[NPU] [Diffusion] support distributed inference pipeline for GLM-Image (#31320)

Co-authored-by: Xiaoyu Zhang <1182563586@qq.com>

docs/docs/sglang-diffusion/disaggregation.mdx

@@ -153,6 +153,118 @@ sglang serve --model-path ... --disagg-role server \
     --decoder-urls  "tcp://10.0.0.5:35000"
 ```
 
+### GLM-Image distributed mode
+
+GLM-Image can batch AR generation in the head and dispatch the resulting prior
+tokens to distributed denoiser workers. Each worker runs prompt/glyph preparation,
+DiT, and VAE decoding locally; no latent or embedding tensors are transferred.
+
+The following 16-device deployment uses devices 0-1 for the external AR server
+and devices 2-15 for 14 independent batch-1 Cache-DiT denoisers.
+
+```bash
+# Run 14 distributed denoisers
+DISAGG_SERVER="tcp://127.0.0.1:19655"
+MODEL_PATH="zai-org/GLM-Image"
+BASE_MASTER_PORT=29005
+
+export SGLANG_CACHE_DIT_FN=2
+export SGLANG_CACHE_DIT_BN=1
+export SGLANG_CACHE_DIT_WARMUP=4
+export SGLANG_CACHE_DIT_RDT=0.4
+export SGLANG_CACHE_DIT_MC=4
+export SGLANG_CACHE_DIT_TAYLORSEER=true
+export SGLANG_CACHE_DIT_TS_ORDER=2
+export SGLANG_CACHE_DIT_ENABLED=true
+
+worker_pids=()
+
+cleanup() {
+    trap - EXIT
+    if ((${#worker_pids[@]})); then
+        kill "${worker_pids[@]}" 2>/dev/null || true
+        wait "${worker_pids[@]}" 2>/dev/null || true
+    fi
+}
+
+trap cleanup EXIT
+trap 'exit 130' INT
+trap 'exit 143' TERM
+
+for i in $(seq 2 15); do
+    scheduler_port=$((19001 + i))
+    master_port=$((BASE_MASTER_PORT + i))
+
+    sglang serve \
+        --model-path "$MODEL_PATH" \
+        --disagg-role denoiser \
+        --disagg-server-addr "$DISAGG_SERVER" \
+        --srt-encoder-url http://127.0.0.1:30020 \
+        --scheduler-port "$scheduler_port" \
+        --master-port "$master_port" \
+        --num-gpus 1 \
+        --base-gpu-id "$i" \
+        --denoiser-sp 1 \
+        --cfg-parallel-size 1 \
+        --batching-max-size 1 \
+        --dit-cpu-offload false \
+        --attention-backend fa &
+    worker_pids+=("$!")
+done
+
+# Stop all denoisers if any worker exits or fails during startup.
+wait -n "${worker_pids[@]}"
+```
+
+Run the external AR server:
+
+```bash
+sglang serve \
+  --model-path ./zai-org/GLM-Image/vision_language_encoder/ \
+  --tokenizer-path ./zai-org/GLM-Image/processor/ \
+  --enable-multimodal \
+  --cuda-graph-max-bs 28 \
+  --device npu \
+  --attention-backend ascend \
+  --disable-fast-image-processor \
+  --tp-size 2 \
+  --host 0.0.0.0 \
+  --port 30020 \
+  --mem-fraction-static 0.8
+```
+
+Run the public head. `--encoder-urls` and `--decoder-urls` are intentionally
+omitted for this topology.
+
+```bash
+sglang serve \
+  --model-path zai-org/GLM-Image \
+  --disagg-role server \
+  --srt-encoder-url http://127.0.0.1:30020 \
+  --srt-encoder-timeout 300 \
+  --denoiser-urls "tcp://127.0.0.1:19003;tcp://127.0.0.1:19004;tcp://127.0.0.1:19005;tcp://127.0.0.1:19006;tcp://127.0.0.1:19007;tcp://127.0.0.1:19008;tcp://127.0.0.1:19009;tcp://127.0.0.1:19010;tcp://127.0.0.1:19011;tcp://127.0.0.1:19012;tcp://127.0.0.1:19013;tcp://127.0.0.1:19014;tcp://127.0.0.1:19015;tcp://127.0.0.1:19016" \
+  --batching-mode dynamic \
+  --batching-max-size 28 \
+  --batching-delay-ms 30 \
+  --enable-batching-metrics \
+  --host 0.0.0.0 \
+  --port 30052 \
+  --scheduler-port 19655 \
+  --output-path ./outputs
+```
+
+Workers return decoded pixels to the head, which saves and serves the final
+files from `--output-path`. The PR benchmark used
+[longtext-bench.zip](https://github.com/user-attachments/files/29779516/longtext-bench.zip):
+
+```bash
+python fetch_images.py \
+  --base-url http://localhost:30052/v1 \
+  --model GLM-Image-distributed-test \
+  --output-dir generated_images \
+  --max-concurrency 56
+```
+
 ## Port Convention
 
 Result endpoints are derived deterministically from the head node's `--scheduler-port` (default: 5555):
@@ -196,6 +308,9 @@ Tensor data between roles (encoder→denoiser, denoiser→decoder) is transferre
 
 **mooncake-transfer-engine** is required for disaggregated diffusion. It provides RDMA for direct GPU-to-GPU data movement.
 
+The GLM-Image distributed mode is an exception: it relays only prior token IDs
+and request metadata over ZMQ and does not require Mooncake.
+
 ```bash
 pip install mooncake-transfer-engine
 ```

python/sglang/multimodal_gen/runtime/disaggregation/orchestrator.py

@@ -7,9 +7,14 @@ import pickle
 import threading
 import time
 from collections import deque
-from dataclasses import dataclass
+from concurrent.futures import Future, ThreadPoolExecutor
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING
 
+import torch
 import zmq
+from transformers import AutoProcessor
+from zmq.utils.monitor import recv_monitor_message
 
 from sglang.multimodal_gen.runtime.disaggregation.dispatch_policy import (
     PoolDispatcher,
@@ -20,6 +25,7 @@ from sglang.multimodal_gen.runtime.disaggregation.request_state import (
 )
 from sglang.multimodal_gen.runtime.disaggregation.roles import RoleType
 from sglang.multimodal_gen.runtime.disaggregation.transport.codec import (
+    send_tensors,
     unpack_tensors,
 )
 from sglang.multimodal_gen.runtime.disaggregation.transport.protocol import (
@@ -31,11 +37,70 @@ from sglang.multimodal_gen.runtime.disaggregation.transport.protocol import (
     encode_transfer_msg,
     is_transfer_message,
 )
+from sglang.multimodal_gen.runtime.pipelines_core.schedule_batch import (
+    OutputBatch,
+    Req,
+)
 from sglang.multimodal_gen.runtime.utils.common import get_zmq_socket
+from sglang.multimodal_gen.runtime.utils.perf_logger import (
+    MemorySnapshot,
+    RequestMetrics,
+)
+
+if TYPE_CHECKING:
+    from sglang.multimodal_gen.runtime.pipelines_core.stages.model_specific_stages.glm_image import (
+        GlmImageAR,
+    )
+    from sglang.multimodal_gen.runtime.server_args import ServerArgs
 
 logger = logging.getLogger(__name__)
 
 
+def _deserialize_request_metrics(data: dict | None) -> RequestMetrics | None:
+    if data is None:
+        return None
+
+    metrics = RequestMetrics(request_id=data["request_id"])
+    metrics.stages = data.get("stages", {})
+    metrics.steps = data.get("steps", [])
+    metrics.total_duration_ms = data.get("total_duration_ms", 0.0)
+    for name, snapshot in data.get("memory_snapshots", {}).items():
+        metrics.memory_snapshots[name] = MemorySnapshot(
+            allocated_mb=snapshot.get("allocated_mb", 0.0),
+            reserved_mb=snapshot.get("reserved_mb", 0.0),
+            peak_allocated_mb=snapshot.get("peak_allocated_mb", 0.0),
+            peak_reserved_mb=snapshot.get("peak_reserved_mb", 0.0),
+            peak_host_anon_mb=snapshot.get("peak_host_anon_mb", 0.0),
+        )
+    return metrics
+
+
+@dataclass
+class _GlmDistributedRequest:
+    """Track one client request as it moves from AR to a denoiser."""
+
+    client_request_id: str
+    req: Req
+    enqueue_time: float
+    worker_idx: int | None = None
+
+
+@dataclass
+class _GlmDistributedModeState:
+    """State used only by the GLM external-AR distributed topology."""
+
+    server_args: "ServerArgs"
+    ar_stage: "GlmImageAR"
+    executor: ThreadPoolExecutor
+    denoiser_worker_available: list[bool]
+    pending_ar_requests: deque[_GlmDistributedRequest] = field(default_factory=deque)
+    pending_denoiser_requests: deque[_GlmDistributedRequest] = field(
+        default_factory=deque
+    )
+    denoiser_requests: dict[str, _GlmDistributedRequest] = field(default_factory=dict)
+    active_ar_batch: tuple[Future, list[_GlmDistributedRequest]] | None = None
+
+
 @dataclass
 class _EncoderTTAEntry:
     request_id: str
@@ -89,9 +154,11 @@ class DiffusionServer:
         dispatch_policy_name: str = "round_robin",
         timeout_s: float = 600.0,
         encoder_capacity: int = 4,
-        denoiser_capacity: int = 2,
+        denoiser_capacity_per_worker: int = 2,
         decoder_capacity: int = 4,
         p2p_mode: bool = True,
+        server_args=None,
+        glm_distributed_mode_enabled: bool = False,
     ):
         self._frontend_endpoint = frontend_endpoint
         self._encoder_work_endpoints = encoder_work_endpoints
@@ -108,9 +175,9 @@ class DiffusionServer:
 
         self._tracker = RequestTracker()
         self._dispatcher = PoolDispatcher(
-            num_encoders=self._num_encoders,
+            num_encoders=max(1, self._num_encoders),
             num_denoisers=self._num_denoisers,
-            num_decoders=self._num_decoders,
+            num_decoders=max(1, self._num_decoders),
             policy_name=dispatch_policy_name,
         )
 
@@ -124,7 +191,7 @@ class DiffusionServer:
 
         # FreeBufferSlots per instance
         self._encoder_free_slots = [encoder_capacity] * self._num_encoders
-        self._denoiser_free_slots = [denoiser_capacity] * self._num_denoisers
+        self._denoiser_free_slots = [denoiser_capacity_per_worker] * self._num_denoisers
         self._decoder_free_slots = [decoder_capacity] * self._num_decoders
 
         # TTA queues per role type
@@ -133,6 +200,24 @@ class DiffusionServer:
         self._decoder_tta: deque[_RoleTTAEntry] = deque()
 
         self._transfer_mode = p2p_mode
+        self._glm_distributed_state: _GlmDistributedModeState | None = None
+
+        if glm_distributed_mode_enabled:
+            from sglang.multimodal_gen.runtime.pipelines_core.stages.model_specific_stages.glm_image import (
+                GlmImageAR,
+            )
+
+            processor = AutoProcessor.from_pretrained(
+                server_args.model_path, subfolder="processor"
+            )
+            self._glm_distributed_state = _GlmDistributedModeState(
+                server_args=server_args,
+                ar_stage=GlmImageAR(processor=processor, vision_language_encoder=None),
+                executor=ThreadPoolExecutor(
+                    max_workers=1, thread_name_prefix="glm-distributed-ar"
+                ),
+                denoiser_worker_available=[False] * self._num_denoisers,
+            )
         self._transfer_state: dict[str, _TransferRequestState] = {}
 
         # Per-instance registration: instance_idx -> {session_id, pool_ptr, pool_size}
@@ -197,6 +282,10 @@ class DiffusionServer:
         if self._thread is not None:
             self._thread.join(timeout=5.0)
             self._thread = None
+        if self._glm_distributed_state is not None:
+            self._glm_distributed_state.executor.shutdown(
+                wait=False, cancel_futures=True
+            )
 
     def _event_loop(self) -> None:
         frontend, _ = get_zmq_socket(
@@ -209,9 +298,16 @@ class DiffusionServer:
             encoder_pushes.append(sock)
 
         denoiser_pushes: list[zmq.Socket] = []
+        denoiser_monitors: list[zmq.Socket] = []
         for i, ep in enumerate(self._denoiser_work_endpoints):
             sock, _ = get_zmq_socket(self._context, zmq.PUSH, ep, bind=False)
             denoiser_pushes.append(sock)
+            if self._glm_distributed_state is not None:
+                denoiser_monitors.append(
+                    sock.get_monitor_socket(
+                        events=zmq.EVENT_CONNECTED | zmq.EVENT_DISCONNECTED
+                    )
+                )
 
         decoder_pushes: list[zmq.Socket] = []
         for i, ep in enumerate(self._decoder_work_endpoints):
@@ -233,6 +329,8 @@ class DiffusionServer:
         poller.register(encoder_result_pull, zmq.POLLIN)
         poller.register(denoiser_result_pull, zmq.POLLIN)
         poller.register(decoder_result_pull, zmq.POLLIN)
+        for monitor in denoiser_monitors:
+            poller.register(monitor, zmq.POLLIN)
 
         self._encoder_pushes = encoder_pushes
         self._denoiser_pushes = denoiser_pushes
@@ -246,6 +344,7 @@ class DiffusionServer:
             + encoder_pushes
             + denoiser_pushes
             + decoder_pushes
+            + denoiser_monitors
         )
 
         try:
@@ -266,7 +365,16 @@ class DiffusionServer:
                 if decoder_result_pull in events:
                     self._handle_role_result(decoder_result_pull, RoleType.DECODER)
 
-                self._drain_all_queues()
+                for worker_idx, monitor in enumerate(denoiser_monitors):
+                    if monitor in events:
+                        self._handle_glm_denoiser_monitor_event(worker_idx, monitor)
+
+                if self._glm_distributed_state is not None:
+                    self._process_glm_ar_batch_result_if_ready()
+                    self._dispatch_glm_ar_batch_if_ready()
+                    self._dispatch_glm_denoiser_requests_if_ready()
+                else:
+                    self._drain_all_queues()
 
         except Exception:
             logger.exception("DiffusionServer event loop error")
@@ -285,7 +393,9 @@ class DiffusionServer:
             self._handle_transfer_result(frames, role)
             return
 
-        if role == RoleType.DECODER:
+        if self._glm_distributed_state is not None and role == RoleType.DENOISER:
+            self._handle_glm_denoiser_result_frames(frames)
+        elif role == RoleType.DECODER:
             self._handle_decoder_result_frames(frames)
         else:
             # Non-transfer frames from encoder/denoiser are error results
@@ -378,6 +488,22 @@ class DiffusionServer:
             self._tracker.transition(request_id, RequestState.ENCODER_WAITING)
         except ValueError:
             pass
+        if self._glm_distributed_state is not None:
+            if (
+                not isinstance(req.prompt, str)
+                or getattr(req, "image_path", None) is not None
+            ):
+                self._complete_with_error(
+                    request_id,
+                    "GLM distributed mode supports one text prompt without image input",
+                )
+                return
+            now = time.monotonic()
+            self._glm_distributed_state.pending_ar_requests.append(
+                _GlmDistributedRequest(request_id, req, now)
+            )
+            return
+
         self._encoder_tta.append(
             _EncoderTTAEntry(
                 request_id=request_id,
@@ -390,11 +516,245 @@ class DiffusionServer:
             request_id,
         )
 
-    def _handle_decoder_result_frames(self, frames: list) -> None:
+    def _dispatch_glm_ar_batch_if_ready(self) -> None:
-        from sglang.multimodal_gen.runtime.pipelines_core.schedule_batch import (
+        """Dispatch one compatible batch to the external AR server."""
-            OutputBatch,
+        state = self._glm_distributed_state
+        assert state is not None
+        if state.active_ar_batch is not None or not state.pending_ar_requests:
+            return
+
+        batch_max_size = (
+            max(1, state.server_args.batching_max_size)
+            if state.server_args.batching_mode == "dynamic"
+            else 1
+        )
+        base = state.pending_ar_requests[0]
+        indices = [0]
+        output_slots = max(1, int(base.req.num_outputs_per_prompt or 1))
+        for index in range(1, len(state.pending_ar_requests)):
+            if output_slots >= batch_max_size:
+                break
+            candidate = state.pending_ar_requests[index]
+            if (base.req.height, base.req.width) == (
+                candidate.req.height,
+                candidate.req.width,
+            ):
+                candidate_outputs = max(
+                    1, int(candidate.req.num_outputs_per_prompt or 1)
+                )
+                if output_slots + candidate_outputs > batch_max_size:
+                    continue
+                indices.append(index)
+                output_slots += candidate_outputs
+
+        waited = time.monotonic() - base.enqueue_time
+        batch_delay_s = state.server_args.batching_delay_ms / 1000.0
+        if output_slots < batch_max_size and waited < batch_delay_s:
+            return
+
+        requests = [state.pending_ar_requests[index] for index in indices]
+        for index in reversed(indices):
+            del state.pending_ar_requests[index]
+
+        for request in requests:
+            try:
+                self._tracker.transition(
+                    request.client_request_id, RequestState.ENCODER_RUNNING
+                )
+            except ValueError:
+                pass
+        state.active_ar_batch = (
+            state.executor.submit(
+                state.ar_stage.generate_and_assign_prior_tokens,
+                [request.req for request in requests],
+                state.server_args,
+                device=torch.device("cpu"),
+            ),
+            requests,
+        )
+        logger.info(
+            "GLM distributed AR dispatched batch size=%d requests, %d outputs",
+            len(requests),
+            output_slots,
         )
 
+    def _process_glm_ar_batch_result_if_ready(self) -> None:
+        """Move a completed AR batch into the denoiser dispatch queue."""
+        state = self._glm_distributed_state
+        assert state is not None
+        active_batch = state.active_ar_batch
+        if active_batch is None or not active_batch[0].done():
+            return
+        future, requests = active_batch
+        state.active_ar_batch = None
+        try:
+            future.result()
+        except Exception as error:
+            for request in requests:
+                self._complete_with_error(
+                    request.client_request_id, f"GLM AR error: {error}"
+                )
+            return
+
+        group_id = f"glm-distributed::{time.monotonic_ns()}"
+        for request_index, request in enumerate(requests):
+            if self._tracker.get(request.client_request_id) is None:
+                continue
+            try:
+                self._tracker.transition(
+                    request.client_request_id, RequestState.ENCODER_DONE
+                )
+                self._tracker.transition(
+                    request.client_request_id, RequestState.DENOISING_WAITING
+                )
+            except ValueError:
+                pass
+            denoiser_req = request.req
+            denoiser_req.request_id = f"{group_id}::request::{request_index}"
+            state.denoiser_requests[denoiser_req.request_id] = request
+            state.pending_denoiser_requests.append(request)
+
+    def _dispatch_glm_denoiser_requests_if_ready(self) -> None:
+        """Dispatch AR-complete requests to available denoisers.
+
+        Each GLM denoiser accepts one request at a time. Dispatch continues until
+        either the pending queue is empty or every connected worker is busy.
+        """
+        from sglang.multimodal_gen.runtime.disaggregation.scheduler_mixin import (
+            extract_transfer_fields,
+        )
+
+        state = self._glm_distributed_state
+        assert state is not None
+        while state.pending_denoiser_requests:
+            available_slots = [
+                slots if state.denoiser_worker_available[index] else 0
+                for index, slots in enumerate(self._denoiser_free_slots)
+            ]
+            worker_idx = self._dispatcher.select_denoiser_with_capacity(available_slots)
+            if worker_idx is None:
+                return
+            request = state.pending_denoiser_requests.popleft()
+            self._denoiser_free_slots[worker_idx] -= 1
+            request.worker_idx = worker_idx
+            tensor_fields, scalar_fields = extract_transfer_fields(request.req)
+            scalar_fields["request_id"] = request.req.request_id
+            send_tensors(
+                self._denoiser_pushes[worker_idx], tensor_fields, scalar_fields
+            )
+            try:
+                self._tracker.transition(
+                    request.client_request_id,
+                    RequestState.DENOISING_RUNNING,
+                    denoiser_instance=worker_idx,
+                )
+            except ValueError:
+                pass
+            logger.info(
+                "GLM distributed mode dispatched request with %d output(s) "
+                "to denoiser[%d]",
+                request.req.num_outputs_per_prompt,
+                worker_idx,
+            )
+
+    def _handle_glm_denoiser_monitor_event(
+        self, worker_idx: int, monitor: zmq.Socket
+    ) -> None:
+        """Update dispatch eligibility when a denoiser connects or disconnects."""
+        state = self._glm_distributed_state
+        assert state is not None
+        event = recv_monitor_message(monitor, flags=zmq.NOBLOCK)["event"]
+        if event == zmq.EVENT_DISCONNECTED:
+            state.denoiser_worker_available[worker_idx] = False
+            self._denoiser_free_slots[worker_idx] = 0
+            replayed_requests = [
+                request
+                for request in state.denoiser_requests.values()
+                if request.worker_idx == worker_idx
+            ]
+            for request in replayed_requests:
+                request.worker_idx = None
+                try:
+                    self._tracker.transition(
+                        request.client_request_id, RequestState.DENOISING_WAITING
+                    )
+                except ValueError:
+                    pass
+            state.pending_denoiser_requests.extendleft(reversed(replayed_requests))
+            logger.warning(
+                "GLM denoiser[%d] disconnected; requeued %d request(s)",
+                worker_idx,
+                len(replayed_requests),
+            )
+        elif event == zmq.EVENT_CONNECTED:
+            registered = worker_idx in self._denoiser_peers
+            state.denoiser_worker_available[worker_idx] = True
+            if not any(
+                request.worker_idx == worker_idx
+                for request in state.denoiser_requests.values()
+            ):
+                self._denoiser_free_slots[worker_idx] = 1
+            logger.info(
+                "GLM denoiser[%d] connected (registered=%s)",
+                worker_idx,
+                registered,
+            )
+
+    def _handle_glm_denoiser_result_frames(self, frames: list) -> None:
+        """Return decoded denoiser output to the originating HTTP request."""
+        state = self._glm_distributed_state
+        assert state is not None
+        tensor_fields, scalar_fields = unpack_tensors(frames, device="cpu")
+        denoiser_request_id = scalar_fields.get("request_id")
+        request = state.denoiser_requests.pop(denoiser_request_id, None)
+        if request is None:
+            logger.warning(
+                "Unknown GLM distributed denoiser result: %s", denoiser_request_id
+            )
+            return
+        if request.worker_idx is not None:
+            self._denoiser_free_slots[request.worker_idx] = 1
+
+        error = scalar_fields.get("error")
+        output = tensor_fields.get("output")
+        total = max(1, int(request.req.num_outputs_per_prompt or 1))
+        output_size = len(output) if output is not None else None
+        if output_size is not None and output_size != total:
+            error = (
+                f"GLM distributed output size mismatch: got {output_size}, "
+                f"expected {total}"
+            )
+            output = None
+        result = OutputBatch(
+            output=output,
+            error=error,
+            metrics=_deserialize_request_metrics(scalar_fields.get("metrics")),
+            metrics_list=[
+                _deserialize_request_metrics(metrics)
+                for metrics in scalar_fields.get("metrics_list", [])
+            ]
+            or None,
+            peak_memory_mb=scalar_fields.get("peak_memory_mb", 0.0),
+            usage=scalar_fields.get("usage"),
+        )
+        with self._lock:
+            identity = self._pending.pop(request.client_request_id, None)
+        if identity is not None:
+            self._frontend.send_multipart([identity, b"", pickle.dumps(result)])
+        try:
+            self._tracker.transition(
+                request.client_request_id, RequestState.DENOISING_DONE
+            )
+            self._tracker.transition(
+                request.client_request_id,
+                RequestState.FAILED if error else RequestState.DONE,
+                error=error,
+            )
+        except ValueError:
+            pass
+        self._tracker.remove(request.client_request_id)
+
+    def _handle_decoder_result_frames(self, frames: list) -> None:
         request_id = self._extract_request_id(frames)
         if request_id is None:
             logger.warning("DiffusionServer: decoder result missing request_id")
@@ -521,10 +881,6 @@ class DiffusionServer:
             return None
 
     def _complete_with_error(self, request_id: str, error_msg: str) -> None:
-        from sglang.multimodal_gen.runtime.pipelines_core.schedule_batch import (
-            OutputBatch,
-        )
-
         logger.error("DiffusionServer: %s — %s", request_id, error_msg)
 
         try:
@@ -578,6 +934,25 @@ class DiffusionServer:
             self._decoder_tta = deque(
                 e for e in self._decoder_tta if e.request_id not in timed_set
             )
+            if self._glm_distributed_state is not None:
+                state = self._glm_distributed_state
+                state.pending_ar_requests = deque(
+                    request
+                    for request in state.pending_ar_requests
+                    if request.client_request_id not in timed_set
+                )
+                state.pending_denoiser_requests = deque(
+                    request
+                    for request in state.pending_denoiser_requests
+                    if request.client_request_id not in timed_set
+                )
+                for denoiser_request_id, request in list(
+                    state.denoiser_requests.items()
+                ):
+                    if request.client_request_id in timed_set:
+                        state.denoiser_requests.pop(denoiser_request_id)
+                        if request.worker_idx is not None:
+                            self._denoiser_free_slots[request.worker_idx] = 1
 
     def _free_slot_for_record(self, record) -> None:
         if (
@@ -667,6 +1042,8 @@ class DiffusionServer:
         prealloc = msg.get("preallocated_slots", [])
         info["free_preallocated_slots"] = list(prealloc)
         peers[idx] = info
+        if role == RoleType.DENOISER and self._glm_distributed_state is not None:
+            self._glm_distributed_state.denoiser_worker_available[idx] = True
 
         logger.info(
             "DiffusionServer transfer: registered %s[%d] work_endpoint=%s "
@@ -1026,10 +1403,6 @@ class DiffusionServer:
             )
 
     def _transfer_return_to_client_from_msg(self, request_id: str, msg: dict) -> None:
-        from sglang.multimodal_gen.runtime.pipelines_core.schedule_batch import (
-            OutputBatch,
-        )
-
         with self._lock:
             client_identity = self._pending.pop(request_id, None)
 
@@ -1054,7 +1427,7 @@ class DiffusionServer:
     def get_stats(self) -> dict:
         with self._lock:
             pending_count = len(self._pending)
-        return {
+        stats = {
             "role": "diffusion_server",
             "transfer_mode": self._transfer_mode,
             "num_encoders": self._num_encoders,
@@ -1074,3 +1447,16 @@ class DiffusionServer:
             "decoder_peers": len(self._decoder_peers),
             "tracker": self._tracker.snapshot(),
         }
+        if self._glm_distributed_state is not None:
+            state = self._glm_distributed_state
+            stats.update(
+                {
+                    "glm_ar_queue_depth": len(state.pending_ar_requests),
+                    "glm_ar_in_flight": state.active_ar_batch is not None,
+                    "glm_denoiser_queue_depth": len(state.pending_denoiser_requests),
+                    "glm_denoiser_worker_available": list(
+                        state.denoiser_worker_available
+                    ),
+                }
+            )
+        return stats

python/sglang/multimodal_gen/runtime/disaggregation/request_state.py

@@ -45,10 +45,14 @@ _VALID_TRANSITIONS: dict[RequestState, set[RequestState]] = {
         RequestState.DENOISING_RUNNING,
     },
     RequestState.DENOISING_WAITING: {RequestState.DENOISING_RUNNING},
-    RequestState.DENOISING_RUNNING: {RequestState.DENOISING_DONE},
+    RequestState.DENOISING_RUNNING: {
+        RequestState.DENOISING_WAITING,
+        RequestState.DENOISING_DONE,
+    },
     RequestState.DENOISING_DONE: {
         RequestState.DECODER_WAITING,
         RequestState.DECODER_RUNNING,
+        RequestState.DONE,
     },
     RequestState.DECODER_WAITING: {RequestState.DECODER_RUNNING},
     RequestState.DECODER_RUNNING: {RequestState.DONE},

python/sglang/multimodal_gen/runtime/disaggregation/scheduler_mixin.py

@@ -30,6 +30,7 @@ from sglang.multimodal_gen.runtime.disaggregation.transport.buffer import (
 )
 from sglang.multimodal_gen.runtime.disaggregation.transport.codec import (
     send_tensors,
+    unpack_tensors,
 )
 from sglang.multimodal_gen.runtime.disaggregation.transport.engine import (
     create_transfer_engine,
@@ -49,6 +50,7 @@ from sglang.multimodal_gen.runtime.disaggregation.transport.protocol import (
     encode_transfer_msg,
     is_transfer_message,
 )
+from sglang.multimodal_gen.runtime.entrypoints.utils import expand_request_outputs
 from sglang.multimodal_gen.runtime.pipelines_core import Req
 from sglang.multimodal_gen.runtime.pipelines_core.diffusion_scheduler_utils import (
     clone_scheduler_runtime,
@@ -65,6 +67,44 @@ if TYPE_CHECKING:
 
 logger = init_logger(__name__)
 
+
+def _advertised_pool_work_endpoint(server_args) -> str:
+    host = server_args.disagg_p2p_hostname or server_args.host or "127.0.0.1"
+    if host == "0.0.0.0":
+        host = server_args.disagg_p2p_hostname or "127.0.0.1"
+    return server_args.pool_work_endpoint.replace("0.0.0.0", host)
+
+
+def _expand_glm_distributed_outputs(req: Req) -> list[Req]:
+    """Split external-AR tokens into requests for sequential denoising."""
+    output_count = max(1, int(req.num_outputs_per_prompt or 1))
+    if output_count == 1:
+        return [req]
+
+    prior_token_ids = req.prior_token_id
+    if not isinstance(prior_token_ids, torch.Tensor) or (
+        prior_token_ids.shape[0] != output_count
+    ):
+        actual_count = (
+            prior_token_ids.shape[0]
+            if isinstance(prior_token_ids, torch.Tensor)
+            else type(prior_token_ids).__name__
+        )
+        raise RuntimeError(
+            "Cannot split GLM-Image AR output for distributed inference: "
+            f"expected {output_count} token rows, got {actual_count}."
+        )
+
+    usage_by_output = req.extra.get("usage_by_output")
+    output_reqs = expand_request_outputs(req)
+    for output_index, output_req in enumerate(output_reqs):
+        output_req.prior_token_id = prior_token_ids[output_index : output_index + 1]
+        output_req.extra.pop("usage_by_output", None)
+        if usage_by_output is not None and output_index < len(usage_by_output):
+            output_req.usage = usage_by_output[output_index]
+    return output_reqs
+
+
 # ---------------------------------------------------------------------------
 # Field extraction: split Req into tensors (transfer buffer) and scalars (JSON)
 # ---------------------------------------------------------------------------
@@ -116,6 +156,9 @@ _SAMPLING_PARAMS_EXCLUDE_FIELDS = frozenset(
     }
 )
 
+# Receivers reconstruct base SamplingParams, so only base defaults can be omitted.
+_BASE_SAMPLING_PARAM_FIELDS = {f.name: f for f in dataclasses.fields(SamplingParams)}
+
 
 def _is_tensor_like(value) -> bool:
     if isinstance(value, torch.Tensor):
@@ -278,7 +321,8 @@ def extract_transfer_fields(req) -> tuple[dict, dict]:
             value = getattr(sp, name, None)
             if value is None:
                 continue
-            if _is_default(value, f):
+            base_field = _BASE_SAMPLING_PARAM_FIELDS.get(name)
+            if base_field is not None and _is_default(value, base_field):
                 continue
             try:
                 scalar_fields[name] = _to_json_serializable(value)
@@ -458,6 +502,23 @@ class SchedulerDisaggMixin:
 
         sa = self.server_args
 
+        if self._is_glm_distributed_mode():
+            self._preallocated_slots = {}
+            register_msg = TransferRegisterMsg(
+                role=self._disagg_role.value,
+                work_endpoint=_advertised_pool_work_endpoint(sa),
+            )
+            self._pool_result_push.send_multipart(encode_transfer_msg(register_msg))
+            self._compute_ready_queue = queue.Queue(maxsize=4)
+            self._recv_prefetch_thread = threading.Thread(
+                target=self._recv_prefetch_loop,
+                daemon=True,
+                name="recv-prefetch-glm-distributed-denoiser",
+            )
+            self._recv_prefetch_thread.start()
+            logger.info("GLM distributed denoiser registered")
+            return
+
         # Pool size: configurable, default 256 MiB
         pool_size = getattr(sa, "disagg_transfer_pool_size", 256 * 1024 * 1024)
 
@@ -522,7 +583,7 @@ class SchedulerDisaggMixin:
             session_id=self._transfer_manager.session_id,
             pool_ptr=self._transfer_manager.pool_data_ptr,
             pool_size=self._transfer_manager.pool_size,
-            work_endpoint=sa.pool_work_endpoint,
+            work_endpoint=_advertised_pool_work_endpoint(sa),
             preallocated_slots=preallocated_slot_info,
         )
         self._pool_result_push.send_multipart(encode_transfer_msg(register_msg))
@@ -623,6 +684,10 @@ class SchedulerDisaggMixin:
                 raw_frames = self._pool_work_pull.recv_multipart()
                 frames = [bytes(f) for f in raw_frames]
 
+                if not is_transfer_message(frames):
+                    self._compute_ready_queue.put(("relay_compute", frames))
+                    continue
+
                 msg = decode_transfer_msg(frames)
                 msg_type = msg.get("msg_type", "")
 
@@ -903,6 +968,18 @@ class SchedulerDisaggMixin:
                         self._broadcast_to_all_ranks(("skip",))
                     self._handle_transfer_msg(data)
 
+                elif msg_type == "relay_compute":
+                    local_device = (
+                        f"{current_platform.device_type}:{self.worker.local_rank}"
+                    )
+                    tensors, scalar_fields = unpack_tensors(data, device=local_device)
+                    request_id = scalar_fields.get("request_id", "unknown")
+                    req = self._build_disagg_req(scalar_fields, tensors)
+                    if is_multi_rank:
+                        self._broadcast_to_all_ranks(("compute",))
+                        self._broadcast_req_to_all_ranks(req)
+                    self._execute_glm_distributed_denoiser_request(req, request_id)
+
                 self._consecutive_error_count = 0
 
             except Exception as e:
@@ -1298,11 +1375,16 @@ class SchedulerDisaggMixin:
         (:meth:`_disagg_non_rank0_event_loop`).
         """
         if self._disagg_role == RoleType.DENOISER:
-            # Initialize scheduler timesteps (same as rank 0)
+            if not self._is_glm_distributed_mode():
-            _init_disagg_request_scheduler(self, req)
+                _init_disagg_request_scheduler(self, req)
 
             with self._disagg_trace_dispatch(req):
-                self.worker.execute_forward([req], return_req=True)
+                if self._is_glm_distributed_mode():
+                    req.save_output = False
+                    req.return_file_paths_only = False
+                    self.worker.execute_forward([req])
+                else:
+                    self.worker.execute_forward([req], return_req=True)
 
         elif self._disagg_role == RoleType.DECODER:
             req.save_output = False
@@ -1456,6 +1538,64 @@ class SchedulerDisaggMixin:
             duration_s,
         )
 
+    def _is_glm_distributed_mode(self: Scheduler) -> bool:
+        """Return whether this scheduler is a GLM distributed denoiser."""
+        return (
+            self._disagg_role == RoleType.DENOISER
+            and self.worker.pipeline.pipeline_name == "GlmImagePipeline"
+            and self.server_args.srt_encoder_url is not None
+        )
+
+    def _execute_glm_distributed_denoiser_request(
+        self: Scheduler, req: Req, request_id: str
+    ) -> None:
+        """Run local preparation, DiT, and VAE, then return decoded pixels."""
+        req.save_output = False
+        start_time = time.monotonic()
+        with self._disagg_trace_dispatch(req):
+            if (
+                self.server_args.pipeline_config.supports_sequential_multi_output_inference()
+                and max(1, int(req.num_outputs_per_prompt or 1)) > 1
+            ):
+                output_reqs = _expand_glm_distributed_outputs(req)
+                output_batches = list(
+                    self.worker.execute_forward_sequentially(output_reqs)
+                )
+                output_batch = self.worker._merge_expanded_output_batches(
+                    output_batches
+                )
+            else:
+                output_batch = self.worker.execute_forward([req])
+
+        tensor_fields = {}
+        scalar_fields = {"request_id": request_id}
+        if output_batch.output is not None:
+            tensor_fields["output"] = output_batch.output
+        if output_batch.error is not None:
+            scalar_fields["error"] = output_batch.error
+        if output_batch.usage is not None:
+            scalar_fields["usage"] = output_batch.usage
+        if output_batch.metrics is not None:
+            scalar_fields["metrics"] = output_batch.metrics.to_dict()
+        if output_batch.metrics_list is not None:
+            scalar_fields["metrics_list"] = [
+                metrics.to_dict() if metrics is not None else None
+                for metrics in output_batch.metrics_list
+            ]
+        scalar_fields["peak_memory_mb"] = output_batch.peak_memory_mb
+        send_tensors(self._pool_result_push, tensor_fields, scalar_fields)
+
+        if self._disagg_metrics:
+            if output_batch.error:
+                self._disagg_metrics.record_request_failed(request_id)
+            else:
+                self._disagg_metrics.record_request_complete(request_id)
+        logger.debug(
+            "GLM distributed denoiser: processed %s in %.2f s",
+            request_id,
+            time.monotonic() - start_time,
+        )
+
     def _disagg_decoder_compute(self: Scheduler, req: Req, request_id: str) -> None:
         """Run decoder compute in transfer mode, send result to DS.
 

python/sglang/multimodal_gen/runtime/launch_server.py

@@ -3,12 +3,9 @@
 import dataclasses
 import multiprocessing as mp
 import os
-import signal
 import sys
-import threading
 import time
 
-import psutil
 import uvicorn
 
 from sglang.multimodal_gen.runtime.disaggregation.orchestrator import (
@@ -24,7 +21,10 @@ from sglang.multimodal_gen.runtime.server_args import (
     prepare_server_args,
     set_global_server_args,
 )
-from sglang.multimodal_gen.runtime.utils.common import is_port_available
+from sglang.multimodal_gen.runtime.utils.common import (
+    is_port_available,
+    kill_process_tree,
+)
 from sglang.multimodal_gen.runtime.utils.logging_utils import configure_logger, logger
 from sglang.multimodal_gen.runtime.utils.trace_wrapper import init_diffusion_tracing
 from sglang.multimodal_gen.utils import kill_itself_when_parent_died
@@ -53,45 +53,6 @@ def _find_available_port(
     )
 
 
-def kill_process_tree(parent_pid, include_parent: bool = True, skip_pid: int = None):
-    """Kill the process and all its child processes."""
-    # Remove sigchld handler to avoid spammy logs.
-    if threading.current_thread() is threading.main_thread():
-        signal.signal(signal.SIGCHLD, signal.SIG_DFL)
-
-    if parent_pid is None:
-        parent_pid = os.getpid()
-        include_parent = False
-
-    try:
-        itself = psutil.Process(parent_pid)
-    except psutil.NoSuchProcess:
-        return
-
-    children = itself.children(recursive=True)
-    for child in children:
-        if child.pid == skip_pid:
-            continue
-        try:
-            child.kill()
-        except psutil.NoSuchProcess:
-            pass
-
-    if include_parent:
-        try:
-            if parent_pid == os.getpid():
-                itself.kill()
-                sys.exit(0)
-
-            itself.kill()
-
-            # Sometime processes cannot be killed with SIGKILL (e.g, PID=1 launched by kubernetes),
-            # so we send an additional signal to kill them.
-            itself.send_signal(signal.SIGQUIT)
-        except psutil.NoSuchProcess:
-            pass
-
-
 def _process_names(processes) -> str:
     return ", ".join(getattr(p, "name", repr(p)) for p in processes)
 
@@ -585,12 +546,12 @@ def launch_http_server_only(server_args):
     )
 
 
-def parse_url_string(url_str: str) -> list[str]:
+def parse_url_string(url_str: str | None) -> list[str]:
     """Parse a semicolon-separated URL string into a list.
 
     Example: "tcp://10.0.0.1:35000;tcp://10.0.0.2:35000" -> ["tcp://...", "tcp://..."]
     """
-    return [u.strip() for u in url_str.split(";") if u.strip()]
+    return [u.strip() for u in (url_str or "").split(";") if u.strip()]
 
 
 def launch_disagg_server(server_args: ServerArgs):
@@ -605,12 +566,23 @@ def launch_disagg_server(server_args: ServerArgs):
         decoder result: scheduler_port + 3
     """
     configure_logger(server_args)
+    set_global_server_args(server_args)
 
-    for name, val in [
+    glm_distributed_mode_enabled = (
-        ("--encoder-urls", server_args.encoder_urls),
+        type(server_args.pipeline_config).__name__ == "GlmImagePipelineConfig"
-        ("--denoiser-urls", server_args.denoiser_urls),
+        and server_args.srt_encoder_url is not None
-        ("--decoder-urls", server_args.decoder_urls),
+        and server_args.encoder_urls is None
-    ]:
+        and server_args.decoder_urls is None
+    )
+    required_urls = [("--denoiser-urls", server_args.denoiser_urls)]
+    if not glm_distributed_mode_enabled:
+        required_urls.extend(
+            [
+                ("--encoder-urls", server_args.encoder_urls),
+                ("--decoder-urls", server_args.decoder_urls),
+            ]
+        )
+    for name, val in required_urls:
         if val is None:
             raise ValueError(f"{name} is required for --disagg-role server")
 
@@ -644,6 +616,9 @@ def launch_disagg_server(server_args: ServerArgs):
         decoder_result_ep,
     )
 
+    denoiser_options = (
+        {"denoiser_capacity_per_worker": 1} if glm_distributed_mode_enabled else {}
+    )
     diffusion_server = DiffusionServer(
         frontend_endpoint=frontend_endpoint,
         encoder_work_endpoints=encoder_work_endpoints,
@@ -654,6 +629,9 @@ def launch_disagg_server(server_args: ServerArgs):
         decoder_result_endpoint=decoder_result_ep,
         dispatch_policy_name=server_args.disagg_dispatch_policy,
         timeout_s=float(server_args.disagg_timeout),
+        server_args=server_args,
+        glm_distributed_mode_enabled=glm_distributed_mode_enabled,
+        **denoiser_options,
     )
     diffusion_server.start()
 
@@ -726,6 +704,44 @@ def launch_disagg_role(server_args: ServerArgs):
         "ulysses_degree": role_par["ulysses_degree"],
         "ring_degree": role_par["ring_degree"],
     }
+    role_tp = role_par["tp_size"] or 1
+    role_sp = role_par["sp_degree"] or 1
+    cfg_degree = (
+        server_args.cfg_parallel_degree if server_args.enable_cfg_parallel else 1
+    )
+    cfg_parallel_explicit = server_args.is_arg_explicitly_set(
+        "enable_cfg_parallel"
+    ) or server_args.is_arg_explicitly_set("cfg_parallel_degree")
+    required_devices = role_tp * role_sp * cfg_degree * server_args.dp_size
+    if not cfg_parallel_explicit and (
+        required_devices > server_args.num_gpus
+        or server_args.num_gpus % required_devices != 0
+    ):
+        logger.warning(
+            "Disabling auto-enabled CFG parallel for %s role because tp=%d, "
+            "sp=%d, cfg=%d, dp=%d is incompatible with %d devices",
+            role_type.value,
+            role_tp,
+            role_sp,
+            cfg_degree,
+            server_args.dp_size,
+            server_args.num_gpus,
+        )
+        role_overrides["enable_cfg_parallel"] = False
+        role_overrides["cfg_parallel_degree"] = 1
+        cfg_degree = 1
+        required_devices = role_tp * role_sp * server_args.dp_size
+
+    if (
+        required_devices > server_args.num_gpus
+        or server_args.num_gpus % required_devices != 0
+    ):
+        raise ValueError(
+            f"Invalid parallelism for {role_type.value} role: "
+            f"tp={role_tp}, sp={role_sp}, cfg={cfg_degree}, "
+            f"dp={server_args.dp_size} requires groups of {required_devices} "
+            f"devices, but num_gpus={server_args.num_gpus}"
+        )
 
     base_dict = {
         f.name: getattr(server_args, f.name)

python/sglang/multimodal_gen/runtime/pipelines/glm_image.py

@@ -12,6 +12,22 @@ from sglang.multimodal_gen.runtime.pipelines_core.stages.model_specific_stages.g
 from sglang.multimodal_gen.runtime.server_args import ServerArgs
 
 
+class GlmImageDenoiserDecodingStage(GlmImageDecodingStage):
+    """Run VAE decoding on the denoiser because this topology has no decoder worker."""
+
+    @property
+    def role_affinity(self) -> RoleType:
+        return RoleType.DENOISER
+
+
+class GlmImageDenoiserPreparationStage(GlmImageBeforeDenoisingStage):
+    """Run DiT preparation on the denoiser because AR runs in the server head."""
+
+    @property
+    def role_affinity(self) -> RoleType:
+        return RoleType.DENOISER
+
+
 class GlmImagePipeline(LoRAPipeline, ComposedPipelineBase):
     pipeline_name = "GlmImagePipeline"
 
@@ -26,6 +42,10 @@ class GlmImagePipeline(LoRAPipeline, ComposedPipelineBase):
     ]
 
     def create_pipeline_stages(self, server_args: ServerArgs):
+        is_glm_distributed_mode = (
+            self._disagg_role == RoleType.DENOISER
+            and server_args.srt_encoder_url is not None
+        )
         self.add_stage(
             GlmImageAR(
                 processor=self.get_module("processor"),
@@ -34,8 +54,13 @@ class GlmImagePipeline(LoRAPipeline, ComposedPipelineBase):
             "glm_image_ar",
         )
 
+        before_denoising_stage_cls = (
+            GlmImageDenoiserPreparationStage
+            if is_glm_distributed_mode
+            else GlmImageBeforeDenoisingStage
+        )
         self.add_stage(
-            GlmImageBeforeDenoisingStage(
+            before_denoising_stage_cls(
                 vae=self.get_module("vae"),
                 text_encoder=self.get_module("text_encoder"),
                 tokenizer=self.get_module("tokenizer"),
@@ -52,14 +77,22 @@ class GlmImagePipeline(LoRAPipeline, ComposedPipelineBase):
             ),
         )
 
-        self.add_stage_factory(
+        if is_glm_distributed_mode:
-            RoleType.DECODER,
+            self.add_stage(
-            lambda: GlmImageDecodingStage(
+                GlmImageDenoiserDecodingStage(
-                vae=self.get_module("vae"),
+                    vae=self.get_module("vae"), pipeline=self
-                pipeline=self,
+                ),
-            ),
+                "decoding_stage",
-            "decoding_stage",
+            )
-        )
+        else:
+            self.add_stage_factory(
+                RoleType.DECODER,
+                lambda: GlmImageDecodingStage(
+                    vae=self.get_module("vae"),
+                    pipeline=self,
+                ),
+                "decoding_stage",
+            )
 
 
 EntryClass = [GlmImagePipeline]

python/sglang/multimodal_gen/runtime/pipelines_core/composed_pipeline_base.py

@@ -270,6 +270,12 @@ class ComposedPipelineBase(ABC):
         extra_allowed_modules = set(
             role_to_pipeline_modules.get(role, {}).get(self.pipeline_name, set())
         )
+        if (
+            role == RoleType.DENOISER
+            and self.pipeline_name == "GlmImagePipeline"
+            and getattr(self.server_args, "srt_encoder_url", None) is not None
+        ):
+            extra_allowed_modules.update({"text_encoder", "tokenizer", "vae"})
 
         if role == RoleType.DENOISER and task_name == "ti2v":
             if self.pipeline_name in {

python/sglang/multimodal_gen/runtime/pipelines_core/schedule_batch.py

@@ -108,6 +108,10 @@ class Req:
     pooled_embeds: list[torch.Tensor] = field(default_factory=list)
     neg_pooled_embeds: list[torch.Tensor] = field(default_factory=list)
 
+    # GLM-Image autoregressive prior tokens
+    prior_token_id: torch.Tensor | None = None
+    prior_token_image_ids: torch.Tensor | list[torch.Tensor] | None = None
+
     # Additional text-related parameters
     max_sequence_length: int | None = None
     prompt_template: dict[str, Any] | None = None

python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/glm_image.py

@@ -15,7 +15,6 @@ from sglang.multimodal_gen.configs.sample.glmimage import (
     GLM_IMAGE_RESOLUTION_ALIGNMENT,
     align_glm_image_resolution,
 )
-from sglang.multimodal_gen.runtime.distributed import get_local_torch_device
 from sglang.multimodal_gen.runtime.managers.forward_context import set_forward_context
 from sglang.multimodal_gen.runtime.managers.memory_managers.component_manager import (
     ComponentUse,
@@ -27,6 +26,7 @@ from sglang.multimodal_gen.runtime.pipelines_core.stages.base import (
     StageParallelismType,
 )
 from sglang.multimodal_gen.runtime.pipelines_core.stages.decoding import DecodingStage
+from sglang.multimodal_gen.runtime.platforms import current_platform
 from sglang.multimodal_gen.runtime.server_args import ServerArgs
 from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
 from sglang.multimodal_gen.runtime.utils.precision import (
@@ -414,7 +414,7 @@ class GlmImageAR(PipelineStage):
             Tuple of the D16 prior token IDs, optional source-image token IDs,
             and optional usage statistics returned by an external AR server.
         """
-        device = get_local_torch_device()
+        device = current_platform.get_local_torch_device()
         _validate_glm_image_resolution_alignment(width, height)
 
         is_text_to_image = image is None or len(image) == 0
@@ -516,8 +516,9 @@ class GlmImageAR(PipelineStage):
         height: int,
         width: int,
         server_args: ServerArgs,
+        device: Optional[torch.device] = None,
     ) -> tuple[list[torch.Tensor], list[dict[str, int] | None]]:
-        device = get_local_torch_device()
+        device = device or current_platform.get_local_torch_device()
         _validate_glm_image_resolution_alignment(width, height)
 
         input_ids = []
@@ -577,27 +578,15 @@ class GlmImageAR(PipelineStage):
             usages.append(_extract_srt_usage(item.get("meta_info")))
         return prior_token_ids, usages
 
-    def run_grouped_requests(
+    def generate_and_assign_prior_tokens(
         self,
         batches: list[Req],
         server_args: ServerArgs,
+        device: Optional[torch.device] = None,
     ) -> list[Req]:
-        can_batch_ar = (
+        """Generate one AR batch and assign its tokens and usage to each request."""
-            len(batches) > 1
-            and server_args.srt_encoder_url is not None
-            and all(
-                isinstance(batch.prompt, str) and batch.image_path is None
-                for batch in batches
-            )
-        )
-        if not can_batch_ar:
-            return super().run_grouped_requests(batches, server_args)
-
         height = batches[0].height
         width = batches[0].width
-        if any(batch.height != height or batch.width != width for batch in batches[1:]):
-            return super().run_grouped_requests(batches, server_args)
-
         start_time = time.time()
         output_counts = [_num_outputs_per_prompt(batch) for batch in batches]
         prompts = [
@@ -616,6 +605,7 @@ class GlmImageAR(PipelineStage):
             height=height,
             width=width,
             server_args=server_args,
+            device=device,
         )
         duration = time.time() - start_time
         logger.info(
@@ -642,6 +632,29 @@ class GlmImageAR(PipelineStage):
             output_offset += output_count
         return batches
 
+    def run_grouped_requests(
+        self,
+        batches: list[Req],
+        server_args: ServerArgs,
+    ) -> list[Req]:
+        can_batch_ar = (
+            len(batches) > 1
+            and server_args.srt_encoder_url is not None
+            and all(
+                isinstance(batch.prompt, str) and batch.image_path is None
+                for batch in batches
+            )
+        )
+        if not can_batch_ar:
+            return super().run_grouped_requests(batches, server_args)
+
+        height = batches[0].height
+        width = batches[0].width
+        if any(batch.height != height or batch.width != width for batch in batches[1:]):
+            return super().run_grouped_requests(batches, server_args)
+
+        return self.generate_and_assign_prior_tokens(batches, server_args)
+
     def iter_sequential_requests(
         self, batch: Req, server_args: ServerArgs
     ) -> Iterator[Req]:
@@ -714,7 +727,7 @@ class GlmImageAR(PipelineStage):
         else:
             ar_condition_images = None
 
-        device = get_local_torch_device()
+        device = current_platform.get_local_torch_device()
 
         if ar_condition_images is not None:
             height = height or ar_condition_images[0].height
@@ -1158,7 +1171,7 @@ class GlmImageBeforeDenoisingStage(PipelineStage):
         height = batch.height
         width = batch.width
 
-        device = get_local_torch_device()
+        device = current_platform.get_local_torch_device()
         batch_size = _num_outputs_per_prompt(batch)
         max_sequence_length = 1024
         seed = getattr(batch, "seed", None)

python/sglang/multimodal_gen/test/run_suite.py

@@ -39,6 +39,7 @@ if current_platform.is_npu():
         DEFAULT_STANDALONE_EST_TIME_SECONDS,
         FILE_SUITES,
         PARAMETRIZED_CASE_GROUPS,
+        STANDALONE_FILE_EST_TIMES,
         STANDALONE_FILES,
         STARTUP_OVERHEAD_SECONDS,
         SUITES,

python/sglang/multimodal_gen/test/server/ascend/test_glm_image_distributed.py

@@ -0,0 +1,315 @@
+"""NPU smoke test for the GLM-Image external-AR distributed topology."""
+
+from __future__ import annotations
+
+import base64
+import os
+import signal
+import subprocess
+import time
+import unittest
+from concurrent.futures import ThreadPoolExecutor
+from pathlib import Path
+
+import requests
+import torch
+
+from sglang.multimodal_gen.test.server.ascend.testcase_configs_npu import (
+    GLM_IMAGE_WEIGHTS_PATH,
+)
+from sglang.multimodal_gen.test.test_utils import find_free_port, wait_for_server_health
+from sglang.test.test_utils import CustomTestCase
+
+HOST = "127.0.0.1"
+_LOG_DIR = Path(os.environ.get("SGLANG_TEST_LOG_DIR", "/tmp"))
+_STARTUP_TIMEOUT_S = float(os.environ.get("SGLANG_GLM_AR_STARTUP_TIMEOUT", "600"))
+# A3 warm AR (26.711s) plus six sequential denoiser outputs (24.1945s each).
+_EXPECTED_MAKESPAN_S = 172.0
+_PERFORMANCE_TOLERANCE = 0.25
+
+
+def _kill_process_tree(proc: subprocess.Popen) -> None:
+    try:
+        os.killpg(os.getpgid(proc.pid), signal.SIGKILL)
+    except (ProcessLookupError, PermissionError):
+        pass
+
+
+def _tail_log(path: Path, lines: int = 80) -> str:
+    if not path.exists():
+        return f"<no log at {path}>"
+    try:
+        return "\n".join(path.read_text(errors="ignore").splitlines()[-lines:])
+    except OSError as error:
+        return f"<log read failed: {error}>"
+
+
+def _wait_for_log(path: Path, message: str, timeout: float) -> None:
+    deadline = time.monotonic() + timeout
+    while time.monotonic() < deadline:
+        if path.exists():
+            try:
+                if message in path.read_text(errors="ignore"):
+                    return
+            except OSError:
+                pass
+        time.sleep(2)
+    raise TimeoutError(f"Missing {message!r} in {path}:\n{_tail_log(path)}")
+
+
+class _GlmDistributedCluster:
+    def __init__(self) -> None:
+        self.model_path = Path(GLM_IMAGE_WEIGHTS_PATH)
+        self.ar_port = find_free_port(HOST)
+        self.denoiser_port = find_free_port(HOST)
+        self.head_port = find_free_port(HOST)
+        self.head_scheduler_port = find_free_port(HOST)
+        self.master_port = find_free_port(HOST)
+        self.processes: list[subprocess.Popen] = []
+        self.log_paths = {
+            "ar": _LOG_DIR / "glm_image_distributed_ar.log",
+            "denoiser": _LOG_DIR / "glm_image_distributed_denoiser.log",
+            "head": _LOG_DIR / "glm_image_distributed_head.log",
+        }
+        self._log_handles: list = []
+
+    def __enter__(self) -> _GlmDistributedCluster:
+        if not self.model_path.is_dir():
+            raise RuntimeError(
+                f"GLM-Image ModelScope cache is missing: {self.model_path}"
+            )
+        try:
+            self._start_ar()
+            self._start_denoiser()
+            self._start_head()
+        except Exception:
+            self.stop()
+            raise
+        return self
+
+    def __exit__(self, *exc) -> None:
+        self.stop()
+
+    def _start_process(self, command: list[str], log_name: str) -> None:
+        log_handle = open(self.log_paths[log_name], "w")
+        self._log_handles.append(log_handle)
+        self.processes.append(
+            subprocess.Popen(
+                command,
+                stdout=log_handle,
+                stderr=subprocess.STDOUT,
+                preexec_fn=os.setsid,
+                env=os.environ.copy(),
+            )
+        )
+
+    def _start_ar(self) -> None:
+        self._start_process(
+            [
+                "sglang",
+                "serve",
+                "--model-path",
+                str(self.model_path / "vision_language_encoder"),
+                "--tokenizer-path",
+                str(self.model_path / "processor"),
+                "--enable-multimodal",
+                "--device",
+                "npu",
+                "--attention-backend",
+                "ascend",
+                "--disable-fast-image-processor",
+                "--tp-size",
+                "1",
+                "--cuda-graph-bs",
+                "2",
+                "--base-gpu-id",
+                "0",
+                "--host",
+                HOST,
+                "--port",
+                str(self.ar_port),
+            ],
+            "ar",
+        )
+        self._wait_for_health("ar", self.ar_port)
+
+    def _start_denoiser(self) -> None:
+        self._start_process(
+            [
+                "sglang",
+                "serve",
+                "--model-path",
+                str(self.model_path),
+                "--disagg-role",
+                "denoiser",
+                "--disagg-server-addr",
+                f"tcp://{HOST}:{self.head_scheduler_port}",
+                "--srt-encoder-url",
+                f"http://{HOST}:{self.ar_port}",
+                "--scheduler-port",
+                str(self.denoiser_port),
+                "--master-port",
+                str(self.master_port),
+                "--num-gpus",
+                "1",
+                "--base-gpu-id",
+                "1",
+                "--denoiser-sp",
+                "1",
+                "--cfg-parallel-size",
+                "1",
+                "--batching-max-size",
+                "1",
+                "--dit-cpu-offload",
+                "false",
+                "--attention-backend",
+                "fa",
+            ],
+            "denoiser",
+        )
+
+    def _start_head(self) -> None:
+        self._start_process(
+            [
+                "sglang",
+                "serve",
+                "--model-path",
+                str(self.model_path),
+                "--disagg-role",
+                "server",
+                "--denoiser-urls",
+                f"tcp://{HOST}:{self.denoiser_port}",
+                "--srt-encoder-url",
+                f"http://{HOST}:{self.ar_port}",
+                "--batching-mode",
+                "dynamic",
+                "--batching-max-size",
+                "2",
+                "--batching-delay-ms",
+                "100",
+                "--scheduler-port",
+                str(self.head_scheduler_port),
+                "--host",
+                HOST,
+                "--port",
+                str(self.head_port),
+            ],
+            "head",
+        )
+        _wait_for_log(
+            self.log_paths["denoiser"], "Role DENOISER ready", _STARTUP_TIMEOUT_S
+        )
+        self._wait_for_health("head", self.head_port)
+
+    def _wait_for_health(self, name: str, port: int) -> None:
+        try:
+            wait_for_server_health(
+                f"http://{HOST}:{port}", path="/v1/models", timeout=_STARTUP_TIMEOUT_S
+            )
+        except Exception as error:
+            raise RuntimeError(
+                f"{name} failed to become healthy:\n{_tail_log(self.log_paths[name])}"
+            ) from error
+
+    def stop(self) -> None:
+        for process in self.processes:
+            _kill_process_tree(process)
+        for log_handle in self._log_handles:
+            log_handle.close()
+        self.processes.clear()
+        self._log_handles.clear()
+
+
+class TestGlmImageDistributedNpu(CustomTestCase):
+    @classmethod
+    def setUpClass(cls) -> None:
+        super().setUpClass()
+        if not hasattr(torch, "npu") or torch.npu.device_count() < 2:
+            raise unittest.SkipTest("requires two Ascend NPUs")
+        cls.cluster = _GlmDistributedCluster()
+        cls.cluster.__enter__()
+
+    @classmethod
+    def tearDownClass(cls) -> None:
+        if hasattr(cls, "cluster"):
+            for name, path in cls.cluster.log_paths.items():
+                print(f"\n=== [glm-image-distributed] {name} log tail ===")
+                print(_tail_log(path))
+            cls.cluster.stop()
+        super().tearDownClass()
+
+    def _generate(self, prompt: str, n: int = 1) -> list[bytes]:
+        response = requests.post(
+            f"http://{HOST}:{self.cluster.head_port}/v1/images/generations",
+            json={
+                "model": str(self.cluster.model_path),
+                "prompt": prompt,
+                "n": n,
+                "size": "1024x1024",
+                "response_format": "b64_json",
+            },
+            timeout=600,
+        )
+        response.raise_for_status()
+        images = response.json()["data"]
+        self.assertEqual(len(images), n)
+        return [base64.b64decode(image["b64_json"]) for image in images]
+
+    def test_external_ar_batching_multi_output_disaggregation_performance(
+        self,
+    ) -> None:
+        self._generate("A warmup landscape")
+        requests_to_generate = [
+            ("A mountain sunrise", 2),
+            ("A city at night", 1),
+            ("A forest lake", 2),
+            ("A desert sunset", 1),
+        ]
+
+        def generate(request: tuple[str, int]) -> tuple[list[bytes], float]:
+            prompt, output_count = request
+            start_time = time.perf_counter()
+            images = self._generate(prompt, n=output_count)
+            return images, time.perf_counter() - start_time
+
+        start_time = time.perf_counter()
+        with ThreadPoolExecutor(max_workers=4) as executor:
+            results = list(executor.map(generate, requests_to_generate))
+        makespan_s = time.perf_counter() - start_time
+        request_latencies_s = [latency_s for _, latency_s in results]
+        images = [image for request_images, _ in results for image in request_images]
+
+        print(
+            "GLM distributed performance: "
+            f"makespan={makespan_s:.2f}s, "
+            f"request_latencies={[round(value, 2) for value in request_latencies_s]}"
+        )
+        self.assertLessEqual(
+            makespan_s,
+            _EXPECTED_MAKESPAN_S * (1 + _PERFORMANCE_TOLERANCE),
+            "GLM external-AR and DiT overlap performance regressed",
+        )
+
+        self.assertTrue(
+            all(images),
+            "each requested output must produce a decoded image",
+        )
+        head_log = self.cluster.log_paths["head"].read_text(errors="ignore")
+        self.assertGreaterEqual(
+            head_log.count(
+                "GLM distributed AR dispatched batch size=2 requests, 2 outputs"
+            ),
+            1,
+            "the two n=1 requests must share an external-AR batch",
+        )
+        self.assertGreaterEqual(
+            head_log.count(
+                "GLM distributed AR dispatched batch size=1 requests, 2 outputs"
+            ),
+            2,
+            "each n=2 request must preserve both outputs through disaggregation",
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()

python/sglang/multimodal_gen/test/server/ascend/testcase_configs_npu.py

@@ -237,14 +237,26 @@ DEFAULT_EST_TIME_SECONDS = 300.0
 STARTUP_OVERHEAD_SECONDS = 120.0
 DEFAULT_STANDALONE_EST_TIME_SECONDS = 300.0
 
+STANDALONE_FILES = {
+    "2-npu": [
+        "ascend/test_glm_image_distributed.py",
+    ],
+}
+
+STANDALONE_FILE_EST_TIMES = {
+    "2-npu": {
+        "ascend/test_glm_image_distributed.py": 900.0,
+    },
+}
+
 SUITES = {
     "1-npu": [
         "ascend/test_server_1_npu.py",
-        # add new 1-npu test files here
+        *STANDALONE_FILES.get("1-npu", []),
     ],
     "2-npu": [
         "ascend/test_server_2_npu.py",
-        # add new 2-npu test files here
+        *STANDALONE_FILES.get("2-npu", []),
     ],
 }
 
@@ -258,6 +270,5 @@ PARAMETRIZED_CASE_GROUPS = {
 }
 
 FILE_SUITES = {}
-STANDALONE_FILES = {}
 COMPONENT_ACCURACY_SUITES = {}
 _UPDATE_WEIGHTS_FROM_DISK_TEST_FILE = None

python/sglang/multimodal_gen/test/unit/test_glm_image_ar.py

@@ -73,7 +73,7 @@ class TestGlmImageARSrtBackend(unittest.TestCase):
 
     @patch(
         "sglang.multimodal_gen.runtime.pipelines_core.stages."
-        "model_specific_stages.glm_image.get_local_torch_device",
+        "model_specific_stages.glm_image.current_platform.get_local_torch_device",
         return_value=torch.device("cpu"),
     )
     @patch(
@@ -102,7 +102,7 @@ class TestGlmImageARSrtBackend(unittest.TestCase):
 
     @patch(
         "sglang.multimodal_gen.runtime.pipelines_core.stages."
-        "model_specific_stages.glm_image.get_local_torch_device",
+        "model_specific_stages.glm_image.current_platform.get_local_torch_device",
         return_value=torch.device("cpu"),
     )
     @patch(
@@ -142,7 +142,7 @@ class TestGlmImageARSrtBackend(unittest.TestCase):
 
     @patch(
         "sglang.multimodal_gen.runtime.pipelines_core.stages."
-        "model_specific_stages.glm_image.get_local_torch_device",
+        "model_specific_stages.glm_image.current_platform.get_local_torch_device",
         return_value=torch.device("cpu"),
     )
     @patch(
@@ -182,7 +182,7 @@ class TestGlmImageARSrtBackend(unittest.TestCase):
 
     @patch(
         "sglang.multimodal_gen.runtime.pipelines_core.stages."
-        "model_specific_stages.glm_image.get_local_torch_device",
+        "model_specific_stages.glm_image.current_platform.get_local_torch_device",
         return_value=torch.device("cpu"),
     )
     @patch(
@@ -207,7 +207,7 @@ class TestGlmImageARSrtBackend(unittest.TestCase):
 
     @patch(
         "sglang.multimodal_gen.runtime.pipelines_core.stages."
-        "model_specific_stages.glm_image.get_local_torch_device",
+        "model_specific_stages.glm_image.current_platform.get_local_torch_device",
         return_value=torch.device("cpu"),
     )
     def test_forward_aligns_runtime_dimensions_before_ar_generation(self, _mock_device):
@@ -248,7 +248,7 @@ class TestGlmImageARSrtBackend(unittest.TestCase):
 
     @patch(
         "sglang.multimodal_gen.runtime.pipelines_core.stages."
-        "model_specific_stages.glm_image.get_local_torch_device",
+        "model_specific_stages.glm_image.current_platform.get_local_torch_device",
         return_value=torch.device("cpu"),
     )
     @patch(
@@ -282,7 +282,7 @@ class TestGlmImageARSrtBackend(unittest.TestCase):
 
     @patch(
         "sglang.multimodal_gen.runtime.pipelines_core.stages."
-        "model_specific_stages.glm_image.get_local_torch_device",
+        "model_specific_stages.glm_image.current_platform.get_local_torch_device",
         return_value=torch.device("cpu"),
     )
     @patch(
@@ -349,7 +349,7 @@ class TestGlmImageARSrtBackend(unittest.TestCase):
 
     @patch(
         "sglang.multimodal_gen.runtime.pipelines_core.stages."
-        "model_specific_stages.glm_image.get_local_torch_device",
+        "model_specific_stages.glm_image.current_platform.get_local_torch_device",
         return_value=torch.device("cpu"),
     )
     def test_generate_prior_tokens_rejects_unaligned_internal_dimensions(
@@ -370,7 +370,7 @@ class TestGlmImageARSrtBackend(unittest.TestCase):
 
     @patch(
         "sglang.multimodal_gen.runtime.pipelines_core.stages."
-        "model_specific_stages.glm_image.get_local_torch_device",
+        "model_specific_stages.glm_image.current_platform.get_local_torch_device",
         return_value=torch.device("cpu"),
     )
     def test_generate_prior_tokens_batch_rejects_unaligned_internal_dimensions(

python/sglang/multimodal_gen/test/unit/test_glm_image_multi_output.py

@@ -110,7 +110,9 @@ def test_ar_stage_generates_one_prior_per_requested_output():
     )
 
     with patch.object(
-        glm_stage, "get_local_torch_device", return_value=torch.device("cpu")
+        glm_stage.current_platform,
+        "get_local_torch_device",
+        return_value=torch.device("cpu"),
     ):
         result = stage.forward(batch, SimpleNamespace())
 
@@ -135,7 +137,9 @@ def test_before_denoising_expands_latents_and_conditions_for_requested_outputs()
     )
 
     with patch.object(
-        glm_stage, "get_local_torch_device", return_value=torch.device("cpu")
+        glm_stage.current_platform,
+        "get_local_torch_device",
+        return_value=torch.device("cpu"),
     ):
         result = stage.forward(batch, SimpleNamespace())