes entered text embeddings"
+                            )
+                        ),
+                        _prepare_mm_embeddings=Mock(return_value=full),
+                        capture_aux_hidden_states=False,
+                        pp_group=NS(is_last_rank=True),
+                        lm_head=object(),
+                        logits_processor=Mock(return_value="ok"),
+                    )
+                    model.prepare_language_model_inputs = lambda ids, fb, emb: (
+                        DeepseekV4ForCausalLM.prepare_language_model_inputs(
+                            model, ids, fb, emb
+                        )
+                    )
+
+                    def body(ids, positions, fb, input_embeds):
+                        model._prepare_mm_embeddings.assert_called_once_with(
+                            batch.input_ids, batch
+                        )
+                        n = len(normalized[rank::size])
+                        torch.testing.assert_close(ids[:n], normalized[rank::size])
+                        torch.testing.assert_close(input_embeds[:n], full[rank::size])
+                        torch.testing.assert_close(
+                            positions[:n], batch.positions[rank::size]
+                        )
+                        self.assertFalse(
+                            (fb.input_ids_global >= MM_PAD_SHIFT_VALUE).any().item()
+                        )
+                        return input_embeds
+
+                    model.model = body
+                    with (
+                        simulated_collective(strategy, batch, full),
+                        patch(RUNNER + ".torch.cuda.current_stream", return_value=None),
+                    ):
+                        result = EagerRunner._execute_extend_cp(
+                            NS(model_runner=NS(model=model)), batch, {}
+                        )
+                    self.assertEqual(result, "ok")
+                    torch.testing.assert_close(
+                        model.logits_processor.call_args.args[0], normalized
+                    )
+                    torch.testing.assert_close(
+                        model.logits_processor.call_args.args[1], full
+                    )
+                    torch.testing.assert_close(batch.input_ids, original)
+                    self.assertFalse(hasattr(batch, "input_ids_global"))
+
+    def test_chunk_prefix_metadata_and_scheduler_hashes_survive_embedder(self):
+        for prefixes, lengths in (([0, 0], [3, 6]), ([16384, 127], [3, 6])):
+            with self.subTest(prefixes=prefixes):
+                ids = torch.tensor([7, 8, 9] + [MM_PAD_SHIFT_VALUE + 17] * 6)
+                original = ids.clone()
+                image = MultimodalInputs(mm_items=[])
+                batch = NS(
+                    mm_inputs=[None, image],
+                    extend_prefix_lens_cpu=prefixes,
+                    extend_seq_lens_cpu=lengths,
+                )
+                full = torch.arange(27, dtype=torch.float32).reshape(9, 3)
+                embedding = object()
+                model = NS(get_input_embeddings=lambda: embedding)
+
+                def embed(**kwargs):
+                    self.assertEqual(kwargs["extend_prefix_lens"], prefixes)
+                    self.assertEqual(kwargs["extend_seq_lens"], lengths)
+                    self.assertIs(kwargs["mm_inputs_list"][1], image)
+                    self.assertEqual(kwargs["mm_inputs_list"][0].mm_items, [])
+                    self.assertIs(kwargs["input_embedding"], embedding)
+                    self.assertNotEqual(kwargs["input_ids"].data_ptr(), ids.data_ptr())
+                    kwargs["input_ids"].zero_()
+                    return full, {}
+
+                with patch(MODEL + ".embed_mm_inputs", side_effect=embed) as mocked:
+                    result = DeepseekV4ForCausalLM._prepare_mm_embeddings(
+                        model, ids, batch
+                    )
+                mocked.assert_called_once()
+                self.assertIs(result, full)
+                self.assertIs(batch.mm_input_embeds, full)
+                torch.testing.assert_close(ids, original)
+
+    def test_vision_enabled_text_batch_skips_image_encoder(self):
+        for mm_inputs in (None, [None, None], []):
+            with self.subTest(mm_inputs=mm_inputs):
+                ids = torch.tensor([4, 5, 6])
+                model = NS(
+                    vision=object(),
+                    config=NS(image_token_id=IMAGE_ID),
+                    _prepare_mm_embeddings=Mock(),
+                )
+                batch = NS(forward_mode=ForwardMode.EXTEND, mm_inputs=mm_inputs)
+                result, embeds = DeepseekV4ForCausalLM.prepare_language_model_inputs(
+                    model, ids, batch
+                )
+                torch.testing.assert_close(result, ids)
+                self.assertIsNone(embeds)
+                model._prepare_mm_embeddings.assert_not_called()
+
+    def test_decode_idle_and_verify_preserve_vocab_ids(self):
+        for mode in (ForwardMode.DECODE, ForwardMode.IDLE, ForwardMode.TARGET_VERIFY):
+            with self.subTest(mode=mode):
+                ids = torch.tensor([4, IMAGE_ID, 6])
+                model = NS(
+                    vision=object(),
+                    config=NS(image_token_id=IMAGE_ID),
+                    _prepare_mm_embeddings=Mock(),
+                )
+                batch = NS(forward_mode=mode, mm_inputs=None)
+                result, embeds = DeepseekV4ForCausalLM.prepare_language_model_inputs(
+                    model, ids, batch
+                )
+                self.assertIs(result, ids)
+                self.assertIsNone(embeds)
+                model._prepare_mm_embeddings.assert_not_called()
+
+    def test_image_embedding_failure_does_not_mutate_scheduler_ids(self):
+        ids = torch.tensor([7, MM_PAD_SHIFT_VALUE + 12, 8])
+        original = ids.clone()
+        batch = NS(
+            mm_inputs=[MultimodalInputs(mm_items=[])],
+            extend_prefix_lens_cpu=[0],
+            extend_seq_lens_cpu=[3],
+        )
+        model = NS(get_input_embeddings=lambda: object())
+
+        def fail(**kwargs):
+            kwargs["input_ids"].zero_()
+            raise RuntimeError("vision failure")
+
+        with patch(MODEL + ".embed_mm_inputs", side_effect=fail):
+            with self.assertRaisesRegex(RuntimeError, "vision failure"):
+                DeepseekV4ForCausalLM._prepare_mm_embeddings(model, ids, batch)
+        torch.testing.assert_close(ids, original)
+        self.assertFalse(hasattr(batch, "mm_input_embeds"))
+
+
+class TestDSV41MultimodalInputs(CustomTestCase):
+    def setUp(self):
+        self.ids = torch.tensor(
+            [7, MM_PAD_SHIFT_VALUE + 12, MM_PAD_SHIFT_VALUE + 12, 9, 10]
+        )
+        self.original = self.ids.clone()
+        self.embeds = torch.arange(15, dtype=torch.float32).reshape(5, 3)
+        self.model = NS(
+            vision=object(),
+            config=NS(image_token_id=129264),
+            _prepare_mm_embeddings=Mock(return_value=self.embeds),
+        )
+        self.batch = NS(
+            input_ids=self.ids,
+            forward_mode=ForwardMode.EXTEND,
+            mm_inputs=[MultimodalInputs(mm_items=[])],
+        )
+
+    def test_prepare_global_embeddings_and_normalized_ids(self):
+        ids, embeds = DeepseekV4ForCausalLM.prepare_language_model_inputs(
+            self.model, self.ids, self.batch
+        )
+        self.assertEqual(ids.tolist(), [7, 129264, 129264, 9, 10])
+        self.assertIs(embeds, self.embeds)
+        self.model._prepare_mm_embeddings.assert_called_once_with(self.ids, self.batch)
+        self.assertTrue(torch.equal(self.ids, self.original))
+
+    def test_reject_preembedded_images(self):
+        with self.assertRaisesRegex(ValueError, "Cannot combine"):
+            DeepseekV4ForCausalLM.prepare_language_model_inputs(
+                self.model, self.ids, self.batch, self.embeds
+            )
+
+    def test_text_only_model_keeps_existing_embeddings(self):
+        self.model.vision = None
+        ids, embeds = DeepseekV4ForCausalLM.prepare_language_model_inputs(
+            self.model, self.ids, self.batch, self.embeds
+        )
+        self.assertIs(ids, self.ids)
+        self.assertIs(embeds, self.embeds)
+        self.model._prepare_mm_embeddings.assert_not_called()
+
+    def test_text_subclass_without_vision_module(self):
+        del self.model.vision
+        ids, embeds = DeepseekV4ForCausalLM.prepare_language_model_inputs(
+            self.model, self.ids, self.batch, self.embeds
+        )
+        self.assertIs(ids, self.ids)
+        self.assertIs(embeds, self.embeds)
+
+    def test_decode_keeps_vocabulary_ids(self):
+        self.batch.forward_mode = ForwardMode.DECODE
+        ids, embeds = DeepseekV4ForCausalLM.prepare_language_model_inputs(
+            self.model, self.ids, self.batch
+        )
+        self.assertIs(ids, self.ids)
+        self.assertIsNone(embeds)
+        self.model._prepare_mm_embeddings.assert_not_called()
+
+    def test_embedding_does_not_mutate_scheduler_hashes(self):
+        self.batch.extend_prefix_lens_cpu = [0]
+        self.batch.extend_seq_lens_cpu = [5]
+        self.model.get_input_embeddings = lambda: None
+
+        def embed(**kwargs):
+            kwargs["input_ids"].zero_()
+            return (self.embeds, {})
+
+        with patch("sglang.srt.models.deepseek_v4.embed_mm_inputs", side_effect=embed):
+            result = DeepseekV4ForCausalLM._prepare_mm_embeddings(
+                self.model, self.ids, self.batch
+            )
+        self.assertTrue(torch.equal(self.ids, self.original))
+        self.assertIs(result, self.batch.mm_input_embeds)
+
+    def test_cp_runner_prepares_before_sharding_and_uses_model_ids_for_logits(self):
+        normalized = torch.tensor([7, 129264, 129264, 9, 10])
+        self.batch.positions = torch.arange(5)
+        self.model.prepare_language_model_inputs = lambda ids, batch, emb: (
+            DeepseekV4ForCausalLM.prepare_language_model_inputs(
+                self.model, ids, batch, emb
+            )
+        )
+        self.model.get_input_embeddings = Mock(
+            side_effect=AssertionError("Raw hashes must not enter text embedding")
+        )
+        self.model.model = Mock(return_value=self.embeds[1::4])
+        self.model.capture_aux_hidden_states = False
+        self.model.pp_group = NS(is_last_rank=True)
+        self.model.lm_head = object()
+        self.model.logits_processor = Mock(return_value="ok")
+
+        @contextmanager
+        def shard(embeds, positions, batch, ids):
+            self.assertIs(embeds, self.embeds)
+            self.assertTrue(torch.equal(ids, normalized))
+            yield (embeds[1::4], positions[1::4], ids[1::4])
+
+        runner = NS(model_runner=NS(model=self.model))
+        module = "sglang.srt.model_executor.runner.eager_runner"
+        with (
+            patch(module + ".cp_shard_model_inputs", side_effect=shard),
+            patch(module + ".cp_gather_after_forward", return_value=self.embeds),
+            patch(module + ".torch.cuda.current_stream", return_value=None),
+        ):
+            result = EagerRunner._execute_extend_cp(runner, self.batch, {})
+        self.assertEqual(result, "ok")
+        args, kwargs = self.model.model.call_args
+        self.assertEqual(args[0].tolist(), [129264])
+        self.assertTrue(torch.equal(kwargs["input_embeds"], self.embeds[1::4]))
+        self.assertTrue(
+            torch.equal(self.model.logits_processor.call_args.args[0], normalized)
+        )
+        self.assertTrue(torch.equal(self.batch.input_ids, self.original))
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/unit/models/test_deepseek_v41_text_cp.py

@@ -0,0 +1,219 @@
+"""Pure-language V4.1 CP input, padding and DSpark state regressions."""
+
+import unittest
+from types import SimpleNamespace as NS
+from unittest.mock import Mock, patch
+
+import torch
+
+from sglang.srt.layers.cp.utils import (
+    cp_gather_after_forward,
+    cp_shard_model_inputs,
+    is_cp_active,
+)
+from sglang.srt.model_executor.runner.eager_runner import EagerRunner
+from sglang.srt.models.deepseek_v4 import DeepseekV4ForCausalLM
+from sglang.test.ci.ci_register import register_cpu_ci
+from sglang.test.dsv41_cp_test_utils import cp_context, simulated_collective
+from sglang.test.test_utils import CustomTestCase
+
+register_cpu_ci(est_time=5, suite="base-a-test-cpu")
+
+
+RUNNER = "sglang.srt.model_executor.runner.eager_runner"
+
+
+class TestDSV41TextCP(CustomTestCase):
+    def test_interleave_roundtrip_mixed_lengths_prefix_and_padding(self):
+        for size in (2, 4):
+            for length in (4, 5, 9, 127, 128, 129):
+                for rank in range(size):
+                    with (
+                        self.subTest(size=size, length=length, rank=rank),
+                        cp_context(size, rank, (1, length - 1), (0, 16384)) as (
+                            strategy,
+                            batch,
+                        ),
+                    ):
+                        embeddings = torch.arange(
+                            length * 3, dtype=torch.float32
+                        ).reshape(length, 3)
+                        original_ids = batch.input_ids.clone()
+                        with cp_shard_model_inputs(
+                            embeddings, batch.positions, batch, batch.input_ids
+                        ) as (local, positions, ids):
+                            count = len(embeddings[rank::size])
+                            torch.testing.assert_close(
+                                local[:count], embeddings[rank::size]
+                            )
+                            torch.testing.assert_close(
+                                positions[:count], batch.positions[rank::size]
+                            )
+                            torch.testing.assert_close(
+                                ids[:count], batch.input_ids[rank::size]
+                            )
+                            self.assertEqual(
+                                torch.count_nonzero(local[count:]).item(), 0
+                            )
+                            self.assertEqual(torch.count_nonzero(ids[count:]).item(), 0)
+                            with simulated_collective(strategy, batch, embeddings):
+                                restored = cp_gather_after_forward(local, batch)
+                            torch.testing.assert_close(
+                                restored, embeddings, rtol=0, atol=0
+                            )
+                        torch.testing.assert_close(batch.input_ids, original_ids)
+                        self.assertFalse(hasattr(batch, "input_ids_global"))
+
+    def test_speculative_state_and_global_ids_restored_on_exception(self):
+        for size in (2, 4):
+            for rank in range(size):
+                for had_global in (False, True):
+                    with (
+                        self.subTest(size=size, rank=rank, had_global=had_global),
+                        cp_context(size, rank) as (_, batch),
+                    ):
+                        full = torch.arange(36, dtype=torch.float32).reshape(9, 4)
+                        batch.spec_info = NS(hidden_states=full)
+                        previous = object()
+                        if had_global:
+                            batch.input_ids_global = previous
+                        with self.assertRaisesRegex(RuntimeError, "injected"):
+                            with cp_shard_model_inputs(
+                                full, batch.positions, batch, batch.input_ids
+                            ):
+                                n = len(full[rank::size])
+                                torch.testing.assert_close(
+                                    batch.spec_info.hidden_states[:n], full[rank::size]
+                                )
+                                # Global MoE IDs are in rank-major order, with padding.
+                                physical = sum(
+                                    batch.attn_cp_metadata.per_rank_actual_token
+                                )
+                                padded = batch.input_ids.new_zeros(physical)
+                                padded[:9] = batch.input_ids
+                                expected = torch.cat(
+                                    [padded[r::size] for r in range(size)]
+                                )
+                                torch.testing.assert_close(
+                                    batch.input_ids_global, expected
+                                )
+                                raise RuntimeError("injected")
+                        self.assertIs(batch.spec_info.hidden_states, full)
+                        if had_global:
+                            self.assertIs(batch.input_ids_global, previous)
+                        else:
+                            self.assertFalse(hasattr(batch, "input_ids_global"))
+
+    def test_short_prompt_falls_back_from_cp(self):
+        with cp_context(4, 0, (1, 2), (0, 0)) as (_, batch):
+            self.assertFalse(is_cp_active(batch))
+
+    def test_runner_text_embedding_and_preembedded_paths(self):
+        for preembedded in (False, True):
+            for rank in range(4):
+                with (
+                    self.subTest(preembedded=preembedded, rank=rank),
+                    cp_context(4, rank) as (strategy, batch),
+                ):
+                    full = torch.arange(27, dtype=torch.float32).reshape(9, 3)
+                    embedding = Mock(return_value=full)
+                    model = NS(
+                        vision=None,
+                        _prepare_mm_embeddings=Mock(
+                            side_effect=AssertionError("Text must not invoke vision")
+                        ),
+                        get_input_embeddings=Mock(return_value=embedding),
+                        pp_group=NS(is_last_rank=True),
+                        lm_head=object(),
+                        capture_aux_hidden_states=False,
+                        logits_processor=Mock(return_value="ok"),
+                    )
+                    model.prepare_language_model_inputs = lambda ids, fb, emb: (
+                        DeepseekV4ForCausalLM.prepare_language_model_inputs(
+                            model, ids, fb, emb
+                        )
+                    )
+
+                    def body(ids, positions, fb, input_embeds):
+                        n = len(full[rank::4])
+                        torch.testing.assert_close(ids[:n], batch.input_ids[rank::4])
+                        torch.testing.assert_close(
+                            positions[:n], batch.positions[rank::4]
+                        )
+                        torch.testing.assert_close(input_embeds[:n], full[rank::4])
+                        return input_embeds
+
+                    model.model = body
+                    with (
+                        simulated_collective(strategy, batch, full),
+                        patch(RUNNER + ".torch.cuda.current_stream", return_value=None),
+                    ):
+                        result = EagerRunner._execute_extend_cp(
+                            NS(model_runner=NS(model=model)),
+                            batch,
+                            {"input_embeds": full} if preembedded else {},
+                        )
+                    self.assertEqual(result, "ok")
+                    if preembedded:
+                        model.get_input_embeddings.assert_not_called()
+                    else:
+                        embedding.assert_called_once_with(batch.input_ids)
+                    model._prepare_mm_embeddings.assert_not_called()
+                    args = model.logits_processor.call_args.args
+                    torch.testing.assert_close(args[0], batch.input_ids)
+                    torch.testing.assert_close(args[1], full)
+
+    def test_dspark_aux_tensor_and_list_gathered_without_pre_norm_override(self):
+        for as_list in (False, True):
+            with self.subTest(as_list=as_list), cp_context(4, 2) as (strategy, batch):
+                full = torch.arange(27, dtype=torch.float32).reshape(9, 3)
+                local = strategy.shard_hidden_states(full, batch)
+                aux = [local.clone(), local.clone()] if as_list else local.clone()
+                model = NS(
+                    get_input_embeddings=lambda: lambda ids: full,
+                    model=Mock(return_value=((local, local.clone()), aux)),
+                    capture_aux_hidden_states=True,
+                    pp_group=NS(is_last_rank=True),
+                    lm_head=object(),
+                    logits_processor=Mock(return_value="ok"),
+                )
+                with (
+                    simulated_collective(strategy, batch, full),
+                    patch(RUNNER + ".torch.cuda.current_stream", return_value=None),
+                ):
+                    EagerRunner._execute_extend_cp(
+                        NS(model_runner=NS(model=model)), batch, {}
+                    )
+                args, kwargs = model.logits_processor.call_args
+                torch.testing.assert_close(args[1], full)
+                for tensor in args[4] if as_list else [args[4]]:
+                    torch.testing.assert_close(tensor, full)
+                self.assertNotIn("hidden_states_before_norm", kwargs)
+
+    def test_target_hidden_states_before_norm_preserved_without_dspark_aux(self):
+        with cp_context(4, 1) as (strategy, batch):
+            full = torch.arange(27, dtype=torch.float32).reshape(9, 3)
+            local = strategy.shard_hidden_states(full, batch)
+            model = NS(
+                get_input_embeddings=lambda: lambda ids: full,
+                model=Mock(return_value=(local, local.clone())),
+                capture_aux_hidden_states=False,
+                pp_group=NS(is_last_rank=True),
+                lm_head=object(),
+                logits_processor=Mock(return_value="ok"),
+            )
+            with (
+                simulated_collective(strategy, batch, full),
+                patch(RUNNER + ".torch.cuda.current_stream", return_value=None),
+            ):
+                EagerRunner._execute_extend_cp(
+                    NS(model_runner=NS(model=model)), batch, {}
+                )
+            torch.testing.assert_close(
+                model.logits_processor.call_args.kwargs["hidden_states_before_norm"],
+                full,
+            )
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/unit/server_args/test_deepseek_v41_text_only_pd.py

@@ -0,0 +1,73 @@
+"""V4.1 language-model-only PD configuration validation."""
+
+import unittest
+from types import SimpleNamespace as NS
+from unittest.mock import patch
+
+from sglang.srt.arg_groups import model_hook
+from sglang.srt.server_args import ServerArgs
+from sglang.test.ci.ci_register import register_cpu_ci
+from sglang.test.test_utils import CustomTestCase
+
+register_cpu_ci(est_time=5, suite="base-a-test-cpu")
+
+
+class TestDSV41TextOnlyPDPolicy(CustomTestCase):
+    def validate(
+        self,
+        model_type="deepseek_v41",
+        mode="prefill",
+        arch="DeepseekV4ForCausalLM",
+        **flags,
+    ):
+        cfg = NS(
+            language_model_only=True,
+            encoder_only=False,
+            language_only=False,
+            enable_prefix_mm_cache=False,
+            enable_broadcast_mm_inputs_process=False,
+            mm_enable_dp_encoder=False,
+            disaggregation_mode=mode,
+        )
+        for name, value in flags.items():
+            setattr(cfg, name, value)
+        model = NS(hf_config=NS(model_type=model_type, architectures=[arch]))
+        args = NS(
+            LANGUAGE_MODEL_ONLY_ARCHITECTURES=ServerArgs.LANGUAGE_MODEL_ONLY_ARCHITECTURES
+        )
+        with (
+            patch.object(model_hook, "resolving_view", return_value=cfg),
+            patch.object(model_hook, "model_config_of", return_value=model),
+        ):
+            model_hook.handle_language_model_only(args)
+
+    def test_v41_modes(self):
+        for mode in ("null", "prefill", "decode"):
+            with self.subTest(mode=mode):
+                self.validate(mode=mode)
+
+    def test_other_models_still_reject_pd(self):
+        with self.assertRaisesRegex(ValueError, "incompatible"):
+            self.validate(model_type="cosmos3", arch="Cosmos3ForConditionalGeneration")
+
+    def test_encoder_options_still_rejected(self):
+        for flag in (
+            "encoder_only",
+            "language_only",
+            "enable_prefix_mm_cache",
+            "enable_broadcast_mm_inputs_process",
+            "mm_enable_dp_encoder",
+        ):
+            with (
+                self.subTest(flag=flag),
+                self.assertRaisesRegex(ValueError, "cannot be combined"),
+            ):
+                self.validate(**{flag: True})
+
+    def test_unknown_arch_rejected(self):
+        with self.assertRaisesRegex(ValueError, "does not support"):
+            self.validate(arch="UnknownArchitecture")
+
+
+if __name__ == "__main__":
+    unittest.main()