kv_events.py 12.1 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
5
6
7
8

import queue
import threading
import time
from abc import ABC, abstractmethod
from collections import deque
9
from collections.abc import Callable
10
from dataclasses import asdict
11
12
from itertools import count
from queue import Queue
13
from typing import Any
14
15
16
17

import msgspec
import zmq

18
from vllm.config.kv_events import KVEventsConfig
19
from vllm.logger import init_logger
20
from vllm.v1.core.kv_cache_utils import ExternalBlockHash
21
22
23
24
25

logger = init_logger(__name__)


class EventBatch(
26
27
28
29
    msgspec.Struct,
    array_like=True,  # type: ignore[call-arg]
    omit_defaults=True,  # type: ignore[call-arg]
    gc=False,  # type: ignore[call-arg]
30
31
32
):
    ts: float
    events: list[Any]
33
    data_parallel_rank: int | None = None
34
35
36


class KVCacheEvent(
37
38
39
40
41
42
    msgspec.Struct,
    array_like=True,  # type: ignore[call-arg]
    omit_defaults=True,  # type: ignore[call-arg]
    gc=False,  # type: ignore[call-arg]
    tag=True,
):
43
44
45
    """Base class for all KV cache-related events"""


46
47
48
MEDIUM_GPU = "GPU"


49
class BlockStored(KVCacheEvent):
50
    block_hashes: list[ExternalBlockHash]
51
    parent_block_hash: ExternalBlockHash | None
52
53
    token_ids: list[int]
    block_size: int
54
55
    lora_id: int | None
    medium: str | None
56
57
58


class BlockRemoved(KVCacheEvent):
59
    block_hashes: list[ExternalBlockHash]
60
    medium: str | None
61
62
63
64
65
66
67


class AllBlocksCleared(KVCacheEvent):
    pass


class KVEventBatch(EventBatch):
68
    events: list[BlockStored | BlockRemoved | AllBlocksCleared]
69
70
71


class EventPublisher(ABC):
72
73
    """Lightweight publisher for EventBatch batches with data parallelism
    support.
74

75
76
    In data parallel setups, each DP rank runs its own EventPublisher instance
    to avoid duplicate events and ensure proper event attribution:
77

78
79
80
    - Each DP rank creates a separate publisher
    - Publishers automatically annotate events with their data_parallel_rank
    - This allows consumers to distinguish events from different DP ranks
81

82
83
84
85
86
87
    The publisher is responsible for adding DP metadata since the scheduler
    operates independently of DP topology and shouldn't need DP awareness.
    """

    def __init__(self, data_parallel_rank: int = 0) -> None:
        self._data_parallel_rank = data_parallel_rank
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119

    @abstractmethod
    def publish(self, events: EventBatch) -> None:
        """Emit events in order.

        Implementations should guarantee at-least-once delivery and
        monotonic ordering (e.g., via sequence numbers).
        """

    @abstractmethod
    def shutdown(self) -> None:
        """Shutdown the publisher."""


class NullEventPublisher(EventPublisher):
    """No-op implementation (default when disabled)."""

    def publish(self, events) -> None:
        return

    def shutdown(self) -> None:
        return


class ZmqEventPublisher(EventPublisher):
    """Reliable PUB/ROUTER publisher with an in-memory replay buffer.

    Spawns a separate thread to handle publishing from a queue.

    Parameters
    ----------
    endpoint:
120
        PUB address. Use `tcp://*:5557` to bind or `tcp://host:5557` to
121
122
123
124
125
126
127
128
129
130
131
132
133
134
        connect.
    replay_endpoint:
        Optional ROUTER address for replay requests. When given, subscribers can
        request missed batches by sending the starting sequence number as an
        8-byte big-endian integer.
    buffer_steps:
        Number of past batches to keep for replay.
    hwm:
        ZeroMQ high-water-mark for PUB socket.
    max_queue_size:
        Maximum number of events to buffer in memory.
    topic:
        Topic to publish events to.
    """
135

136
137
138
139
140
    SHUTDOWN_TIMEOUT: float = 1.0
    END_SEQ = (-1).to_bytes(8, "big", signed=True)

    def __init__(
        self,
141
        data_parallel_rank: int,
142
        endpoint: str = "tcp://*:5557",
143
        replay_endpoint: str | None = None,
144
145
146
147
148
149
        buffer_steps: int = 10_000,
        hwm: int = 100_000,
        max_queue_size: int = 100_000,
        topic: str = "",
    ) -> None:
        # Storage
150
        super().__init__(data_parallel_rank)
151
        self._event_queue = Queue[EventBatch | None](maxsize=max_queue_size)
152
153
154
155
        self._buffer = deque[tuple[int, bytes]](maxlen=buffer_steps)

        # ZMQ sockets
        self._ctx = zmq.Context.instance()
156
157
        self._pub: zmq.Socket | None = None
        self._replay: zmq.Socket | None = None
158
159
160
161
        self._dp_rank = data_parallel_rank

        self._endpoint = self.offset_endpoint_port(endpoint, self._dp_rank)
        self._replay_endpoint = self.offset_endpoint_port(
162
163
            replay_endpoint, self._dp_rank
        )
164
        self._hwm = hwm
165
        self._socket_setup()
166
167
168

        # Payload
        self._seq_gen = count()
169
        self._topic_bytes = topic.encode("utf-8")
170
171
172
173
174

        # Thread
        self._running = True
        logger.info("Starting ZMQ publisher thread")

175
176
177
        self._thread = threading.Thread(
            target=self._publisher_thread, daemon=True, name="zmq-publisher"
        )
178
179
180
181
182
        self._thread.start()

    def publish(self, events: EventBatch) -> None:
        if not self._running:
            raise RuntimeError("Publisher is closed")
183
184
        if events.data_parallel_rank is None:
            events.data_parallel_rank = self._data_parallel_rank
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
        self._event_queue.put(events)

    def shutdown(self) -> None:
        """Stop the publisher thread and clean up resources."""
        self._running = False
        self._event_queue.put_nowait(None)

        start = time.time()
        pending_items = True
        while pending_items and (time.time() - start < self.SHUTDOWN_TIMEOUT):
            pending_items = not self._event_queue.empty()
            if pending_items:
                time.sleep(0.1)

        if pending_items:
            logger.warning(
                "Warning: Queue still has %s items after %s seconds timeout",
                self._event_queue.qsize(),
                self.SHUTDOWN_TIMEOUT,
            )

        if self._thread.is_alive():
            self._thread.join(timeout=self.SHUTDOWN_TIMEOUT)

        # Clean up ZMQ resources
        try:
            if self._pub is not None:
                self._pub.close(linger=0)
            if self._replay is not None:
                self._replay.close(linger=0)
        finally:
            pass  # Do not terminate context; other sockets may use it

    def _socket_setup(self) -> None:
        """Initialize sockets
        https://pyzmq.readthedocs.io/en/v19.0.0/morethanbindings.html#thread-safety
        """
        if self._pub is None:
            self._pub = self._ctx.socket(zmq.PUB)
            self._pub.set_hwm(self._hwm)
            # Heuristic: bind if wildcard / * present, else connect.
            # bind stable, connect volatile convention
227
228
229
230
231
232
            if self._endpoint is not None and (
                "*" in self._endpoint
                or "::" in self._endpoint
                or self._endpoint.startswith("ipc://")
                or self._endpoint.startswith("inproc://")
            ):
233
                self._pub.bind(self._endpoint)
234
            elif self._endpoint is not None:
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
                self._pub.connect(self._endpoint)

        # Set up replay socket: use ROUTER
        # 1) handles multiple REQ clients (identities)
        # 2) lets us send back one request → many replies (streamed events)
        # 3) works in our non‑blocking poll loop alongside PUB
        if self._replay_endpoint is not None:
            self._replay = self._ctx.socket(zmq.ROUTER)
            self._replay.bind(self._replay_endpoint)

    def _publisher_thread(self) -> None:
        """Background thread that processes the event queue."""
        self._pack = msgspec.msgpack.Encoder()

        assert self._pub is not None  # narrows type for mypy

        while self._running or self._event_queue.qsize() > 0:
            # --- replay (non-critical) ---------------------------------
            if self._replay is not None and self._replay.poll(0):
                try:
                    self._service_replay()
                except Exception as e:
                    logger.exception("Error in replay: %s", e)

            # --- main queue (critical) ---------------------------------
            try:
                event = self._event_queue.get(timeout=0.1)
                if event is None:
                    break  # Sentinel received, exit thread
            except queue.Empty:
                continue

            try:
                seq = next(self._seq_gen)

                payload = self._pack.encode(event)
                seq_bytes = seq.to_bytes(8, "big")
272
                self._pub.send_multipart((self._topic_bytes, seq_bytes, payload))
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298

                self._buffer.append((seq, payload))
                self._event_queue.task_done()

            except Exception as e:
                # Publishing failed;  back-off a bit to avoid a tight error loop
                logger.exception("Error in publisher thread: %s", e)
                time.sleep(0.1)

    def _service_replay(self) -> None:
        """If a replay request is waiting, send buffered batches."""
        assert self._replay is not None  # narrows type for mypy

        frame = self._replay.recv_multipart()
        if len(frame) != 3:
            logger.warning("Invalid replay request: %s", frame)
            return
        client_id, _, start_seq_bytes = frame
        start_seq = int.from_bytes(start_seq_bytes, "big")

        for seq, buf in self._buffer:
            if seq >= start_seq:
                # [identity, empty_delim, seq_bytes, payload]
                # (identity, empty_delim) are stripped off by the router
                # receiving payload is (seq_bytes, payload)
                self._replay.send_multipart(
299
300
                    (client_id, b"", seq.to_bytes(8, "big"), buf)
                )
301
302
303
304
        # Send end of sequence marker
        # receiving payload is (-1, b""")
        self._replay.send_multipart((client_id, b"", self.END_SEQ, b""))

305
    @staticmethod
306
    def offset_endpoint_port(
307
308
        endpoint: str | None, data_parallel_rank: int
    ) -> str | None:
309
        """Helper function to offset the port in an endpoint by
310
311
312
            the data parallel rank.

        Args:
313
            endpoint: The endpoint string
314
315
316
317
                (e.g., "tcp://*:5557" or "inproc://cache")
            data_parallel_rank: The data parallel rank to offset by

        Returns:
318
            The endpoint with the port offset by data_parallel_rank
319
320
321
322
323
324
325
326
327
328
329
330
331
                or suffix appended
        """
        # Do nothing if input is None or data_parallel_rank is 0
        if not endpoint or data_parallel_rank == 0:
            return endpoint

        if "inproc" in endpoint:
            return f"{endpoint}_dp{data_parallel_rank}"
        if "tcp" in endpoint:
            if endpoint and ":" in endpoint:
                # Get everything after the last colon (the port)
                last_colon_idx = endpoint.rfind(":")
                base_addr = endpoint[:last_colon_idx]
332
                base_port = int(endpoint[last_colon_idx + 1 :])
333
334
335
336
337
                new_port = base_port + data_parallel_rank
                return f"{base_addr}:{new_port}"
            return endpoint
        raise ValueError("Invalid endpoint: must contain 'inproc' or 'tcp'")

338
339
340
341
342
343
344
345

class EventPublisherFactory:
    _registry: dict[str, Callable[..., EventPublisher]] = {
        "null": NullEventPublisher,
        "zmq": ZmqEventPublisher,
    }

    @classmethod
346
    def register_publisher(cls, name: str, ctor: Callable[..., EventPublisher]) -> None:
347
348
349
350
351
        if name in cls._registry:
            raise KeyError(f"publisher '{name}' already registered")
        cls._registry[name] = ctor

    @classmethod
352
    def create(
353
        cls, config: KVEventsConfig | None, data_parallel_rank: int = 0
354
    ) -> EventPublisher:
355
        """Create publisher from a config mapping."""
356
        if config is None or config.publisher == "null":
357
358
            return NullEventPublisher()

359
        config_dict = asdict(config)
360

361
        kind = config_dict.pop("publisher")
362
363
364
365
366
        config_dict.pop("enable_kv_cache_events")
        try:
            constructor = cls._registry[kind]
        except KeyError as exc:
            raise ValueError(f"Unknown event publisher '{kind}'") from exc
367
        return constructor(data_parallel_rank=data_parallel_rank, **config_dict)