[PD] Add support for different TP sizes per DP rank (sgl-project#5922)

Signed-off-by: Shangming Cai <[email protected]>

Author: ShangmingCai

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

@@ -37,6 +37,7 @@ def __init__(
         args: KVArgs,
         disaggregation_mode: DisaggregationMode,
         server_args: ServerArgs,
+        is_mla_backend: Optional[bool] = False,
     ): ...
 
 

python/sglang/srt/disaggregation/decode.py

@@ -37,6 +37,7 @@
     ReqToMetadataIdxAllocator,
     TransferBackend,
     get_kv_class,
+    is_mla_backend,
     kv_to_page_indices,
     poll_and_all_reduce,
 )
@@ -86,6 +87,7 @@ def __init__(
         self.req_to_token_pool = req_to_token_pool
         self.token_to_kv_pool_allocator = token_to_kv_pool_allocator
         self.token_to_kv_pool = token_to_kv_pool_allocator.get_kvcache()
+        self.is_mla_backend = is_mla_backend(self.token_to_kv_pool)
         self.aux_dtype = aux_dtype
         self.metadata_buffers = metadata_buffers
         self.req_to_metadata_buffer_idx_allocator = req_to_metadata_buffer_idx_allocator
@@ -131,7 +133,10 @@ def _init_kv_manager(self) -> BaseKVManager:
         kv_args.gpu_id = self.scheduler.gpu_id
         kv_manager_class = get_kv_class(self.transfer_backend, KVClassType.MANAGER)
         kv_manager = kv_manager_class(
-            kv_args, DisaggregationMode.DECODE, self.scheduler.server_args
+            kv_args,
+            DisaggregationMode.DECODE,
+            self.scheduler.server_args,
+            self.is_mla_backend,
         )
         return kv_manager
 

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

@@ -83,6 +83,7 @@ def __init__(
         args: KVArgs,
         disaggregation_mode: DisaggregationMode,
         server_args: ServerArgs,
+        is_mla_backend: Optional[bool] = False,
     ):
         try:
             from nixl._api import nixl_agent

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

@@ -33,6 +33,7 @@
     ReqToMetadataIdxAllocator,
     TransferBackend,
     get_kv_class,
+    is_mla_backend,
     kv_to_page_indices,
     kv_to_page_num,
     poll_and_all_reduce,
@@ -68,6 +69,7 @@ def __init__(
         scheduler: Scheduler,
     ):
         self.token_to_kv_pool = token_to_kv_pool
+        self.is_mla_backend = is_mla_backend(token_to_kv_pool)
         self.aux_dtype = aux_dtype
 
         self.metadata_buffers = metadata_buffers
@@ -114,7 +116,7 @@ def _init_kv_manager(self) -> BaseKVManager:
             kv_args,
             DisaggregationMode.PREFILL,
             self.scheduler.server_args,
-            # self.scheduler.disagg_launch_done,
+            self.is_mla_backend,
         )
         return kv_manager
 

python/sglang/srt/disaggregation/prefill.py

@@ -105,3 +105,47 @@ def kv_to_page_indices(kv_indices: np.ndarray, page_size: int):
 def kv_to_page_num(num_kv_indices: int, page_size: int):
     # ceil(num_kv_indices / page_size)
     return (num_kv_indices + page_size - 1) // page_size
+
+
+@dataclasses.dataclass
+class PDRegistryRequest:
+    """A request to register a machine itself to the LB."""
+
+    mode: str
+    registry_url: str
+    bootstrap_port: Optional[int] = None
+
+    def __post_init__(self):
+        if self.mode == "prefill" and self.bootstrap_port is None:
+            raise ValueError("Bootstrap port must be set in PREFILL mode.")
+        elif self.mode == "decode" and self.bootstrap_port is not None:
+            raise ValueError("Bootstrap port must not be set in DECODE mode.")
+        elif self.mode not in ["prefill", "decode"]:
+            raise ValueError(
+                f"Invalid mode: {self.mode}. Must be 'prefill' or 'decode'."
+            )
+
+
+def register_disaggregation_server(
+    mode: str, server_port: int, bootstrap_port: int, pdlb_url: str
+):
+    boostrap_port = bootstrap_port if mode == "prefill" else None
+    registry_request = PDRegistryRequest(
+        mode=mode,
+        registry_url=f"http://{get_ip()}:{server_port}",
+        bootstrap_port=boostrap_port,
+    )
+    res = requests.post(
+        f"{pdlb_url}/register",
+        json=dataclasses.asdict(registry_request),
+    )
+    if res.status_code != 200:
+        warnings.warn(
+            f"Failed to register disaggregation server: {res.status_code} {res.text}"
+        )
+
+
+def is_mla_backend(target_kv_pool) -> bool:
+    from sglang.srt.mem_cache.memory_pool import MLATokenToKVPool
+
+    return isinstance(target_kv_pool, MLATokenToKVPool)

python/sglang/srt/disaggregation/utils.py

@@ -83,17 +83,26 @@ class TestFile:
         TestFile("test_triton_moe_channel_fp8_kernel.py", 25),
     ],
     "per-commit-2-gpu": [
-        TestFile("models/lora/test_lora_tp.py", 300),
-        TestFile("test_data_parallelism.py", 90),
-        TestFile("test_dp_attention.py", 90),
-        TestFile("test_expert_distribution.py", 100),
-        TestFile("test_eplb.py", 100),
-        TestFile("test_mla_tp.py", 420),
-        TestFile("test_moe_ep.py", 220),
-        TestFile("test_patch_torch.py", 30),
-        TestFile("test_update_weights_from_distributed.py", 100),
-        TestFile("test_verl_engine.py", 100),
-        TestFile("test_two_batch_overlap.py", 100),
+        TestFile("models/lora/test_lora_tp.py", 116),
+        TestFile("test_data_parallelism.py", 73),
+        TestFile("test_dp_attention.py", 137),
+        TestFile("test_mla_tp.py", 170),
+        TestFile("test_moe_ep.py", 181),
+        TestFile("test_patch_torch.py", 19),
+        TestFile("test_update_weights_from_distributed.py", 103),
+        TestFile("test_verl_engine.py", 64),
+    ],
+    "per-commit-8-gpu": [
+        # Disabled deepep tests temporarily because it takes too much time.
+        # TODO: re-enable them after reducing the test time with compilation cache and smaller models.
+        # TestFile("test_deepep_intranode.py", 50),
+        # TestFile("test_deepep_low_latency.py", 50),
+        # TestFile("test_moe_deepep_eval_accuracy_large.py", 250),
+        TestFile("test_disaggregation.py", 210),
+        TestFile("test_local_attn.py", 250),
+        TestFile("test_disaggregation_different_tp.py", 210),
+        TestFile("test_full_deepseek_v3.py", 250),
+        TestFile("test_pp_single_node.py", 150),
     ],
     "nightly": [
         TestFile("test_nightly_gsm8k_eval.py"),

test/srt/run_suite.py

@@ -0,0 +1,151 @@
+import os
+import subprocess
+import time
+import unittest
+from types import SimpleNamespace
+
+import requests
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
+from sglang.test.test_utils import (
+    DEFAULT_MODEL_NAME_FOR_TEST_MLA,
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    popen_launch_pd_server,
+    run_with_timeout,
+)
+
+
+class TestDisaggregationMooncakeDifferentTP(CustomTestCase):
+    @classmethod
+    def setUpClass(cls):
+        # Temporarily disable JIT DeepGEMM
+        cls.original_jit_deepgemm = os.environ.get("SGL_ENABLE_JIT_DEEPGEMM")
+        os.environ["SGL_ENABLE_JIT_DEEPGEMM"] = "false"
+
+        cls.model = DEFAULT_MODEL_NAME_FOR_TEST_MLA
+        cls.base_host = "127.0.0.1"
+        cls.base_port = int(DEFAULT_URL_FOR_TEST.split(":")[-1])
+        cls.lb_url = DEFAULT_URL_FOR_TEST
+        cls.prefill_url = f"http://{cls.base_host}:{cls.base_port + 100}"
+        cls.decode_url = f"http://{cls.base_host}:{cls.base_port + 200}"
+
+        run_with_timeout(cls.start_prefill, timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH)
+        run_with_timeout(cls.start_decode, timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH)
+
+        cls.wait_server_ready(cls.prefill_url + "/health")
+        cls.wait_server_ready(cls.decode_url + "/health")
+
+        lb_command = [
+            "python3",
+            "-m",
+            "sglang.srt.disaggregation.mini_lb",
+            "--prefill",
+            cls.prefill_url,
+            "--decode",
+            cls.decode_url,
+            "--host",
+            cls.base_host,
+            "--port",
+            str(cls.base_port),
+        ]
+
+        print("Starting load balancer:", " ".join(lb_command))
+        cls.process_lb = subprocess.Popen(
+            lb_command, stdout=subprocess.PIPE, stderr=subprocess.PIPE
+        )
+        cls.wait_server_ready(cls.lb_url + "/health")
+
+    @classmethod
+    def start_prefill(cls):
+        prefill_args = [
+            "--trust-remote-code",
+            "--disaggregation-mode",
+            "prefill",
+            "--host",
+            cls.base_host,
+            "--port",
+            str(cls.base_port + 100),
+            "--tp",
+            "4",
+        ]
+        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",
+            "--host",
+            cls.base_host,
+            "--port",
+            str(cls.base_port + 200),
+            "--tp",
+            "2",
+            "--base-gpu-id",
+            "4",
+        ]
+        cls.process_decode = popen_launch_pd_server(
+            cls.model,
+            cls.decode_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=decode_args,
+        )
+
+    @classmethod
+    def wait_server_ready(cls, url, timeout=60):
+        start_time = time.time()
+        while True:
+            try:
+                response = requests.get(url)
+                if response.status_code == 200:
+                    print(f"Server {url} is ready")
+                    return
+            except Exception:
+                pass
+
+            if time.time() - start_time > timeout:
+                raise RuntimeError(f"Server {url} failed to start in {timeout}s")
+            time.sleep(1)
+
+    @classmethod
+    def tearDownClass(cls):
+        # Restore JIT DeepGEMM environment variable
+        if cls.original_jit_deepgemm is not None:
+            os.environ["SGL_ENABLE_JIT_DEEPGEMM"] = cls.original_jit_deepgemm
+        else:
+            os.environ.pop("SGL_ENABLE_JIT_DEEPGEMM", None)
+
+        for process in [cls.process_lb, cls.process_decode, cls.process_prefill]:
+            if process:
+                try:
+                    kill_process_tree(process.pid)
+                except Exception as e:
+                    print(f"Error killing process {process.pid}: {e}")
+
+    def test_gsm8k(self):
+        args = SimpleNamespace(
+            num_shots=5,
+            data_path=None,
+            num_questions=200,
+            max_new_tokens=512,
+            parallel=128,
+            host="http://127.0.0.1",
+            port=int(self.lb_url.split(":")[-1]),
+        )
+        metrics = run_eval_few_shot_gsm8k(args)
+        print(f"Evaluation metrics: {metrics}")
+
+        self.assertGreater(metrics["accuracy"], 0.60)
+
+
+if __name__ == "__main__":
+    unittest.main()