[PD] Support KV transfer with MORI-IO (#14626)

Co-authored-by: cwortman-amd <cwortman@amd.com>
2026-01-29 15:22:41 +08:00
parent d3cdee0a04
commit cbf90d70ff
6 changed files with 1519 additions and 1 deletions
--- a/docker/rocm.Dockerfile
+++ b/docker/rocm.Dockerfile
@@ -59,6 +59,16 @@ ARG TILELANG_COMMIT="ebf4a7cb8881432165ae8760e99d209d905c704a"
 ARG FHT_REPO="https://github.com/jeffdaily/fast-hadamard-transform.git"
 ARG FHT_BRANCH="rocm"
 ARG FHT_COMMIT="46efb7d776d38638fc39f3c803eaee3dd7016bd1"
+
+ARG ENABLE_MORI=0
+ARG NIC_BACKEND=none
+
+ARG MORI_REPO="https://github.com/ROCm/mori.git"
+ARG MORI_COMMIT="b0dce4beebeb1f26c784eee17d5fd9785ee9447f"
+
+# AMD AINIC apt repo settings
+ARG AINIC_VERSION=1.117.5
+ARG UBUNTU_CODENAME=jammy
 USER root

 # Install some basic utilities
@@ -283,6 +293,73 @@ RUN python3 -m pip install --no-cache-dir \
    py-spy \
    pre-commit

+# -----------------------
+# MORI (optional)
+RUN /bin/bash -lc 'set -euo pipefail; \
+  if [ "${ENABLE_MORI}" != "1" ]; then \
+    echo "[MORI] Skipping (ENABLE_MORI=${ENABLE_MORI})"; \
+    exit 0; \
+  fi; \
+  echo "[MORI] Enabling MORI (NIC_BACKEND=${NIC_BACKEND})"; \
+  \
+  # Base deps for MORI build
+  apt-get update && apt-get install -y --no-install-recommends \
+      build-essential \
+      g++ \
+      jq \
+      libopenmpi-dev \
+      libpci-dev \
+      initramfs-tools \
+  && rm -rf /var/lib/apt/lists/*; \
+  \
+  # NIC backend deps
+  case "${NIC_BACKEND}" in \
+    # default: mlx5
+    none) \
+      export USE_IONIC="OFF"; \
+      export USE_BNXT="OFF"; \
+      ;; \
+    # AMD NIC
+    ainic) \
+      export USE_IONIC="ON"; \
+      export USE_BNXT="OFF"; \
+      apt-get update && apt-get install -y --no-install-recommends ca-certificates curl gnupg apt-transport-https && \
+      rm -rf /var/lib/apt/lists/* && mkdir -p /etc/apt/keyrings; \
+      curl -fsSL https://repo.radeon.com/rocm/rocm.gpg.key | gpg --dearmor > /etc/apt/keyrings/amdainic.gpg; \
+      echo "deb [arch=amd64 signed-by=/etc/apt/keyrings/amdainic.gpg] https://repo.radeon.com/amdainic/pensando/ubuntu/${AINIC_VERSION} ${UBUNTU_CODENAME} main" \
+        > /etc/apt/sources.list.d/amdainic.list; \
+      apt-get update && apt-get install -y --no-install-recommends \
+          libionic-dev \
+          ionic-common \
+      ; \
+      rm -rf /var/lib/apt/lists/*; \
+      ;; \
+    # TODO: Add Broadcom bnxt packages/repos here later.
+    # bnxt) \
+    #   export USE_IONIC="OFF"; \
+    #   export USE_BNXT="ON"; \
+    #   echo "[MORI] NIC_BACKEND=bnxt: USE_BNXT=ON. Add Broadcom bnxt packages/repos here later."; \
+    #   ;; \
+    *) \
+      echo "ERROR: unknown NIC_BACKEND=${NIC_BACKEND}. Use one of: none, ainic"; \
+      exit 2; \
+      ;; \
+  esac; \
+  \
+  # Build/install MORI
+  export MORI_GPU_ARCHS="${GPU_ARCH_LIST}"; \
+  echo "[MORI] MORI_GPU_ARCHS=${MORI_GPU_ARCHS} USE_IONIC=${USE_IONIC} USE_BNXT=${USE_BNXT}"; \
+  rm -rf /sgl-workspace/mori; \
+  git clone "${MORI_REPO}" /sgl-workspace/mori; \
+  cd /sgl-workspace/mori; \
+  git checkout "${MORI_COMMIT}"; \
+  git submodule update --init --recursive; \
+  python3 setup.py develop; \
+  python3 -c "import os, torch; print(os.path.join(os.path.dirname(torch.__file__), \"lib\"))" > /etc/ld.so.conf.d/torch.conf; \
+  ldconfig; \
+  echo "export PYTHONPATH=/sgl-workspace/mori:\${PYTHONPATH}" >> /etc/bash.bashrc; \
+  echo "[MORI] Done."'
+
 # -----------------------
 # Performance environment variable.

--- a/python/sglang/srt/disaggregation/mori/init.py
+++ b/python/sglang/srt/disaggregation/mori/init.py
@@ -0,0 +1,6 @@
+from sglang.srt.disaggregation.mori.conn import (
+    MoriKVBootstrapServer,
+    MoriKVManager,
+    MoriKVReceiver,
+    MoriKVSender,
+)
--- a/python/sglang/srt/disaggregation/mori/conn.py
+++ b/python/sglang/srt/disaggregation/mori/conn.py
--- a/python/sglang/srt/disaggregation/utils.py
+++ b/python/sglang/srt/disaggregation/utils.py
@@ -231,6 +231,7 @@ class MetadataBuffers:

 class TransferBackend(Enum):
    MOONCAKE = "mooncake"
+    MORI = "mori"
    NIXL = "nixl"
    ASCEND = "ascend"
    FAKE = "fake"
@@ -266,6 +267,23 @@ def get_kv_class(
            KVClassType.BOOTSTRAP_SERVER: MooncakeKVBootstrapServer,
        }
        return class_mapping.get(class_type)
+    elif transfer_backend == TransferBackend.MORI:
+        from sglang.srt.disaggregation.base import KVArgs
+        from sglang.srt.disaggregation.mori import (
+            MoriKVBootstrapServer,
+            MoriKVManager,
+            MoriKVReceiver,
+            MoriKVSender,
+        )
+
+        class_mapping = {
+            KVClassType.KVARGS: KVArgs,
+            KVClassType.MANAGER: MoriKVManager,
+            KVClassType.SENDER: MoriKVSender,
+            KVClassType.RECEIVER: (MoriKVReceiver),
+            KVClassType.BOOTSTRAP_SERVER: MoriKVBootstrapServer,
+        }
+        return class_mapping.get(class_type)
    elif transfer_backend == TransferBackend.ASCEND:
        from sglang.srt.disaggregation.ascend import (
            AscendKVBootstrapServer,
--- a/python/sglang/srt/server_args.py
+++ b/python/sglang/srt/server_args.py
@@ -143,7 +143,7 @@ ATTENTION_BACKEND_CHOICES = [

 LORA_BACKEND_CHOICES = ["triton", "csgmv", "ascend", "torch_native"]

-DISAGG_TRANSFER_BACKEND_CHOICES = ["mooncake", "nixl", "ascend", "fake"]
+DISAGG_TRANSFER_BACKEND_CHOICES = ["mooncake", "nixl", "ascend", "fake", "mori"]

 ENCODER_TRANSFER_BACKEND_CHOICES = ["zmq_to_scheduler", "zmq_to_tokenizer", "mooncake"]

--- a/test/manual/test_mori_transfer_engine_e2e.py
+++ b/test/manual/test_mori_transfer_engine_e2e.py
@@ -0,0 +1,281 @@
+import os
+import subprocess
+import unittest
+
+import requests
+
+from sglang.test.server_fixtures.disaggregation_fixture import (
+    PDDisaggregationServerBase,
+)
+from sglang.test.test_utils import (
+    DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    popen_launch_pd_server,
+)
+
+
+class TestMoriTransferEngineE2E(PDDisaggregationServerBase):
+    """
+    Run:
+        SGLANG_MORI_MANUAL_E2E=1 python3 test/manual/test_mori_transfer_engine_e2e.py
+
+    Optional:
+    - SGLANG_MORI_E2E_TEST_MODEL: override model (defaults to a small test model)
+    - SGLANG_TEST_PD_DISAGG_DEVICES: RDMA devices string, e.g. "mlx5_roce0,mlx5_roce4"
+    """
+
+    @classmethod
+    def setUpClass(cls):
+        if os.environ.get("SGLANG_MORI_MANUAL_E2E", "") not in ("1", "true", "True"):
+            raise unittest.SkipTest(
+                "Set SGLANG_MORI_MANUAL_E2E=1 to run this manual MORI E2E test."
+            )
+
+        try:
+            import torch
+
+            if not torch.cuda.is_available():
+                raise unittest.SkipTest("torch.cuda is not available.")
+        except Exception as e:
+            raise unittest.SkipTest(f"torch is not available/usable: {e}")
+
+        # Force the disaggregation fixture to use MORI backend in local/manual runs.
+        os.environ["SGLANG_TEST_PD_DISAGG_BACKEND"] = "mori"
+
+        super().setUpClass()
+
+        cls.model = os.environ.get(
+            "SGLANG_MORI_E2E_TEST_MODEL", DEFAULT_SMALL_MODEL_NAME_FOR_TEST
+        )
+
+        cls.start_prefill()
+        cls.start_decode()
+
+        cls.wait_server_ready(
+            cls.prefill_url + "/health", timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
+        )
+        cls.wait_server_ready(
+            cls.decode_url + "/health", timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
+        )
+
+        cls.launch_lb()
+
+    @classmethod
+    def tearDownClass(cls):
+        os.environ.pop("SGLANG_TEST_PD_DISAGG_BACKEND", None)
+        super().tearDownClass()
+
+    @classmethod
+    def launch_lb(cls):
+        lb_command = [
+            "python3",
+            "-m",
+            "sglang_router.launch_router",
+            "--pd-disaggregation",
+            "--mini-lb",
+            "--prefill",
+            cls.prefill_url,
+            "--decode",
+            cls.decode_url,
+            "--host",
+            cls.base_host,
+            "--port",
+            cls.lb_port,
+        ]
+        print("Starting load balancer:", " ".join(lb_command))
+        cls.process_lb = subprocess.Popen(lb_command, stdout=None, stderr=None)
+        cls.wait_server_ready(
+            cls.lb_url + "/health", timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
+        )
+
+    @classmethod
+    def start_prefill(cls):
+        prefill_args = [
+            "--trust-remote-code",
+            "--disaggregation-mode",
+            "prefill",
+            "--tp",
+            "1",
+        ]
+        prefill_args += cls.transfer_backend + cls.rdma_devices
+        cls.process_prefill = popen_launch_pd_server(
+            cls.model,
+            cls.prefill_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=prefill_args,
+        )
+
+    @classmethod
+    def start_decode(cls):
+        decode_args = [
+            "--trust-remote-code",
+            "--disaggregation-mode",
+            "decode",
+            "--tp",
+            "1",
+            "--base-gpu-id",
+            "1",
+        ]
+        decode_args += cls.transfer_backend + cls.rdma_devices
+        cls.process_decode = popen_launch_pd_server(
+            cls.model,
+            cls.decode_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=decode_args,
+        )
+
+    def test_generate_smoke(self):
+        resp = requests.post(
+            self.lb_url + "/generate",
+            json={
+                "text": "Hello",
+                "sampling_params": {"temperature": 0, "max_new_tokens": 8},
+            },
+            timeout=120,
+        )
+        self.assertEqual(resp.status_code, 200, resp.text)
+        out = resp.json()
+        self.assertIn("text", out)
+        self.assertIsInstance(out["text"], str)
+        self.assertGreater(len(out["text"]), 0)
+
+
+class TestMoriTransferEngineTPMismatchE2E(PDDisaggregationServerBase):
+    """Manual MORI PD-disaggregation E2E with TP mismatch.
+
+    Scenario:
+    - prefill: tp=2 (GPU 0-1)
+    - decode:  tp=4 (GPU 2-5)
+
+    Manual-only and requires >= 6 visible GPUs.
+    """
+
+    _PORT_DELTA = 10
+
+    @classmethod
+    def setUpClass(cls):
+        if os.environ.get("SGLANG_MORI_MANUAL_E2E", "") not in ("1", "true", "True"):
+            raise unittest.SkipTest(
+                "Set SGLANG_MORI_MANUAL_E2E=1 to run this manual MORI E2E test."
+            )
+
+        try:
+            import torch
+
+            if not torch.cuda.is_available():
+                raise unittest.SkipTest("torch.cuda is not available.")
+            if torch.cuda.device_count() < 6:
+                raise unittest.SkipTest(
+                    "TP-mismatch test requires >= 6 visible GPUs (prefill tp=2 + decode tp=4)."
+                )
+        except Exception as e:
+            raise unittest.SkipTest(f"torch is not available/usable: {e}")
+
+        os.environ["SGLANG_TEST_PD_DISAGG_BACKEND"] = "mori"
+        super().setUpClass()
+
+        # Shift ports to avoid clashing with TestMoriTransferEngineE2E.
+        cls.lb_port = str(int(cls.lb_port) + cls._PORT_DELTA)
+        cls.prefill_port = str(int(cls.prefill_port) + cls._PORT_DELTA)
+        cls.decode_port = str(int(cls.decode_port) + cls._PORT_DELTA)
+        cls.prefill_url = f"http://{cls.base_host}:{cls.prefill_port}"
+        cls.decode_url = f"http://{cls.base_host}:{cls.decode_port}"
+        cls.lb_url = f"http://{cls.base_host}:{cls.lb_port}"
+
+        cls.model = os.environ.get(
+            "SGLANG_MORI_E2E_TEST_MODEL", DEFAULT_SMALL_MODEL_NAME_FOR_TEST
+        )
+
+        cls.start_prefill()
+        cls.start_decode()
+
+        cls.wait_server_ready(
+            cls.prefill_url + "/health", timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
+        )
+        cls.wait_server_ready(
+            cls.decode_url + "/health", timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
+        )
+        cls.launch_lb()
+
+    @classmethod
+    def tearDownClass(cls):
+        os.environ.pop("SGLANG_TEST_PD_DISAGG_BACKEND", None)
+        super().tearDownClass()
+
+    @classmethod
+    def launch_lb(cls):
+        lb_command = [
+            "python3",
+            "-m",
+            "sglang_router.launch_router",
+            "--pd-disaggregation",
+            "--mini-lb",
+            "--prefill",
+            cls.prefill_url,
+            "--decode",
+            cls.decode_url,
+            "--host",
+            cls.base_host,
+            "--port",
+            cls.lb_port,
+        ]
+        print("Starting load balancer:", " ".join(lb_command))
+        cls.process_lb = subprocess.Popen(lb_command, stdout=None, stderr=None)
+        cls.wait_server_ready(
+            cls.lb_url + "/health", timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
+        )
+
+    @classmethod
+    def start_prefill(cls):
+        prefill_args = [
+            "--trust-remote-code",
+            "--disaggregation-mode",
+            "prefill",
+            "--tp",
+            "2",
+        ]
+        prefill_args += cls.transfer_backend + cls.rdma_devices
+        cls.process_prefill = popen_launch_pd_server(
+            cls.model,
+            cls.prefill_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=prefill_args,
+        )
+
+    @classmethod
+    def start_decode(cls):
+        decode_args = [
+            "--trust-remote-code",
+            "--disaggregation-mode",
+            "decode",
+            "--tp",
+            "4",
+            "--base-gpu-id",
+            "2",
+        ]
+        decode_args += cls.transfer_backend + cls.rdma_devices
+        cls.process_decode = popen_launch_pd_server(
+            cls.model,
+            cls.decode_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=decode_args,
+        )
+
+    def test_generate_smoke_tp_mismatch(self):
+        resp = requests.post(
+            self.lb_url + "/generate",
+            json={
+                "text": "Hello",
+                "sampling_params": {"temperature": 0, "max_new_tokens": 8},
+            },
+            timeout=120,
+        )
+        self.assertEqual(resp.status_code, 200, resp.text)
+        out = resp.json()
+        self.assertIn("text", out)
+        self.assertIsInstance(out["text"], str)
+        self.assertGreater(len(out["text"]), 0)
+
+
+if __name__ == "__main__":
+    unittest.main()