[PD] support spec decode (#6507)

Co-authored-by: SangBin Cho <rkooo567@gmail.com>

.pre-commit-config.yaml

@@ -23,7 +23,7 @@ repos:
     hooks:
       - id: isort
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.11.2
+    rev: v0.11.7
     hooks:
       - id: ruff
         args: [--select=F401, --fixable=F401]

python/sglang/srt/disaggregation/decode.py

@@ -47,7 +47,7 @@ from sglang.srt.disaggregation.utils import (
 from sglang.srt.managers.schedule_batch import FINISH_ABORT
 from sglang.srt.mem_cache.base_prefix_cache import BasePrefixCache
 from sglang.srt.mem_cache.memory_pool import ReqToTokenPool, TokenToKVPoolAllocator
-from sglang.srt.model_executor.forward_batch_info import ForwardMode
+from sglang.srt.model_executor.forward_batch_info import CaptureHiddenMode, ForwardMode
 from sglang.srt.sampling.sampling_batch_info import SamplingBatchInfo
 
 logger = logging.getLogger(__name__)
@@ -76,6 +76,7 @@ class DecodePreallocQueue:
         self,
         req_to_token_pool: ReqToTokenPool,
         token_to_kv_pool_allocator: TokenToKVPoolAllocator,
+        draft_token_to_kv_pool: Optional[KVCache],
         req_to_metadata_buffer_idx_allocator: ReqToMetadataIdxAllocator,
         metadata_buffers: List[torch.Tensor],
         aux_dtype: torch.dtype,
@@ -91,6 +92,7 @@ class DecodePreallocQueue:
         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.draft_token_to_kv_pool = draft_token_to_kv_pool
         self.is_mla_backend = is_mla_backend(self.token_to_kv_pool)
         self.aux_dtype = aux_dtype
         self.metadata_buffers = metadata_buffers
@@ -119,6 +121,14 @@ class DecodePreallocQueue:
             self.token_to_kv_pool.get_contiguous_buf_infos()
         )
 
+        if self.draft_token_to_kv_pool is not None:
+            draft_kv_data_ptrs, draft_kv_data_lens, draft_kv_item_lens = (
+                self.draft_token_to_kv_pool.get_contiguous_buf_infos()
+            )
+            kv_data_ptrs += draft_kv_data_ptrs
+            kv_data_lens += draft_kv_data_lens
+            kv_item_lens += draft_kv_item_lens
+
         kv_args.kv_data_ptrs = kv_data_ptrs
         kv_args.kv_data_lens = kv_data_lens
         kv_args.kv_item_lens = kv_item_lens

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

@@ -51,6 +51,7 @@ def group_concurrent_contiguous(
     return src_groups, dst_groups
 
 
+# prefill
 @dataclasses.dataclass
 class TransferKVChunk:
     room: int
@@ -60,6 +61,7 @@ class TransferKVChunk:
     prefill_aux_index: Optional[int]
 
 
+# decode
 @dataclasses.dataclass
 class TransferInfo:
     room: int
@@ -93,6 +95,7 @@ class TransferInfo:
         )
 
 
+# decode
 @dataclasses.dataclass
 class KVArgsRegisterInfo:
     room: str

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

@@ -61,7 +61,8 @@ class MooncakeTransferEngine:
         self, session_id: str, buffer: int, peer_buffer_address: int, length: int
     ) -> int:
         """Synchronously transfer data to the specified address."""
-
+        # the first time: based on session_id (which contains remote_ip) to construct a queue pair, and cache the queue pair
+        # later: based on the cached queue pair to send data
         ret = self.engine.transfer_sync_write(
             session_id, buffer, peer_buffer_address, length
         )

python/sglang/srt/disaggregation/prefill.py

@@ -61,6 +61,7 @@ class PrefillBootstrapQueue:
     def __init__(
         self,
         token_to_kv_pool: KVCache,
+        draft_token_to_kv_pool: Optional[KVCache],
         req_to_metadata_buffer_idx_allocator: ReqToMetadataIdxAllocator,
         metadata_buffers: List[torch.Tensor],
         aux_dtype: torch.dtype,
@@ -72,6 +73,8 @@ class PrefillBootstrapQueue:
         scheduler: Scheduler,
     ):
         self.token_to_kv_pool = token_to_kv_pool
+        self.draft_token_to_kv_pool = draft_token_to_kv_pool
+
         self.is_mla_backend = is_mla_backend(token_to_kv_pool)
         self.aux_dtype = aux_dtype
 
@@ -98,6 +101,16 @@ class PrefillBootstrapQueue:
             self.token_to_kv_pool.get_contiguous_buf_infos()
         )
 
+        if self.draft_token_to_kv_pool is not None:
+            # We should also transfer draft model kv cache. The indices are
+            # always shared with a target model.
+            draft_kv_data_ptrs, draft_kv_data_lens, draft_kv_item_lens = (
+                self.draft_token_to_kv_pool.get_contiguous_buf_infos()
+            )
+            kv_data_ptrs += draft_kv_data_ptrs
+            kv_data_lens += draft_kv_data_lens
+            kv_item_lens += draft_kv_item_lens
+
         kv_args.kv_data_ptrs = kv_data_ptrs
         kv_args.kv_data_lens = kv_data_lens
         kv_args.kv_item_lens = kv_item_lens

python/sglang/srt/managers/scheduler.py

@@ -591,6 +591,11 @@ class Scheduler(
             self.disagg_decode_prealloc_queue = DecodePreallocQueue(
                 req_to_token_pool=self.req_to_token_pool,
                 token_to_kv_pool_allocator=self.token_to_kv_pool_allocator,
+                draft_token_to_kv_pool=(
+                    None
+                    if self.draft_worker is None
+                    else self.draft_worker.model_runner.token_to_kv_pool
+                ),
                 req_to_metadata_buffer_idx_allocator=req_to_metadata_buffer_idx_allocator,
                 metadata_buffers=metadata_buffers,
                 aux_dtype=aux_dtype,
@@ -624,6 +629,11 @@ class Scheduler(
 
             self.disagg_prefill_bootstrap_queue = PrefillBootstrapQueue(
                 token_to_kv_pool=self.token_to_kv_pool_allocator.get_kvcache(),
+                draft_token_to_kv_pool=(
+                    None
+                    if self.draft_worker is None
+                    else self.draft_worker.model_runner.token_to_kv_pool
+                ),
                 req_to_metadata_buffer_idx_allocator=req_to_metadata_buffer_idx_allocator,
                 metadata_buffers=metadata_buffers,
                 aux_dtype=aux_dtype,
@@ -1409,6 +1419,13 @@ class Scheduler(
                 self.running_batch.batch_is_full = True
                 break
 
+            if self.disaggregation_mode == DisaggregationMode.PREFILL:
+                # In prefill mode, prealloc queue and transfer queue can also take memory,
+                # so we need to check if the available size for the actual available size.
+                if len(adder.can_run_list) >= self.req_to_token_pool.available_size():
+                    self.running_batch.batch_is_full = True
+                    break
+
             req.init_next_round_input(
                 None if prefix_computed else self.tree_cache,
                 self.enable_hierarchical_cache,

test/srt/run_suite.py

@@ -115,7 +115,7 @@ suites = {
         # 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), # disabled since we have different_tp test
+        TestFile("test_disaggregation.py", 210),
         TestFile("test_disaggregation_different_tp.py", 210),
         TestFile("test_full_deepseek_v3.py", 250),
     ],

test/srt/test_disaggregation.py

@@ -8,6 +8,8 @@ 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_EAGLE_DRAFT_MODEL_FOR_TEST,
+    DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST,
     DEFAULT_MODEL_NAME_FOR_TEST,
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
     DEFAULT_URL_FOR_TEST,
@@ -17,7 +19,9 @@ from sglang.test.test_utils import (
 )
 
 
-class TestDisaggregationMooncake(CustomTestCase):
+# skip the test because we have different_tp test
+@unittest.skip("skip the test because we have different_tp test")
+class TestDisaggregationAccuracy(CustomTestCase):
     @classmethod
     def setUpClass(cls):
         cls.model = DEFAULT_MODEL_NAME_FOR_TEST
@@ -65,6 +69,8 @@ class TestDisaggregationMooncake(CustomTestCase):
             str(cls.base_port + 100),
             "--tp",
             "4",
+            # "--disaggregation-ib-device",
+            # "mlx5_roce0,mlx5_roce1,mlx5_roce2,mlx5_roce3",
         ]
         cls.process_prefill = popen_launch_pd_server(
             cls.model,
@@ -87,6 +93,8 @@ class TestDisaggregationMooncake(CustomTestCase):
             "4",
             "--base-gpu-id",
             "4",
+            # "--disaggregation-ib-device",
+            # "mlx5_roce4,mlx5_roce5,mlx5_roce6,mlx5_roce7",
         ]
         cls.process_decode = popen_launch_pd_server(
             cls.model,
@@ -136,5 +144,138 @@ class TestDisaggregationMooncake(CustomTestCase):
         self.assertGreater(metrics["accuracy"], 0.62)
 
 
+class TestDisaggregationSpecAccuracy(CustomTestCase):
+
+    @classmethod
+    def setUpClass(cls):
+        super().setUpClass()
+        cls.model = DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST
+        cls.draft_model = DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST
+        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}"
+        cls.spec_args = [
+            "--speculative-algorithm",
+            "EAGLE",
+            "--speculative-draft-model-path",
+            cls.draft_model,
+            "--speculative-num-steps",
+            "3",
+            "--speculative-eagle-topk",
+            "4",
+            "--speculative-num-draft-tokens",
+            "16",
+            "--cuda-graph-max-bs",
+            "8",
+        ]
+
+        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 wait_server_ready(cls, url, timeout=60):
+        start_time = time.perf_counter()
+        while True:
+            try:
+                response = requests.get(url)
+                if response.status_code == 200:
+                    print(f"Server {url} is ready")
+                    return
+            except Exception:
+                pass
+
+            if time.perf_counter() - start_time > timeout:
+                raise RuntimeError(f"Server {url} failed to start in {timeout}s")
+            time.sleep(1)
+
+    @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",
+            # "--disaggregation-ib-device",
+            # "mlx5_roce0,mlx5_roce1,mlx5_roce2,mlx5_roce3",
+        ] + cls.spec_args
+
+        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",
+            "4",
+            "--base-gpu-id",
+            "4",
+            # "--disaggregation-ib-device",
+            # "mlx5_roce4,mlx5_roce5,mlx5_roce6,mlx5_roce7",
+        ] + cls.spec_args
+        cls.process_decode = popen_launch_pd_server(
+            cls.model,
+            cls.decode_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=decode_args,
+        )
+
+    def test_gsm8k(self):
+        args = SimpleNamespace(
+            num_shots=5,
+            data_path=None,
+            num_questions=200,
+            max_new_tokens=512,
+            parallel=4,  # TODO: 128 crashes the decode
+            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"], 0.20)
+
+
 if __name__ == "__main__":
     unittest.main()