scheduler.py 118 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
# Copyright 2023-2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
14
15
"""A scheduler that manages a tensor parallel GPU worker."""

16
import datetime
17
import faulthandler
18
import logging
19
import os
20
import signal
21
import sys
Lianmin Zheng's avatar
Lianmin Zheng committed
22
import threading
23
import time
24
from collections import defaultdict, deque
Lianmin Zheng's avatar
Lianmin Zheng committed
25
from concurrent import futures
26
from dataclasses import dataclass
27
from pathlib import Path
28
from types import SimpleNamespace
29
from typing import Dict, List, Optional, Tuple, Union
30

31
import psutil
32
import setproctitle
33
import torch
34
import zmq
35
from torch.distributed import barrier
36

37
from sglang.global_config import global_config
Lianmin Zheng's avatar
Lianmin Zheng committed
38
from sglang.srt.configs.model_config import ModelConfig
39
from sglang.srt.constants import GPU_MEMORY_TYPE_KV_CACHE, GPU_MEMORY_TYPE_WEIGHTS
40
41
42
43
from sglang.srt.constrained.base_grammar_backend import (
    INVALID_GRAMMAR_OBJ,
    create_grammar_backend,
)
Byron Hsu's avatar
Byron Hsu committed
44
45
46
47
48
from sglang.srt.disaggregation.decode import (
    DecodePreallocQueue,
    DecodeTransferQueue,
    SchedulerDisaggregationDecodeMixin,
)
49
from sglang.srt.disaggregation.kv_events import EventPublisherFactory, KVEventBatch
Byron Hsu's avatar
Byron Hsu committed
50
51
52
53
54
55
from sglang.srt.disaggregation.prefill import (
    PrefillBootstrapQueue,
    SchedulerDisaggregationPrefillMixin,
)
from sglang.srt.disaggregation.utils import (
    DisaggregationMode,
56
    MetadataBuffers,
Byron Hsu's avatar
Byron Hsu committed
57
    ReqToMetadataIdxAllocator,
58
    TransferBackend,
59
    prepare_abort,
Byron Hsu's avatar
Byron Hsu committed
60
)
61
from sglang.srt.distributed import get_pp_group, get_world_group
fzyzcjy's avatar
fzyzcjy committed
62
from sglang.srt.eplb.expert_distribution import get_global_expert_distribution_recorder
xm:D's avatar
xm:D committed
63
64
65
66
67
from sglang.srt.hf_transformers_utils import (
    get_processor,
    get_tokenizer,
    get_tokenizer_from_processor,
)
68
from sglang.srt.layers.dp_attention import compute_dp_attention_world_info
69
70
71
from sglang.srt.layers.logits_processor import LogitsProcessorOutput
from sglang.srt.managers.io_struct import (
    AbortReq,
72
    CloseSessionReqInput,
73
    ExpertDistributionReq,
74
    ExpertDistributionReqOutput,
75
76
    FlushCacheReqInput,
    FlushCacheReqOutput,
77
78
    GetInternalStateReq,
    GetInternalStateReqOutput,
79
80
    GetWeightsByNameReqInput,
    GetWeightsByNameReqOutput,
81
    HealthCheckOutput,
82
83
    InitWeightsUpdateGroupReqInput,
    InitWeightsUpdateGroupReqOutput,
84
85
    LoadLoRAAdapterReqInput,
    LoadLoRAAdapterReqOutput,
86
87
    OpenSessionReqInput,
    OpenSessionReqOutput,
88
    ProfileReq,
89
90
    ProfileReqOutput,
    ProfileReqType,
91
92
93
94
    ReleaseMemoryOccupationReqInput,
    ReleaseMemoryOccupationReqOutput,
    ResumeMemoryOccupationReqInput,
    ResumeMemoryOccupationReqOutput,
95
96
    RpcReqInput,
    RpcReqOutput,
97
98
    SetInternalStateReq,
    SetInternalStateReqOutput,
99
100
    SlowDownReqInput,
    SlowDownReqOutput,
101
102
    TokenizedEmbeddingReqInput,
    TokenizedGenerateReqInput,
103
104
    UnloadLoRAAdapterReqInput,
    UnloadLoRAAdapterReqOutput,
Chayenne's avatar
Chayenne committed
105
106
    UpdateWeightFromDiskReqInput,
    UpdateWeightFromDiskReqOutput,
107
108
    UpdateWeightsFromDistributedReqInput,
    UpdateWeightsFromDistributedReqOutput,
109
110
    UpdateWeightsFromTensorReqInput,
    UpdateWeightsFromTensorReqOutput,
111
)
112
from sglang.srt.managers.mm_utils import init_embedding_cache
113
114
from sglang.srt.managers.schedule_batch import (
    FINISH_ABORT,
Mick's avatar
Mick committed
115
    MultimodalInputs,
116
117
    Req,
    ScheduleBatch,
118
    global_server_args_dict,
119
)
120
121
122
123
124
from sglang.srt.managers.schedule_policy import (
    AddReqResult,
    PrefillAdder,
    SchedulePolicy,
)
125
126
127
from sglang.srt.managers.scheduler_output_processor_mixin import (
    SchedulerOutputProcessorMixin,
)
128
from sglang.srt.managers.session_controller import Session
129
from sglang.srt.managers.tp_worker import TpModelWorker
130
from sglang.srt.managers.tp_worker_overlap_thread import TpModelWorkerClient
131
from sglang.srt.managers.utils import validate_input_length
tarinkk's avatar
tarinkk committed
132
from sglang.srt.mem_cache.chunk_cache import ChunkCache, SWAChunkCache
133
from sglang.srt.mem_cache.hiradix_cache import HiRadixCache
134
from sglang.srt.mem_cache.radix_cache import RadixCache
Hanming Lu's avatar
Hanming Lu committed
135
from sglang.srt.mem_cache.swa_radix_cache import SWARadixCache
136
from sglang.srt.metrics.collector import SchedulerMetricsCollector, SchedulerStats
Lianmin Zheng's avatar
Lianmin Zheng committed
137
from sglang.srt.model_executor.forward_batch_info import ForwardMode, PPProxyTensors
138
from sglang.srt.reasoning_parser import ReasoningParser
139
from sglang.srt.server_args import PortArgs, ServerArgs
140
from sglang.srt.speculative.spec_info import SpeculativeAlgorithm
141
from sglang.srt.torch_memory_saver_adapter import TorchMemorySaverAdapter
142
from sglang.srt.two_batch_overlap import TboDPAttentionPreparer
143
from sglang.srt.utils import (
144
    DeepEPMode,
145
    DynamicGradMode,
146
    broadcast_pyobj,
fzyzcjy's avatar
fzyzcjy committed
147
    configure_gc_logger,
148
    configure_logger,
Lianmin Zheng's avatar
Lianmin Zheng committed
149
    disable_request_logging,
150
    get_available_gpu_memory,
151
    get_bool_env_var,
152
    get_zmq_socket,
153
    is_cpu,
Lianmin Zheng's avatar
Lianmin Zheng committed
154
    kill_itself_when_parent_died,
155
    point_to_point_pyobj,
156
    pyspy_dump_schedulers,
157
158
    require_mlp_sync,
    require_mlp_tp_gather,
159
    set_gpu_proc_affinity,
160
161
162
    set_random_seed,
    suppress_other_loggers,
)
163
from sglang.utils import TypeBasedDispatcher, get_exception_traceback
164
165
166

logger = logging.getLogger(__name__)

167
# Test retract decode for debugging purposes
168
169
TEST_RETRACT = get_bool_env_var("SGLANG_TEST_RETRACT")
RECORD_STEP_TIME = get_bool_env_var("SGLANG_RECORD_STEP_TIME")
170
GRAMMAR_TIMEOUT = float(os.environ.get("SGLANG_GRAMMAR_TIMEOUT", 300))
171

172
173
_is_cpu = is_cpu()

174

175
176
@dataclass
class GenerationBatchResult:
177
178
179
    logits_output: Optional[LogitsProcessorOutput]
    pp_hidden_states_proxy_tensors: Optional[torch.Tensor]
    next_token_ids: Optional[List[int]]
180
181
    extend_input_len_per_req: List[int]
    extend_logprob_start_len_per_req: List[int]
182
    bid: int
183
    can_run_cuda_graph: bool
184
185
186
187
188
189
190
191


@dataclass
class EmbeddingBatchResult:
    embeddings: torch.Tensor
    bid: int


192
193
194
195
196
197
198
199
200
201
202
203
class KvMetrics:
    def __init__(self):
        self.request_active_slots = None
        self.request_total_slots = None
        self.kv_active_blocks = None
        self.kv_total_blocks = None
        self.num_requests_waiting = None
        self.gpu_cache_usage_perc = None
        self.gpu_prefix_cache_hit_rate = None
        self.data_parallel_rank = None


204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
class IdleSleeper:
    """
    In setups which have long inactivity periods it is desirable to reduce
    system power consumption when sglang does nothing. This would lead not only
    to power savings, but also to more CPU thermal headroom when a request
    eventually comes. This is important in cases when multiple GPUs are connected
    as each GPU would otherwise pin one thread at 100% CPU usage.

    The simplest solution is to use zmq.Poller on all sockets that may receive
    data that needs handling immediately.
    """

    def __init__(self, sockets):
        self.poller = zmq.Poller()
        for s in sockets:
            self.poller.register(s, zmq.POLLIN)

    def maybe_sleep(self):
        self.poller.poll(1000)


Byron Hsu's avatar
Byron Hsu committed
225
226
227
228
229
class Scheduler(
    SchedulerOutputProcessorMixin,
    SchedulerDisaggregationDecodeMixin,
    SchedulerDisaggregationPrefillMixin,
):
230
231
232
233
234
235
236
237
    """A scheduler that manages a tensor parallel GPU worker."""

    def __init__(
        self,
        server_args: ServerArgs,
        port_args: PortArgs,
        gpu_id: int,
        tp_rank: int,
238
        pp_rank: int,
239
        dp_rank: Optional[int],
240
241
    ):
        # Parse args
242
        self.server_args = server_args
243
        self.tp_rank = tp_rank
244
        self.pp_rank = pp_rank
245
        self.dp_rank = dp_rank
246
        self.tp_size = server_args.tp_size
247
248
        self.pp_size = server_args.pp_size
        self.dp_size = server_args.dp_size
249
        self.schedule_policy = server_args.schedule_policy
250
        self.enable_lora = server_args.enable_lora
251
        self.max_loras_per_batch = server_args.max_loras_per_batch
252
        self.enable_overlap = not server_args.disable_overlap_schedule
253
        self.skip_tokenizer_init = server_args.skip_tokenizer_init
254
        self.enable_metrics = server_args.enable_metrics
255
256
257
        self.enable_metrics_for_all_schedulers = (
            server_args.enable_metrics_for_all_schedulers
        )
258
        self.enable_kv_cache_events = server_args.kv_events_config is not None
259
        self.stream_interval = server_args.stream_interval
260
261
262
        self.spec_algorithm = SpeculativeAlgorithm.from_string(
            server_args.speculative_algorithm
        )
263
264
        self.gpu_id = gpu_id
        self.enable_hierarchical_cache = server_args.enable_hierarchical_cache
265
        self.enable_hicache_storage = server_args.hicache_storage_backend is not None
Lianmin Zheng's avatar
Lianmin Zheng committed
266
        self.page_size = server_args.page_size
267
268
        self.dp_size = server_args.dp_size
        self.attn_tp_rank, self.attn_tp_size, self.attn_dp_rank = (
269
270
271
272
273
274
275
276
            compute_dp_attention_world_info(
                server_args.enable_dp_attention,
                self.tp_rank,
                self.tp_size,
                self.dp_size,
            )
        )

277
278
        # Init inter-process communication
        context = zmq.Context(2)
279
280
        self.idle_sleeper = None

281
        if self.pp_rank == 0 and self.attn_tp_rank == 0:
282
            self.recv_from_tokenizer = get_zmq_socket(
Lianmin Zheng's avatar
Lianmin Zheng committed
283
                context, zmq.PULL, port_args.scheduler_input_ipc_name, False
284
            )
285
            self.send_to_tokenizer = get_zmq_socket(
Lianmin Zheng's avatar
Lianmin Zheng committed
286
                context, zmq.PUSH, port_args.tokenizer_ipc_name, False
287
            )
288

289
            if server_args.skip_tokenizer_init:
290
                # Directly send to the TokenizerManager
291
                self.send_to_detokenizer = get_zmq_socket(
Lianmin Zheng's avatar
Lianmin Zheng committed
292
                    context, zmq.PUSH, port_args.tokenizer_ipc_name, False
293
294
                )
            else:
295
                # Send to the DetokenizerManager
296
                self.send_to_detokenizer = get_zmq_socket(
Lianmin Zheng's avatar
Lianmin Zheng committed
297
                    context, zmq.PUSH, port_args.detokenizer_ipc_name, False
298
                )
299
300
301
302

            self.recv_from_rpc = get_zmq_socket(
                context, zmq.DEALER, port_args.rpc_ipc_name, False
            )
303
304
305
306
307
308
309
            if self.server_args.sleep_on_idle:
                self.idle_sleeper = IdleSleeper(
                    [
                        self.recv_from_tokenizer,
                        self.recv_from_rpc,
                    ]
                )
310
        else:
311
            self.recv_from_tokenizer = None
312
            self.recv_from_rpc = None
313
314
            self.send_to_tokenizer = SimpleNamespace(send_pyobj=lambda x: None)
            self.send_to_detokenizer = SimpleNamespace(send_pyobj=lambda x: None)
315

316
317
318
319
320
        if self.current_scheduler_metrics_enabled():
            self.send_metrics_from_scheduler = get_zmq_socket(
                context, zmq.PUSH, port_args.metrics_ipc_name, False
            )

321
        # Init tokenizer
322
        self.init_tokenizer()
323

324
325
326
327
328
329
330
331
332
        # Set reasoning_parser and think_end_id if --reasoning_parser is enabled
        if self.server_args.reasoning_parser and self.tokenizer:
            reasoning_parser = ReasoningParser(
                model_type=self.server_args.reasoning_parser, stream_reasoning=False
            )
            self.tokenizer.think_end_id = self.tokenizer.encode(
                reasoning_parser.detector.think_end_token, add_special_tokens=False
            )[0]

333
334
335
336
        # Check whether overlap can be enabled
        if not self.is_generation:
            self.enable_overlap = False
            logger.info("Overlap scheduler is disabled for embedding models.")
337

338
        # Launch a tensor parallel worker
339
        if self.enable_overlap:
340
            TpWorkerClass = TpModelWorkerClient
341
342
        else:
            TpWorkerClass = TpModelWorker
343

344
        self.tp_worker = TpWorkerClass(
345
            server_args=server_args,
346
347
            gpu_id=gpu_id,
            tp_rank=tp_rank,
348
            pp_rank=pp_rank,
349
            dp_rank=dp_rank,
350
            nccl_port=port_args.nccl_port,
351
        )
352

353
        # Launch a draft worker for speculative decoding
354
355
356
357
358
359
360
361
362
363
364
365
366
367
        if self.spec_algorithm.is_eagle():
            from sglang.srt.speculative.eagle_worker import EAGLEWorker

            self.draft_worker = EAGLEWorker(
                gpu_id=gpu_id,
                tp_rank=tp_rank,
                server_args=server_args,
                nccl_port=port_args.nccl_port,
                target_worker=self.tp_worker,
                dp_rank=dp_rank,
            )
        else:
            self.draft_worker = None

368
        # Get token and memory info from the model worker
369
370
371
372
        (
            self.max_total_num_tokens,
            self.max_prefill_tokens,
            self.max_running_requests,
373
            self.max_req_len,
374
375
            self.max_req_input_len,
            self.random_seed,
376
            self.device,
377
378
379
380
381
            worker_global_server_args_dict,
            _,
            _,
            _,
        ) = self.tp_worker.get_worker_info()
382
383
384
385
386
387
388
389
        if global_server_args_dict["max_micro_batch_size"] is None:
            global_server_args_dict["max_micro_batch_size"] = max(
                self.max_running_requests // server_args.pp_size, 1
            )

        self.tp_group = self.tp_worker.get_tp_group()
        self.tp_cpu_group = self.tp_group.cpu_group
        self.attn_tp_group = self.tp_worker.get_attention_tp_group()
390
        self.attn_tp_cpu_group = self.tp_worker.get_attention_tp_cpu_group()
391
392
393
        self.pp_group = get_pp_group()
        self.world_group = get_world_group()

394
        self.pad_input_ids_func = self.tp_worker.get_pad_input_ids_func()
395
        global_server_args_dict.update(worker_global_server_args_dict)
396
        set_random_seed(self.random_seed)
397

Hanming Lu's avatar
Hanming Lu committed
398
399
400
401
402
403
404
405
        # Hybrid
        self.is_hybrid = self.tp_worker.is_hybrid
        if self.is_hybrid:
            self.sliding_window_size = self.tp_worker.sliding_window_size
            self.full_tokens_per_layer, self.swa_tokens_per_layer = (
                self.tp_worker.get_tokens_per_layer_info()
            )

406
        # Print debug info
407
        if tp_rank == 0:
408
409
410
            avail_mem = get_available_gpu_memory(
                self.device, self.gpu_id, empty_cache=False
            )
411
412
413
414
415
            logger.info(
                f"max_total_num_tokens={self.max_total_num_tokens}, "
                f"chunked_prefill_size={server_args.chunked_prefill_size}, "
                f"max_prefill_tokens={self.max_prefill_tokens}, "
                f"max_running_requests={self.max_running_requests}, "
416
417
                f"context_len={self.model_config.context_len}, "
                f"available_gpu_mem={avail_mem:.2f} GB"
418
            )
419

Lianmin Zheng's avatar
Lianmin Zheng committed
420
        # Init memory pool and cache
421
        self.init_memory_pool_and_cache()
422
423
424

        # Init running status
        self.waiting_queue: List[Req] = []
425
        # The running decoding batch for continuous batching
Lianmin Zheng's avatar
Lianmin Zheng committed
426
        self.running_batch: ScheduleBatch = ScheduleBatch(reqs=[], batch_is_full=False)
427
        # The current forward batch
Lianmin Zheng's avatar
Lianmin Zheng committed
428
        self.cur_batch: Optional[ScheduleBatch] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
429
        # The last forward batch
430
        self.last_batch: Optional[ScheduleBatch] = None
Lianmin Zheng's avatar
Lianmin Zheng committed
431
432
        self.forward_ct = 0
        self.forward_ct_decode = 0
433
        self.num_generated_tokens = 0
Liangsheng Yin's avatar
Liangsheng Yin committed
434
        self.last_prefill_tokens = 0
435
436
        self.last_decode_stats_tic = time.perf_counter()
        self.last_prefill_stats_tic = time.perf_counter()
437
        self.return_health_check_ct = 0
438
439
440
441
442
        self.num_retracted_reqs: int = 0
        self.num_paused_reqs: int = 0
        self.kv_transfer_speed_gb_s: float = 0.0
        self.kv_transfer_latency_ms: float = 0.0
        self.sessions: Dict[str, Session] = {}
443
        self.current_stream = torch.get_device_module(self.device).current_stream()
444
445
        if self.device == "cpu":
            self.current_stream.synchronize = lambda: None  # No-op for CPU
446
        self.forward_sleep_time = None
447

448
449
        # Init chunked prefill
        self.chunked_prefill_size = server_args.chunked_prefill_size
450
451
        if self.chunked_prefill_size <= 0:  # -1 means disable
            self.chunked_prefill_size = None
452
        self.chunked_req = None
453
454
455
456
        self.is_mixed_chunk = (
            self.chunked_prefill_size is not None and server_args.enable_mixed_chunk
        )

Lianmin Zheng's avatar
Lianmin Zheng committed
457
        # Init the grammar backend for constrained generation
458
        self.grammar_queue: List[Req] = []
459
        if not server_args.skip_tokenizer_init:
460
461
462
            self.grammar_backend = create_grammar_backend(
                server_args, self.tokenizer, self.model_config.vocab_size
            )
Lianmin Zheng's avatar
Lianmin Zheng committed
463
464
        else:
            self.grammar_backend = None
465

466
        # Init schedule policy and new token estimation
467
        self.policy = SchedulePolicy(
Lianmin Zheng's avatar
Lianmin Zheng committed
468
469
470
            self.schedule_policy,
            self.tree_cache,
            self.enable_hierarchical_cache,
471
        )
472
473
474
        assert (
            server_args.schedule_conservativeness >= 0
        ), "Invalid schedule_conservativeness"
475
476
        self.init_new_token_ratio = min(
            global_config.default_init_new_token_ratio
477
478
            * server_args.schedule_conservativeness,
            1.0,
479
        )
480
481
482
483
484
485
486
487
488
489
        self.min_new_token_ratio = min(
            self.init_new_token_ratio
            * global_config.default_min_new_token_ratio_factor,
            1.0,
        )
        self.new_token_ratio_decay = (
            self.init_new_token_ratio - self.min_new_token_ratio
        ) / global_config.default_new_token_ratio_decay_steps
        self.new_token_ratio = self.init_new_token_ratio

Lianmin Zheng's avatar
Lianmin Zheng committed
490
491
492
493
        # Init watchdog thread
        self.watchdog_timeout = server_args.watchdog_timeout
        t = threading.Thread(target=self.watchdog_thread, daemon=True)
        t.start()
494
        self.parent_process = psutil.Process().parent()
495
496

        # Init memory saver, profiler and metric stats
497
498
499
        self.memory_saver_adapter = TorchMemorySaverAdapter.create(
            enable=server_args.enable_memory_saver
        )
500
        self.init_profier()
501
502

        # Init metrics stats
503
        self.init_metrics(tp_rank, pp_rank, dp_rank)
504
        self.init_kv_events(server_args.kv_events_config)
505

506
507
        # Init request dispatcher
        self._request_dispatcher = TypeBasedDispatcher(
508
509
510
            [
                (TokenizedGenerateReqInput, self.handle_generate_request),
                (TokenizedEmbeddingReqInput, self.handle_embedding_request),
511
                (FlushCacheReqInput, self.flush_cache_wrapped),
512
                (AbortReq, self.abort_request),
513
514
                (OpenSessionReqInput, self.open_session),
                (CloseSessionReqInput, self.close_session),
515
516
517
518
519
520
521
522
                (UpdateWeightFromDiskReqInput, self.update_weights_from_disk),
                (InitWeightsUpdateGroupReqInput, self.init_weights_update_group),
                (
                    UpdateWeightsFromDistributedReqInput,
                    self.update_weights_from_distributed,
                ),
                (UpdateWeightsFromTensorReqInput, self.update_weights_from_tensor),
                (GetWeightsByNameReqInput, self.get_weights_by_name),
523
524
                (ReleaseMemoryOccupationReqInput, self.release_memory_occupation),
                (ResumeMemoryOccupationReqInput, self.resume_memory_occupation),
525
                (SlowDownReqInput, self.slow_down),
526
                (ProfileReq, self.profile),
527
                (GetInternalStateReq, self.get_internal_state),
528
                (SetInternalStateReq, self.set_internal_state),
529
                (RpcReqInput, self.handle_rpc_request),
530
                (ExpertDistributionReq, self.expert_distribution_handle),
531
532
                (LoadLoRAAdapterReqInput, self.load_lora_adapter),
                (UnloadLoRAAdapterReqInput, self.unload_lora_adapter),
533
534
535
            ]
        )

536
        # Init disaggregation
Byron Hsu's avatar
Byron Hsu committed
537
538
539
540
541
        self.disaggregation_mode = DisaggregationMode(
            self.server_args.disaggregation_mode
        )
        self.init_disaggregation()

fzyzcjy's avatar
fzyzcjy committed
542
543
544
        if get_bool_env_var("SGLANG_GC_LOG"):
            configure_gc_logger()

545
546
547
    def current_scheduler_metrics_enabled(self):
        return self.attn_tp_rank == 0 or self.enable_metrics_for_all_schedulers

548
549
550
551
    def maybe_sleep_on_idle(self):
        if self.idle_sleeper is not None:
            self.idle_sleeper.maybe_sleep()

552
553
    def init_tokenizer(self):
        server_args = self.server_args
Lianmin Zheng's avatar
Lianmin Zheng committed
554

555
        self.model_config = ModelConfig.from_server_args(server_args)
556
        self.is_generation = self.model_config.is_generation
557

558
559
560
561
562
563
564
565
566
        if server_args.skip_tokenizer_init:
            self.tokenizer = self.processor = None
        else:
            if self.model_config.is_multimodal:
                self.processor = get_processor(
                    server_args.tokenizer_path,
                    tokenizer_mode=server_args.tokenizer_mode,
                    trust_remote_code=server_args.trust_remote_code,
                    revision=server_args.revision,
567
                    use_fast=not server_args.disable_fast_image_processor,
568
                )
xm:D's avatar
xm:D committed
569
                self.tokenizer = get_tokenizer_from_processor(self.processor)
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
            else:
                self.tokenizer = get_tokenizer(
                    server_args.tokenizer_path,
                    tokenizer_mode=server_args.tokenizer_mode,
                    trust_remote_code=server_args.trust_remote_code,
                    revision=server_args.revision,
                )

    def init_memory_pool_and_cache(self):
        server_args = self.server_args

        self.req_to_token_pool, self.token_to_kv_pool_allocator = (
            self.tp_worker.get_memory_pool()
        )

        if (
            server_args.chunked_prefill_size is not None
            and server_args.disable_radix_cache
        ):
Hanming Lu's avatar
Hanming Lu committed
589
            if self.is_hybrid:
tarinkk's avatar
tarinkk committed
590
591
592
593
                ChunkCacheClass = SWAChunkCache
            else:
                ChunkCacheClass = ChunkCache
            self.tree_cache = ChunkCacheClass(
594
595
                req_to_token_pool=self.req_to_token_pool,
                token_to_kv_pool_allocator=self.token_to_kv_pool_allocator,
596
                page_size=self.page_size,
597
598
599
600
601
602
            )
        else:
            if self.enable_hierarchical_cache:
                self.tree_cache = HiRadixCache(
                    req_to_token_pool=self.req_to_token_pool,
                    token_to_kv_pool_allocator=self.token_to_kv_pool_allocator,
603
604
605
606
607
                    tp_cache_group=(
                        self.attn_tp_cpu_group
                        if self.server_args.enable_dp_attention
                        else self.tp_cpu_group
                    ),
608
                    page_size=self.page_size,
609
                    hicache_ratio=server_args.hicache_ratio,
Zhiqiang Xie's avatar
Zhiqiang Xie committed
610
611
                    hicache_size=server_args.hicache_size,
                    hicache_write_policy=server_args.hicache_write_policy,
612
613
614
615
616
617
                    hicache_io_backend=(
                        "direct"
                        if server_args.attention_backend
                        == "fa3"  # hot fix for incompatibility
                        else server_args.hicache_io_backend
                    ),
618
                    hicache_storage_backend=server_args.hicache_storage_backend,
619
                )
620
621
622
                self.tp_worker.register_hicache_layer_transfer_counter(
                    self.tree_cache.cache_controller.layer_done_counter
                )
Hanming Lu's avatar
Hanming Lu committed
623
624
625
626
627
628
629
630
631
632
633
            elif self.is_hybrid:
                assert (
                    self.server_args.disaggregation_mode == "null"
                ), "Hybrid mode does not support disaggregation yet"
                self.tree_cache = SWARadixCache(
                    req_to_token_pool=self.req_to_token_pool,
                    token_to_kv_pool_allocator=self.token_to_kv_pool_allocator,
                    sliding_window_size=self.sliding_window_size,
                    page_size=self.page_size,
                    disable=server_args.disable_radix_cache,
                )
634

635
636
637
638
            else:
                self.tree_cache = RadixCache(
                    req_to_token_pool=self.req_to_token_pool,
                    token_to_kv_pool_allocator=self.token_to_kv_pool_allocator,
Lianmin Zheng's avatar
Lianmin Zheng committed
639
                    page_size=self.page_size,
640
                    disable=server_args.disable_radix_cache,
641
                    enable_kv_cache_events=self.enable_kv_cache_events,
642
643
644
645
646
647
648
649
650
651
652
653
                )

        self.decode_mem_cache_buf_multiplier = (
            1
            if self.spec_algorithm.is_none()
            else (
                server_args.speculative_num_draft_tokens
                + (
                    server_args.speculative_eagle_topk
                    * server_args.speculative_num_steps
                )
            )
654
        )
655

656
657
658
        embedding_cache_size = int(os.environ.get("SGLANG_VLM_CACHE_SIZE_MB", "100"))
        init_embedding_cache(embedding_cache_size * 1024 * 1024)

659
660
661
662
663
    def init_profier(self):
        self.torch_profiler = None
        self.torch_profiler_output_dir: Optional[str] = None
        self.profiler_activities: Optional[List[str]] = None
        self.profile_id: Optional[str] = None
664
        self.profiler_start_forward_ct: Optional[int] = None
665
666
667
668
669
670
671
672
673
674
        self.profiler_target_forward_ct: Optional[int] = None
        self.profiler_target_prefill_ct: Optional[int] = None
        self.profiler_target_decode_ct: Optional[int] = None
        self.profiler_prefill_ct: Optional[int] = None
        self.profiler_decode_ct: Optional[int] = None
        self.profile_by_stage: bool = False
        self.profile_steps: Optional[int] = None
        self.profile_in_progress: bool = False
        self.rpd_profiler = None

675
    def init_metrics(self, tp_rank: int, pp_rank: int, dp_rank: Optional[int]):
676
        self.last_gen_throughput: float = 0.0
Lianmin Zheng's avatar
Lianmin Zheng committed
677
        self.last_input_throughput: float = 0.0
678
679
680
681
682
        self.step_time_dict = defaultdict(list)  # Dict[batch size -> step time]
        self.spec_num_total_accepted_tokens = 0
        self.spec_num_total_forward_ct = 0
        self.cum_spec_accept_length = 0
        self.cum_spec_accept_count = 0
683
        self.total_retracted_reqs = 0
684
685
686
        self.stats = SchedulerStats()
        if self.enable_metrics:
            engine_type = "unified"
687
688
689
690
691
692
693
694
695
            labels = {
                "model_name": self.server_args.served_model_name,
                "engine_type": engine_type,
                "tp_rank": tp_rank,
                "pp_rank": pp_rank,
            }
            if dp_rank is not None:
                labels["dp_rank"] = dp_rank
            self.metrics_collector = SchedulerMetricsCollector(labels=labels)
Lianmin Zheng's avatar
Lianmin Zheng committed
696

697
698
    def init_kv_events(self, kv_events_config: Optional[str]):
        if self.enable_kv_cache_events:
699
700
701
            self.kv_event_publisher = EventPublisherFactory.create(
                kv_events_config, self.attn_dp_rank
            )
702

Byron Hsu's avatar
Byron Hsu committed
703
    def init_disaggregation(self):
704
705
706
707
        self.transfer_backend = TransferBackend(
            self.server_args.disaggregation_transfer_backend
        )

Byron Hsu's avatar
Byron Hsu committed
708
709
710
711
        if (
            self.disaggregation_mode == DisaggregationMode.DECODE
        ):  # *2 for the headroom.
            buffer_size = (self.req_to_token_pool.size) * 2
Byron Hsu's avatar
Byron Hsu committed
712
            self.req_to_metadata_buffer_idx_allocator = ReqToMetadataIdxAllocator(
Byron Hsu's avatar
Byron Hsu committed
713
714
                buffer_size
            )
715
716
            self.disagg_metadata_buffers = MetadataBuffers(
                buffer_size,
717
718
                hidden_size=self.model_config.hf_text_config.hidden_size,
                dtype=self.model_config.dtype,
719
720
                custom_mem_pool=self.token_to_kv_pool_allocator.get_kvcache().maybe_get_custom_mem_pool(),
            )
Byron Hsu's avatar
Byron Hsu committed
721
722
723

            # The decode requests polling kv cache
            self.disagg_decode_transfer_queue = DecodeTransferQueue(
724
                gloo_group=self.attn_tp_cpu_group,
Byron Hsu's avatar
Byron Hsu committed
725
                req_to_metadata_buffer_idx_allocator=self.req_to_metadata_buffer_idx_allocator,
726
                tp_rank=self.tp_rank,
727
                metadata_buffers=self.disagg_metadata_buffers,
Byron Hsu's avatar
Byron Hsu committed
728
729
                scheduler=self,
                tree_cache=self.tree_cache,
Byron Hsu's avatar
Byron Hsu committed
730
731
732
733
734
735
            )

            # The decode requests pending for pre-allocation
            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,
Byron Hsu's avatar
Byron Hsu committed
736
737
738
739
740
                draft_token_to_kv_pool=(
                    None
                    if self.draft_worker is None
                    else self.draft_worker.model_runner.token_to_kv_pool
                ),
Byron Hsu's avatar
Byron Hsu committed
741
                req_to_metadata_buffer_idx_allocator=self.req_to_metadata_buffer_idx_allocator,
742
                metadata_buffers=self.disagg_metadata_buffers,
Byron Hsu's avatar
Byron Hsu committed
743
744
745
                scheduler=self,
                transfer_queue=self.disagg_decode_transfer_queue,
                tree_cache=self.tree_cache,
746
                gloo_group=self.attn_tp_cpu_group,
Byron Hsu's avatar
Byron Hsu committed
747
748
                tp_rank=self.tp_rank,
                tp_size=self.tp_size,
749
750
                dp_size=self.server_args.dp_size,
                gpu_id=self.gpu_id,
Byron Hsu's avatar
Byron Hsu committed
751
                bootstrap_port=self.server_args.disaggregation_bootstrap_port,
752
753
                max_total_num_tokens=self.max_total_num_tokens,
                prefill_pp_size=self.server_args.disaggregation_prefill_pp,
754
                num_reserved_decode_tokens=self.server_args.num_reserved_decode_tokens,
755
                transfer_backend=self.transfer_backend,
Byron Hsu's avatar
Byron Hsu committed
756
            )
Liangsheng Yin's avatar
Liangsheng Yin committed
757

Byron Hsu's avatar
Byron Hsu committed
758
759
760
        elif self.disaggregation_mode == DisaggregationMode.PREFILL:
            # *2 for the headroom.
            buffer_size = self.max_running_requests * 2
Byron Hsu's avatar
Byron Hsu committed
761
            self.req_to_metadata_buffer_idx_allocator = ReqToMetadataIdxAllocator(
Byron Hsu's avatar
Byron Hsu committed
762
763
                buffer_size
            )
764
765
            self.disagg_metadata_buffers = MetadataBuffers(
                buffer_size,
766
767
                hidden_size=self.model_config.hf_text_config.hidden_size,
                dtype=self.model_config.dtype,
768
769
                custom_mem_pool=self.token_to_kv_pool_allocator.get_kvcache().maybe_get_custom_mem_pool(),
            )
Byron Hsu's avatar
Byron Hsu committed
770

Liangsheng Yin's avatar
Liangsheng Yin committed
771
            self.disagg_prefill_bootstrap_queue = PrefillBootstrapQueue(
Byron Hsu's avatar
Byron Hsu committed
772
                token_to_kv_pool=self.token_to_kv_pool_allocator.get_kvcache(),
Byron Hsu's avatar
Byron Hsu committed
773
774
775
776
777
                draft_token_to_kv_pool=(
                    None
                    if self.draft_worker is None
                    else self.draft_worker.model_runner.token_to_kv_pool
                ),
Byron Hsu's avatar
Byron Hsu committed
778
                req_to_metadata_buffer_idx_allocator=self.req_to_metadata_buffer_idx_allocator,
779
                metadata_buffers=self.disagg_metadata_buffers,
Byron Hsu's avatar
Byron Hsu committed
780
781
                tp_rank=self.tp_rank,
                tp_size=self.tp_size,
Byron Hsu's avatar
Byron Hsu committed
782
                gpu_id=self.gpu_id,
Byron Hsu's avatar
Byron Hsu committed
783
                bootstrap_port=self.server_args.disaggregation_bootstrap_port,
784
                gloo_group=self.attn_tp_cpu_group,
Byron Hsu's avatar
Byron Hsu committed
785
786
787
                max_total_num_tokens=self.max_total_num_tokens,
                decode_tp_size=self.server_args.disaggregation_decode_tp,
                decode_dp_size=self.server_args.disaggregation_decode_dp,
788
                scheduler=self,
Byron Hsu's avatar
Byron Hsu committed
789
790
791
                pp_rank=self.pp_rank,
                pp_size=self.pp_size,
                transfer_backend=self.transfer_backend,
Byron Hsu's avatar
Byron Hsu committed
792
793
            )
            # The prefill requests that are in the middle of kv sending
794
            self.disagg_prefill_inflight_queue: List[Req] = []
Byron Hsu's avatar
Byron Hsu committed
795

796
    @DynamicGradMode()
797
    def event_loop_normal(self):
798
        """A normal scheduler loop."""
799
        while True:
Lianmin Zheng's avatar
Lianmin Zheng committed
800
801
            recv_reqs = self.recv_requests()
            self.process_input_requests(recv_reqs)
802

803
            batch = self.get_next_batch_to_run()
Lianmin Zheng's avatar
Lianmin Zheng committed
804
            self.cur_batch = batch
805
806
807
808

            if batch:
                result = self.run_batch(batch)
                self.process_batch_result(batch, result)
Lianmin Zheng's avatar
Lianmin Zheng committed
809
            else:
Lianmin Zheng's avatar
Lianmin Zheng committed
810
                # When the server is idle, do self-check and re-init some states
Lianmin Zheng's avatar
Lianmin Zheng committed
811
                self.check_memory()
Hanming Lu's avatar
Hanming Lu committed
812
                self.check_tree_cache()
813
                self.new_token_ratio = self.init_new_token_ratio
814
                self.maybe_sleep_on_idle()
815
816

            self.last_batch = batch
817

818
    @DynamicGradMode()
Lianmin Zheng's avatar
Lianmin Zheng committed
819
    def event_loop_overlap(self):
820
        """A scheduler loop that overlaps the CPU processing and GPU computation."""
821
        self.result_queue = deque()
Lianmin Zheng's avatar
Lianmin Zheng committed
822
823
824
825
826
827
828

        while True:
            recv_reqs = self.recv_requests()
            self.process_input_requests(recv_reqs)

            batch = self.get_next_batch_to_run()
            self.cur_batch = batch
829

Lianmin Zheng's avatar
Lianmin Zheng committed
830
            if batch:
831
                batch.launch_done = threading.Event()
Lianmin Zheng's avatar
Lianmin Zheng committed
832
                result = self.run_batch(batch)
833
                self.result_queue.append((batch.copy(), result))
Lianmin Zheng's avatar
Lianmin Zheng committed
834

835
                if self.last_batch is None:
836
                    # Create a dummy first batch to start the pipeline for overlap schedule.
837
838
839
840
841
842
                    # It is now used for triggering the sampling_info_done event.
                    tmp_batch = ScheduleBatch(
                        reqs=None,
                        forward_mode=ForwardMode.DUMMY_FIRST,
                        next_batch_sampling_info=self.tp_worker.cur_sampling_info,
                    )
843
                    self.process_batch_result(tmp_batch, None, batch.launch_done)
844

Lianmin Zheng's avatar
Lianmin Zheng committed
845
            if self.last_batch:
846
                # Process the results of the last batch
847
                tmp_batch, tmp_result = self.result_queue.popleft()
848
849
850
                tmp_batch.next_batch_sampling_info = (
                    self.tp_worker.cur_sampling_info if batch else None
                )
851
852
853
854
                # NOTE: we should use current launched batch's launch_done event Instead of the last batch's
                self.process_batch_result(
                    tmp_batch, tmp_result, batch.launch_done if batch else None
                )
Lianmin Zheng's avatar
Lianmin Zheng committed
855
            elif batch is None:
Lianmin Zheng's avatar
Lianmin Zheng committed
856
                # When the server is idle, do self-check and re-init some states
Lianmin Zheng's avatar
Lianmin Zheng committed
857
                self.check_memory()
Hanming Lu's avatar
Hanming Lu committed
858
                self.check_tree_cache()
859
                self.new_token_ratio = self.init_new_token_ratio
860
                self.maybe_sleep_on_idle()
Lianmin Zheng's avatar
Lianmin Zheng committed
861
862
863

            self.last_batch = batch

864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
    @DynamicGradMode()
    def event_loop_pp(self):
        """A non-overlap scheduler loop for pipeline parallelism."""
        mbs = [None] * self.pp_size
        last_mbs = [None] * self.pp_size
        self.running_mbs = [
            ScheduleBatch(reqs=[], batch_is_full=False) for _ in range(self.pp_size)
        ]
        bids = [None] * self.pp_size
        pp_outputs: Optional[PPProxyTensors] = None
        while True:
            server_is_idle = True
            for mb_id in range(self.pp_size):
                self.running_batch = self.running_mbs[mb_id]
                self.last_batch = last_mbs[mb_id]

                recv_reqs = self.recv_requests()
                self.process_input_requests(recv_reqs)
                mbs[mb_id] = self.get_next_batch_to_run()
                self.running_mbs[mb_id] = self.running_batch

                self.cur_batch = mbs[mb_id]
                if self.cur_batch:
                    server_is_idle = False
                    result = self.run_batch(self.cur_batch)

890
                # (last rank) send the outputs to the next step
891
892
893
894
895
896
                if self.pp_group.is_last_rank:
                    if self.cur_batch:
                        next_token_ids, bids[mb_id] = (
                            result.next_token_ids,
                            result.bid,
                        )
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
                        if self.cur_batch.return_logprob:
                            pp_outputs = PPProxyTensors(
                                {
                                    "next_token_ids": next_token_ids,
                                    "extend_input_len_per_req": result.extend_input_len_per_req,
                                    "extend_logprob_start_len_per_req": result.extend_logprob_start_len_per_req,
                                }
                                | (
                                    {
                                        f"logits_output.{k}": v
                                        for k, v in result.logits_output.__dict__.items()
                                    }
                                    if result.logits_output is not None
                                    else {}
                                )
                            )
                        else:
                            pp_outputs = PPProxyTensors(
                                {
                                    "next_token_ids": next_token_ids,
                                }
                            )
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
                        # send the output from the last round to let the next stage worker run post processing
                        self.pp_group.send_tensor_dict(
                            pp_outputs.tensors,
                            all_gather_group=self.attn_tp_group,
                        )

                # receive outputs and post-process (filter finished reqs) the coming microbatch
                next_mb_id = (mb_id + 1) % self.pp_size
                next_pp_outputs = None
                if mbs[next_mb_id] is not None:
                    next_pp_outputs: Optional[PPProxyTensors] = PPProxyTensors(
                        self.pp_group.recv_tensor_dict(
                            all_gather_group=self.attn_tp_group
                        )
                    )
                    mbs[next_mb_id].output_ids = next_pp_outputs["next_token_ids"]
935
936
937
938
939
940
941
942
943
                    logits_output_args = {
                        k[len("logits_output.") :]: v
                        for k, v in next_pp_outputs.tensors.items()
                        if k.startswith("logits_output.")
                    }
                    if len(logits_output_args) > 0:
                        logits_output = LogitsProcessorOutput(**logits_output_args)
                    else:
                        logits_output = None
944
                    output_result = GenerationBatchResult(
945
                        logits_output=logits_output,
946
947
                        pp_hidden_states_proxy_tensors=None,
                        next_token_ids=next_pp_outputs["next_token_ids"],
948
949
950
951
952
953
                        extend_input_len_per_req=next_pp_outputs.tensors.get(
                            "extend_input_len_per_req", None
                        ),
                        extend_logprob_start_len_per_req=next_pp_outputs.tensors.get(
                            "extend_logprob_start_len_per_req", None
                        ),
954
                        bid=bids[next_mb_id],
955
                        can_run_cuda_graph=result.can_run_cuda_graph,
956
957
958
959
                    )
                    self.process_batch_result(mbs[next_mb_id], output_result)
                    last_mbs[next_mb_id] = mbs[next_mb_id]

960
                # (not last rank)
961
962
963
                if not self.pp_group.is_last_rank:
                    if self.cur_batch:
                        bids[mb_id] = result.bid
964
965
                    # carry the outputs to the next stage
                    # send the outputs from the last round to let the next stage worker run post processing
966
967
968
969
970
971
972
                    if pp_outputs:
                        self.pp_group.send_tensor_dict(
                            pp_outputs.tensors,
                            all_gather_group=self.attn_tp_group,
                        )

                    # send out reqs to the next stage
973
                    dp_offset = self.attn_dp_rank * self.attn_tp_size
974
975
976
977
                    if self.attn_tp_rank == 0:
                        point_to_point_pyobj(
                            recv_reqs,
                            self.pp_rank * self.tp_size + dp_offset,
978
                            self.world_group.device_group,
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
                            self.pp_rank * self.tp_size + dp_offset,
                            (self.pp_rank + 1) * self.tp_size + dp_offset,
                        )

                    # send out proxy tensors to the next stage
                    if self.cur_batch:
                        self.pp_group.send_tensor_dict(
                            result.pp_hidden_states_proxy_tensors,
                            all_gather_group=self.attn_tp_group,
                        )

                pp_outputs = next_pp_outputs

            # When the server is idle, self-check and re-init some states
            if server_is_idle:
                self.check_memory()
Hanming Lu's avatar
Hanming Lu committed
995
                self.check_tree_cache()
996
                self.new_token_ratio = self.init_new_token_ratio
997
                self.maybe_sleep_on_idle()
998

999
1000
    def recv_requests(self) -> List[Req]:
        """Receive results at tp_rank = 0 and broadcast it to all other TP ranks."""
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
        if self.pp_rank == 0:
            if self.attn_tp_rank == 0:
                recv_reqs = []

                while True:
                    try:
                        recv_req = self.recv_from_tokenizer.recv_pyobj(zmq.NOBLOCK)
                    except zmq.ZMQError:
                        break
                    recv_reqs.append(recv_req)

                while True:
                    try:
                        recv_rpc = self.recv_from_rpc.recv_pyobj(zmq.NOBLOCK)
                    except zmq.ZMQError:
                        break
                    recv_reqs.append(recv_rpc)
            else:
                recv_reqs = None
Lianmin Zheng's avatar
Lianmin Zheng committed
1020
        else:
1021
            if self.attn_tp_rank == 0:
1022
                dp_offset = self.attn_dp_rank * self.attn_tp_size
1023
1024
1025
                recv_reqs = point_to_point_pyobj(
                    [],
                    self.pp_rank * self.tp_size + dp_offset,
1026
                    self.world_group.device_group,
1027
1028
1029
1030
1031
                    (self.pp_rank - 1) * self.tp_size + dp_offset,
                    self.pp_rank * self.tp_size + dp_offset,
                )
            else:
                recv_reqs = None
1032

1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
        if self.server_args.enable_dp_attention:
            if self.attn_tp_rank == 0:
                work_reqs = [
                    req
                    for req in recv_reqs
                    if isinstance(
                        req, (TokenizedGenerateReqInput, TokenizedEmbeddingReqInput)
                    )
                ]
                control_reqs = [
                    req
                    for req in recv_reqs
                    if not isinstance(
                        req, (TokenizedGenerateReqInput, TokenizedEmbeddingReqInput)
                    )
                ]
            else:
                work_reqs = None
                control_reqs = None

            if self.attn_tp_size != 1:
                work_reqs = broadcast_pyobj(
                    work_reqs,
1056
                    self.attn_tp_group.rank,
1057
                    self.attn_tp_cpu_group,
1058
                    src=self.attn_tp_group.ranks[0],
1059
1060
1061
                )
            if self.tp_size != 1:
                control_reqs = broadcast_pyobj(
1062
1063
1064
1065
                    control_reqs,
                    self.tp_group.rank,
                    self.tp_cpu_group,
                    src=self.tp_group.ranks[0],
1066
1067
1068
                )
            recv_reqs = work_reqs + control_reqs
        elif self.tp_size != 1:
1069
1070
1071
1072
1073
1074
            recv_reqs = broadcast_pyobj(
                recv_reqs,
                self.tp_group.rank,
                self.tp_cpu_group,
                src=self.tp_group.ranks[0],
            )
1075
1076
        return recv_reqs

Lianmin Zheng's avatar
Lianmin Zheng committed
1077
    def process_input_requests(self, recv_reqs: List):
1078
        for recv_req in recv_reqs:
1079
1080
            # If it is a health check generation request and there are running requests, ignore it.
            if is_health_check_generate_req(recv_req) and (
Lianmin Zheng's avatar
Lianmin Zheng committed
1081
                self.chunked_req is not None or not self.running_batch.is_empty()
1082
1083
1084
1085
            ):
                self.return_health_check_ct += 1
                continue

1086
            output = self._request_dispatcher(recv_req)
1087
            if output is not None:
1088
1089
1090
1091
1092
                if isinstance(output, RpcReqOutput):
                    if self.recv_from_rpc is not None:
                        self.recv_from_rpc.send_pyobj(output)
                else:
                    self.send_to_tokenizer.send_pyobj(output)
1093
1094
1095
1096
1097

    def handle_generate_request(
        self,
        recv_req: TokenizedGenerateReqInput,
    ):
1098
        # Create a new request
1099
1100
1101
1102
1103
        if (
            recv_req.session_params is None
            or recv_req.session_params.id is None
            or recv_req.session_params.id not in self.sessions
        ):
Rin Intachuen's avatar
Rin Intachuen committed
1104
1105
1106
1107
1108
1109
            if recv_req.input_embeds is not None:
                # Generate fake input_ids based on the length of input_embeds
                seq_length = len(recv_req.input_embeds)
                fake_input_ids = [1] * seq_length
                recv_req.input_ids = fake_input_ids

1110
1111
1112
1113
            if recv_req.bootstrap_port is None:
                # Use default bootstrap port
                recv_req.bootstrap_port = self.server_args.disaggregation_bootstrap_port

1114
1115
1116
1117
1118
            req = Req(
                recv_req.rid,
                recv_req.input_text,
                recv_req.input_ids,
                recv_req.sampling_params,
Lianmin Zheng's avatar
Lianmin Zheng committed
1119
1120
                return_logprob=recv_req.return_logprob,
                top_logprobs_num=recv_req.top_logprobs_num,
1121
                token_ids_logprob=recv_req.token_ids_logprob,
Lianmin Zheng's avatar
Lianmin Zheng committed
1122
                stream=recv_req.stream,
1123
                lora_path=recv_req.lora_path,
Rin Intachuen's avatar
Rin Intachuen committed
1124
                input_embeds=recv_req.input_embeds,
Lianmin Zheng's avatar
Lianmin Zheng committed
1125
                custom_logit_processor=recv_req.custom_logit_processor,
1126
                return_hidden_states=recv_req.return_hidden_states,
1127
                eos_token_ids=self.model_config.hf_eos_token_id,
1128
                bootstrap_host=recv_req.bootstrap_host,
1129
                bootstrap_port=recv_req.bootstrap_port,
1130
                bootstrap_room=recv_req.bootstrap_room,
1131
                data_parallel_rank=recv_req.data_parallel_rank,
1132
1133
            )
            req.tokenizer = self.tokenizer
Lianmin Zheng's avatar
Lianmin Zheng committed
1134

1135
1136
1137
            if self.disaggregation_mode != DisaggregationMode.NULL:
                # Invalid request for disaggregated mode
                if recv_req.bootstrap_room is None:
1138
                    error_msg = (
1139
1140
1141
                        f"Invalid request: Disaggregated request received without "
                        f"boostrap room id. {req.rid=}"
                    )
1142
1143
                    logger.error(error_msg)
                    prepare_abort(req, error_msg)
1144
1145
1146
                    self.stream_output([req], req.return_logprob)
                    return

1147
1148
1149
1150
            if (
                recv_req.session_params is not None
                and recv_req.session_params.id is not None
            ):
1151
                req.set_finish_with_abort(
1152
                    f"Invalid request: session id {recv_req.session_params.id} does not exist"
1153
                )
1154
                self._add_request_to_queue(req)
1155
1156
                return
        else:
1157
1158
            # Create a new request from a previous session
            session = self.sessions[recv_req.session_params.id]
1159
            req = session.create_req(recv_req, self.tokenizer)
1160
            if isinstance(req.finished_reason, FINISH_ABORT):
1161
                self._add_request_to_queue(req)
1162
                return
1163

1164
        # Handle multimodal inputs
Mick's avatar
Mick committed
1165
1166
        if recv_req.mm_inputs is not None:
            image_inputs = MultimodalInputs.from_dict(recv_req.mm_inputs)
1167
            # Expand a single image token into multiple dummy tokens for receiving image embeddings
1168
            req.origin_input_ids = self.pad_input_ids_func(
1169
                req.origin_input_ids, image_inputs
1170
            )
1171
            req.extend_image_inputs(image_inputs)
1172

1173
            if len(req.origin_input_ids) >= self.max_req_input_len:
1174
1175
1176
1177
1178
                req.set_finish_with_abort(
                    error_msg=(
                        "Multimodal prompt is too long after expanding multimodal tokens. "
                        f"After expanding {len(req.origin_input_ids_unpadded)=} => {len(req.origin_input_ids)} >= {self.max_req_input_len}."
                    )
1179
                )
1180
                self._add_request_to_queue(req)
1181
1182
                return

1183
        # Validate prompt length
1184
1185
1186
1187
1188
1189
        error_msg = validate_input_length(
            req,
            self.max_req_input_len,
            self.server_args.allow_auto_truncate,
        )
        if error_msg:
1190
            req.set_finish_with_abort(error_msg)
1191
            self._add_request_to_queue(req)
1192
            return
1193

1194
        # Copy more attributes
1195
        if recv_req.logprob_start_len == -1 or not recv_req.return_logprob:
1196
1197
1198
1199
1200
            # By default, only return the logprobs for output tokens
            req.logprob_start_len = len(req.origin_input_ids) - 1
        else:
            req.logprob_start_len = recv_req.logprob_start_len

1201
        if req.logprob_start_len >= len(req.origin_input_ids):
1202
            error_msg = f"{req.logprob_start_len=} is higher than the number of input tokens {len(req.origin_input_ids)=}. Please use a smaller logprob_start_len."
1203
            req.logprob_start_len = len(req.origin_input_ids) - 1
1204
            req.set_finish_with_abort(error_msg)
1205
1206
1207
            self._add_request_to_queue(req)
            return

1208
1209
1210
1211
1212
1213
        req.sampling_params.max_new_tokens = min(
            (
                req.sampling_params.max_new_tokens
                if req.sampling_params.max_new_tokens is not None
                else 1 << 30
            ),
1214
            self.max_req_len - len(req.origin_input_ids) - 1,
1215
1216
        )

1217
1218
1219
1220
1221
        # Init grammar cache for this request
        add_to_grammar_queue = False
        if (
            req.sampling_params.json_schema is not None
            or req.sampling_params.regex is not None
1222
            or req.sampling_params.ebnf is not None
1223
            or req.sampling_params.structural_tag is not None
1224
1225
1226
1227
1228
1229
        ):
            assert self.grammar_backend is not None
            if req.sampling_params.json_schema is not None:
                key = ("json", req.sampling_params.json_schema)
            elif req.sampling_params.regex is not None:
                key = ("regex", req.sampling_params.regex)
1230
1231
            elif req.sampling_params.ebnf is not None:
                key = ("ebnf", req.sampling_params.ebnf)
1232
1233
            elif req.sampling_params.structural_tag:
                key = ("structural_tag", req.sampling_params.structural_tag)
1234

1235
1236
1237
1238
1239
            value, cache_hit = self.grammar_backend.get_cached_or_future_value(key)
            req.grammar = value

            if not cache_hit:
                req.grammar_key = key
1240
                add_to_grammar_queue = True
1241
1242
1243
1244
            else:
                if value is INVALID_GRAMMAR_OBJ:  # We hit a cached invalid grammar.
                    error_msg = f"Invalid grammar request with cache hit: {key=}"
                    req.set_finish_with_abort(error_msg)
1245
1246

        if add_to_grammar_queue:
1247
            req.queue_time_start = time.perf_counter()
1248
1249
            self.grammar_queue.append(req)
        else:
1250
1251
1252
            self._add_request_to_queue(req)

    def _add_request_to_queue(self, req: Req):
1253
        req.queue_time_start = time.perf_counter()
Byron Hsu's avatar
Byron Hsu committed
1254
        if self.disaggregation_mode == DisaggregationMode.PREFILL:
Byron Hsu's avatar
Byron Hsu committed
1255
1256
1257
            self.disagg_prefill_bootstrap_queue.add(
                req, self.model_config.num_key_value_heads
            )
Byron Hsu's avatar
Byron Hsu committed
1258
1259
1260
        elif self.disaggregation_mode == DisaggregationMode.DECODE:
            self.disagg_decode_prealloc_queue.add(req)
        else:
1261
1262
1263
1264
1265
1266
1267
1268
1269
            if self.enable_hicache_storage:
                req.init_next_round_input(self.tree_cache)
                last_hash = req.last_host_node.get_last_hash_value()
                matched_len = len(req.prefix_indices) + req.host_hit_length
                if (matched_len > 0 and last_hash is not None) or matched_len == 0:
                    new_input_tokens = req.fill_ids[matched_len:]
                    self.tree_cache.prefetch_from_storage(
                        req.rid, req.last_host_node, new_input_tokens, last_hash
                    )
Byron Hsu's avatar
Byron Hsu committed
1270
1271
            self.waiting_queue.append(req)

1272
    def _extend_requests_to_queue(self, reqs: List[Req], is_retracted: bool = False):
1273
        if self.disaggregation_mode == DisaggregationMode.PREFILL:
Byron Hsu's avatar
Byron Hsu committed
1274
1275
1276
            self.disagg_prefill_bootstrap_queue.extend(
                reqs, self.model_config.num_key_value_heads
            )
1277
1278
        elif self.disaggregation_mode == DisaggregationMode.DECODE:
            # If this is a decode server, we put the request to the decode pending prealloc queue
1279
            self.disagg_decode_prealloc_queue.extend(reqs, is_retracted)
Byron Hsu's avatar
Byron Hsu committed
1280
1281
        else:
            self.waiting_queue.extend(reqs)
1282
1283
1284

    def handle_embedding_request(
        self,
1285
        recv_req: TokenizedEmbeddingReqInput,
1286
1287
1288
1289
1290
1291
    ):
        req = Req(
            recv_req.rid,
            recv_req.input_text,
            recv_req.input_ids,
            recv_req.sampling_params,
woodx's avatar
woodx committed
1292
            token_type_ids=recv_req.token_type_ids,
1293
1294
1295
        )
        req.tokenizer = self.tokenizer

1296
1297
        # Handle multimodal inputs
        if recv_req.image_inputs is not None:
Mick's avatar
Mick committed
1298
            image_inputs = MultimodalInputs.from_dict(recv_req.image_inputs)
1299
1300
1301
1302
1303
1304
1305
            # Expand a single image token into multiple dummy tokens for receiving image embeddings
            req.origin_input_ids = self.pad_input_ids_func(
                req.origin_input_ids, image_inputs
            )
            req.extend_image_inputs(image_inputs)

            if len(req.origin_input_ids) >= self.max_req_input_len:
1306
1307
1308
1309
1310
                req.set_finish_with_abort(
                    error_msg=(
                        "Multimodal prompt is too long after expanding multimodal tokens. "
                        f"After expanding {len(req.origin_input_ids_unpadded)=} => {len(req.origin_input_ids)} >= {self.max_req_input_len}."
                    )
1311
                )
1312
                self._add_request_to_queue(req)
1313
1314
                return

1315
        # Validate prompts length
1316
        error_msg = validate_input_length(
1317
1318
1319
1320
            req,
            self.max_req_input_len,
            self.server_args.allow_auto_truncate,
        )
1321
        if error_msg:
1322
            self._add_request_to_queue(req)
1323
            return
1324

1325
1326
        # Copy more attributes
        req.logprob_start_len = len(req.origin_input_ids) - 1
1327
        self._add_request_to_queue(req)
1328

1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
    def _emit_kv_metrics(self):
        kv_metrics = KvMetrics()
        kv_metrics.request_active_slots = self.stats.num_running_reqs
        kv_metrics.request_total_slots = self.max_running_requests
        kv_metrics.kv_active_blocks = int(
            self.stats.token_usage * self.max_total_num_tokens
        )
        kv_metrics.kv_total_blocks = self.max_total_num_tokens
        kv_metrics.num_requests_waiting = self.stats.num_queue_reqs
        kv_metrics.gpu_cache_usage_perc = self.stats.token_usage
        kv_metrics.gpu_prefix_cache_hit_rate = self.stats.cache_hit_rate
        kv_metrics.data_parallel_rank = self.dp_rank if self.dp_rank is not None else 0

        if not self.send_metrics_from_scheduler.closed:
            self.send_metrics_from_scheduler.send_pyobj(kv_metrics)

1345
1346
1347
1348
    def log_prefill_stats(
        self,
        adder: PrefillAdder,
        can_run_list: List[Req],
1349
        running_bs: int,
1350
    ):
1351
1352
        gap_latency = time.perf_counter() - self.last_prefill_stats_tic
        self.last_prefill_stats_tic = time.perf_counter()
Liangsheng Yin's avatar
Liangsheng Yin committed
1353
1354
        self.last_input_throughput = self.last_prefill_tokens / gap_latency
        self.last_prefill_tokens = adder.log_input_tokens
Lianmin Zheng's avatar
Lianmin Zheng committed
1355

Hanming Lu's avatar
Hanming Lu committed
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
        if self.is_hybrid:
            (
                full_num_used,
                swa_num_used,
                full_token_usage,
                swa_token_usage,
                _,
                _,
                _,
                _,
            ) = self._get_swa_token_info()
            num_used = max(full_num_used, swa_num_used)
            token_usage = max(full_token_usage, swa_token_usage)
            token_msg = (
                f"full token usage: {full_token_usage:.2f}, "
                f"swa token usage: {swa_token_usage:.2f}, "
            )
        else:
            num_used, token_usage, _, _ = self._get_token_info()
            token_msg = f"token usage: {token_usage:.2f}, "
1376

1377
        num_new_seq = len(can_run_list)
1378
        f = (
1379
            f"Prefill batch. "
1380
            f"#new-seq: {num_new_seq}, "
1381
1382
            f"#new-token: {adder.log_input_tokens}, "
            f"#cached-token: {adder.log_hit_tokens}, "
Hanming Lu's avatar
Hanming Lu committed
1383
            f"{token_msg}"
1384
        )
Liangsheng Yin's avatar
Liangsheng Yin committed
1385
1386
1387
1388

        if self.disaggregation_mode == DisaggregationMode.PREFILL:
            f += f"#unbootstrapped-req: {len(self.disagg_prefill_bootstrap_queue.queue)}, "
            f += f"#queue-req: {len(self.waiting_queue)}, "
fzyzcjy's avatar
fzyzcjy committed
1389
            f += f"#transferring-req: {len(self.disagg_prefill_inflight_queue)}, "
1390
            f += f"input throughput (token/s): {self.last_input_throughput:.2f}, "
Liangsheng Yin's avatar
Liangsheng Yin committed
1391
        else:
Liangsheng Yin's avatar
Liangsheng Yin committed
1392
            f += f"#running-req: {running_bs}, "
1393
1394
            f += f"#queue-req: {len(self.waiting_queue)}, "

1395
        logger.info(f)
1396
1397

        if self.enable_metrics:
1398
1399
1400
            cache_hit_rate = adder.log_hit_tokens / (
                adder.log_input_tokens + adder.log_hit_tokens
            )
1401
1402
            self.stats.num_running_reqs = running_bs
            self.stats.num_used_tokens = num_used
Hanming Lu's avatar
Hanming Lu committed
1403
            self.stats.token_usage = round(token_usage, 2)
1404
1405
            self.stats.num_queue_reqs = len(self.waiting_queue)
            self.stats.cache_hit_rate = cache_hit_rate
1406
1407
1408
1409
1410
1411

            total_queue_latency = 0
            for req in can_run_list:
                total_queue_latency += req.queue_time_end - req.queue_time_start
            self.stats.avg_request_queue_latency = total_queue_latency / num_new_seq

1412
            self.metrics_collector.log_stats(self.stats)
1413
            self._emit_kv_metrics()
1414
        self._publish_kv_events()
1415

1416
1417
1418
    def log_decode_stats(
        self, can_run_cuda_graph: bool, running_batch: ScheduleBatch = None
    ):
1419
1420
        batch = running_batch or self.running_batch

1421
1422
        gap_latency = time.perf_counter() - self.last_decode_stats_tic
        self.last_decode_stats_tic = time.perf_counter()
1423
1424
        self.last_gen_throughput = self.num_generated_tokens / gap_latency
        self.num_generated_tokens = 0
1425
        num_running_reqs = len(batch.reqs)
Hanming Lu's avatar
Hanming Lu committed
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
        if self.is_hybrid:
            (
                full_num_used,
                swa_num_used,
                full_token_usage,
                swa_token_usage,
                _,
                _,
                _,
                _,
            ) = self._get_swa_token_info()
            num_used = max(full_num_used, swa_num_used)
            token_usage = max(full_token_usage, swa_token_usage)
            token_msg = (
                f"#full token: {full_num_used}, "
                f"full token usage: {full_token_usage:.2f}, "
                f"#swa token: {swa_num_used}, "
                f"swa token usage: {swa_token_usage:.2f}, "
            )
        else:
            num_used, token_usage, _, _ = self._get_token_info()
            token_msg = f"#token: {num_used}, " f"token usage: {token_usage:.2f}, "
1448
1449
1450
1451
1452

        if RECORD_STEP_TIME:
            self.step_time_dict[num_running_reqs].append(
                gap_latency / self.server_args.decode_log_interval
            )
Lianmin Zheng's avatar
Lianmin Zheng committed
1453

Hanming Lu's avatar
Hanming Lu committed
1454
        msg = f"Decode batch. #running-req: {num_running_reqs}, {token_msg}"
Liangsheng Yin's avatar
Liangsheng Yin committed
1455

1456
        if self.spec_algorithm.is_none():
1457
            spec_accept_length = 0
1458
        else:
1459
            spec_accept_length = (
1460
1461
                self.spec_num_total_accepted_tokens / self.spec_num_total_forward_ct
            )
1462
1463
            self.cum_spec_accept_length += self.spec_num_total_accepted_tokens
            self.cum_spec_accept_count += self.spec_num_total_forward_ct
1464
            self.spec_num_total_accepted_tokens = self.spec_num_total_forward_ct = 0
Liangsheng Yin's avatar
Liangsheng Yin committed
1465
1466
1467
            msg += f"accept len: {spec_accept_length:.2f}, "

        if self.disaggregation_mode == DisaggregationMode.DECODE:
1468
            msg += f"pre-allocated usage: {self.disagg_decode_prealloc_queue.num_tokens_pre_allocated / self.max_total_num_tokens:.2f}, "
1469
            msg += f"#retracted-req: {len(self.disagg_decode_prealloc_queue.retracted_queue)}, "
Liangsheng Yin's avatar
Liangsheng Yin committed
1470
1471

        msg += (
1472
            f"cuda graph: {can_run_cuda_graph}, "
Liangsheng Yin's avatar
Liangsheng Yin committed
1473
            f"gen throughput (token/s): {self.last_gen_throughput:.2f}, "
1474
            f"#queue-req: {len(self.waiting_queue)}, "
Liangsheng Yin's avatar
Liangsheng Yin committed
1475
        )
1476
1477

        logger.info(msg)
1478
1479
1480
        if self.enable_metrics:
            self.stats.num_running_reqs = num_running_reqs
            self.stats.num_used_tokens = num_used
Hanming Lu's avatar
Hanming Lu committed
1481
            self.stats.token_usage = round(token_usage, 2)
1482
1483
            self.stats.cache_hit_rate = 0.0
            self.stats.gen_throughput = self.last_gen_throughput
1484
            self.stats.num_queue_reqs = len(self.waiting_queue)
1485
            self.stats.num_grammar_queue_reqs = len(self.grammar_queue)
1486
            self.stats.spec_accept_length = spec_accept_length
1487
            self.stats.total_retracted_reqs = self.total_retracted_reqs
1488
            self.metrics_collector.log_stats(self.stats)
1489
            self._emit_kv_metrics()
1490
        self._publish_kv_events()
1491

Lianmin Zheng's avatar
Lianmin Zheng committed
1492
    def check_memory(self):
Hanming Lu's avatar
Hanming Lu committed
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
        if self.is_hybrid:
            (
                full_num_used,
                swa_num_used,
                _,
                _,
                full_available_size,
                full_evictable_size,
                swa_available_size,
                swa_evictable_size,
            ) = self._get_swa_token_info()
            memory_leak = full_num_used != 0 or swa_num_used != 0
            token_msg = (
                f"{self.full_tokens_per_layer=}, {full_available_size=}, {full_evictable_size=}, {self.tree_cache.full_protected_size()=}\n"
                f"{self.swa_tokens_per_layer=}, {swa_available_size=}, {swa_evictable_size=}, {self.tree_cache.swa_protected_size()=}\n"
            )
tarinkk's avatar
tarinkk committed
1509
        else:
Hanming Lu's avatar
Hanming Lu committed
1510
1511
1512
1513
1514
1515
            _, _, available_size, evictable_size = self._get_token_info()
            protected_size = self.tree_cache.protected_size()
            memory_leak = (available_size + evictable_size) != (
                self.max_total_num_tokens
                if not self.enable_hierarchical_cache
                else self.max_total_num_tokens - protected_size
Lianmin Zheng's avatar
Lianmin Zheng committed
1516
            )
Hanming Lu's avatar
Hanming Lu committed
1517
1518
1519
1520
            token_msg = f"{self.max_total_num_tokens=}, {available_size=}, {evictable_size=}, {protected_size=}\n"

        if memory_leak:
            msg = "token_to_kv_pool_allocator memory leak detected! " f"{token_msg}"
Lianmin Zheng's avatar
Lianmin Zheng committed
1521
            raise ValueError(msg)
Lianmin Zheng's avatar
Lianmin Zheng committed
1522

1523
1524
1525
1526
1527
1528
1529
1530
        if self.disaggregation_mode == DisaggregationMode.DECODE:
            req_total_size = (
                self.req_to_token_pool.size + self.req_to_token_pool.pre_alloc_size
            )
        else:
            req_total_size = self.req_to_token_pool.size

        if len(self.req_to_token_pool.free_slots) != req_total_size:
1531
            msg = (
1532
                "req_to_token_pool memory leak detected!"
1533
1534
                f"available_size={len(self.req_to_token_pool.free_slots)}, "
                f"total_size={self.req_to_token_pool.size}\n"
Lianmin Zheng's avatar
Lianmin Zheng committed
1535
            )
Lianmin Zheng's avatar
Lianmin Zheng committed
1536
            raise ValueError(msg)
Lianmin Zheng's avatar
Lianmin Zheng committed
1537

1538
1539
        if (
            self.enable_metrics
1540
            and self.current_scheduler_metrics_enabled()
1541
            and time.perf_counter() > self.metrics_collector.last_log_time + 30
1542
1543
        ):
            # During idle time, also collect metrics every 30 seconds.
Hanming Lu's avatar
Hanming Lu committed
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
            if self.is_hybrid:
                (
                    full_num_used,
                    swa_num_used,
                    full_token_usage,
                    swa_token_usage,
                    _,
                    _,
                    _,
                    _,
                ) = self._get_swa_token_info()
                num_used = max(full_num_used, swa_num_used)
                token_usage = max(full_token_usage, swa_token_usage)
            else:
                num_used, token_usage, _, _ = self._get_token_info()
Lianmin Zheng's avatar
Lianmin Zheng committed
1559
            num_running_reqs = len(self.running_batch.reqs)
1560
1561
            self.stats.num_running_reqs = num_running_reqs
            self.stats.num_used_tokens = num_used
Hanming Lu's avatar
Hanming Lu committed
1562
            self.stats.token_usage = round(token_usage, 2)
1563
1564
            self.stats.gen_throughput = 0
            self.stats.num_queue_reqs = len(self.waiting_queue)
1565
            self.stats.num_grammar_queue_reqs = len(self.grammar_queue)
1566
            self.metrics_collector.log_stats(self.stats)
1567
        self._publish_kv_events()
1568

Hanming Lu's avatar
Hanming Lu committed
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
    def check_tree_cache(self):
        if self.is_hybrid and isinstance(self.tree_cache, SWARadixCache):
            self.tree_cache.sanity_check()

    def _get_token_info(self):
        available_size = self.token_to_kv_pool_allocator.available_size()
        evictable_size = self.tree_cache.evictable_size()
        num_used = self.max_total_num_tokens - (available_size + evictable_size)
        token_usage = num_used / self.max_total_num_tokens
        return num_used, token_usage, available_size, evictable_size

    def _get_swa_token_info(self):
        full_available_size = self.token_to_kv_pool_allocator.full_available_size()
        full_evictable_size = self.tree_cache.full_evictable_size()
        swa_available_size = self.token_to_kv_pool_allocator.swa_available_size()
        swa_evictable_size = self.tree_cache.swa_evictable_size()
        full_num_used = self.full_tokens_per_layer - (
            full_available_size + full_evictable_size
        )
        swa_num_used = self.swa_tokens_per_layer - (
            swa_available_size + swa_evictable_size
        )
        full_token_usage = full_num_used / self.full_tokens_per_layer
        swa_token_usage = swa_num_used / self.swa_tokens_per_layer
        return (
            full_num_used,
            swa_num_used,
            full_token_usage,
            swa_token_usage,
            full_available_size,
            full_evictable_size,
            swa_available_size,
            swa_evictable_size,
        )

1604
    def get_next_batch_to_run(self) -> Optional[ScheduleBatch]:
1605
        # Merge the prefill batch into the running batch
1606
1607
1608
1609
1610
1611
1612
1613
        chunked_req_to_exclude = set()
        if self.chunked_req:
            # Move the chunked request out of the batch so that we can merge
            # only finished requests to running_batch.
            chunked_req_to_exclude.add(self.chunked_req)
            self.tree_cache.cache_unfinished_req(self.chunked_req)
            # chunked request keeps its rid but will get a new req_pool_idx
            self.req_to_token_pool.free(self.chunked_req.req_pool_idx)
Lianmin Zheng's avatar
Lianmin Zheng committed
1614
        if self.last_batch and self.last_batch.forward_mode.is_extend():
1615
1616
1617
1618
            if self.last_batch.chunked_req is not None:
                # In the context pipeline parallelism, after the last chunk, the current microbatch still track outdated chunked_req.
                # We need to discard it.
                chunked_req_to_exclude.add(self.last_batch.chunked_req)
Lianmin Zheng's avatar
Lianmin Zheng committed
1619

1620
            # Filter batch
1621
            last_bs = self.last_batch.batch_size()
1622
1623
1624
            self.last_batch.filter_batch(
                chunked_req_to_exclude=list(chunked_req_to_exclude)
            )
1625
            if self.last_batch.batch_size() < last_bs:
Lianmin Zheng's avatar
Lianmin Zheng committed
1626
                self.running_batch.batch_is_full = False
1627

1628
            # Merge the new batch into the running batch
1629
            if not self.last_batch.is_empty():
Lianmin Zheng's avatar
Lianmin Zheng committed
1630
                if self.running_batch.is_empty():
1631
1632
                    self.running_batch = self.last_batch
                else:
Lianmin Zheng's avatar
Lianmin Zheng committed
1633
                    # Merge running_batch with prefill batch
1634
                    self.running_batch.merge_batch(self.last_batch)
1635

1636
        new_batch = self.get_new_batch_prefill()
1637

1638
1639
1640
1641
1642
        need_dp_attn_preparation = require_mlp_sync(self.server_args)

        if need_dp_attn_preparation and not self.spec_algorithm.is_none():
            # In speculative decoding, prefill batches and decode batches cannot be processed in the same DP attention group.
            # We prepare idle batches in advance to skip preparing decode batches when there are prefill batches in the group.
1643
            new_batch = self.prepare_mlp_sync_batch(new_batch)
1644
1645
1646
            need_dp_attn_preparation = new_batch is None

        if new_batch is not None:
1647
1648
1649
1650
            # Run prefill first if possible
            ret = new_batch
        else:
            # Run decode
Lianmin Zheng's avatar
Lianmin Zheng committed
1651
            if not self.running_batch.is_empty():
1652
                self.running_batch = self.update_running_batch(self.running_batch)
Lianmin Zheng's avatar
Lianmin Zheng committed
1653
1654
1655
                ret = self.running_batch if not self.running_batch.is_empty() else None
            else:
                ret = None
1656

1657
1658
        # Handle DP attention
        if need_dp_attn_preparation:
1659
            ret = self.prepare_mlp_sync_batch(ret)
1660
1661

        return ret
1662

1663
1664
1665
1666
1667
1668
    def get_num_allocatable_reqs(self, running_bs):
        res = global_server_args_dict["max_micro_batch_size"] - running_bs
        if self.pp_size > 1:
            res = min(res, self.req_to_token_pool.available_size())
        return res

Lianmin Zheng's avatar
Lianmin Zheng committed
1669
    def get_new_batch_prefill(self) -> Optional[ScheduleBatch]:
Lianmin Zheng's avatar
Lianmin Zheng committed
1670
        # Check if the grammar is ready in the grammar queue
1671
        if self.grammar_queue:
1672
            self.move_ready_grammar_requests()
1673

Lianmin Zheng's avatar
Lianmin Zheng committed
1674
1675
        # Handle the cases where prefill is not allowed
        if (
Lianmin Zheng's avatar
Lianmin Zheng committed
1676
            self.running_batch.batch_is_full or len(self.waiting_queue) == 0
1677
        ) and self.chunked_req is None:
Lianmin Zheng's avatar
Lianmin Zheng committed
1678
1679
            return None

Lianmin Zheng's avatar
Lianmin Zheng committed
1680
        running_bs = len(self.running_batch.reqs)
1681
        # Ignore the check if self.chunked_req is not None.
1682
1683
1684
1685
1686
        # In the non-PP case, when self.chunked_req is not None, num_allocatable_reqs should always be greater than 0,
        # as the space for the chunked request has just been released.
        # In PP case, a chunked req can start in one microbatch and end in another microbatch, so the max_running_requests per microbatch should not be strict.
        # Instead, we should always allow chunked request to be added, otherwise, there will be a memory leak.
        if self.get_num_allocatable_reqs(running_bs) <= 0 and not self.chunked_req:
Lianmin Zheng's avatar
Lianmin Zheng committed
1687
            self.running_batch.batch_is_full = True
1688
1689
            return None

1690
        if self.enable_hierarchical_cache:
1691
            self.tree_cache.check_hicache_events()
1692

1693
        # Get priority queue
1694
        self.policy.calc_priority(self.waiting_queue)
1695

Lianmin Zheng's avatar
Lianmin Zheng committed
1696
        # Prefill policy
1697
        adder = PrefillAdder(
1698
            self.page_size,
1699
            self.tree_cache,
1700
            self.token_to_kv_pool_allocator,
1701
1702
1703
1704
            self.running_batch,
            self.new_token_ratio,
            self.max_prefill_tokens,
            self.chunked_prefill_size,
1705
            running_bs if self.is_mixed_chunk else 0,
1706
1707
        )

Lianmin Zheng's avatar
Lianmin Zheng committed
1708
        if self.chunked_req is not None:
1709
1710
            self.chunked_req.init_next_round_input()
            self.chunked_req = adder.add_chunked_req(self.chunked_req)
1711

1712
        if self.enable_lora:
Lianmin Zheng's avatar
Lianmin Zheng committed
1713
1714
            lora_set = set([req.lora_path for req in self.running_batch.reqs])

1715
        # Get requests from the waiting queue to a new prefill batch
1716
1717
        for req in self.waiting_queue:
            if (
1718
                self.enable_lora
1719
1720
1721
1722
1723
1724
1725
                and len(
                    lora_set
                    | set([req.lora_path for req in adder.can_run_list])
                    | set([req.lora_path])
                )
                > self.max_loras_per_batch
            ):
Lianmin Zheng's avatar
Lianmin Zheng committed
1726
                self.running_batch.batch_is_full = True
1727
1728
                break

1729
            if len(adder.can_run_list) >= self.get_num_allocatable_reqs(running_bs):
Lianmin Zheng's avatar
Lianmin Zheng committed
1730
                self.running_batch.batch_is_full = True
1731
                break
1732

Byron Hsu's avatar
Byron Hsu committed
1733
1734
1735
1736
1737
1738
1739
            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

1740
1741
1742
            if self.enable_hicache_storage:
                self.tree_cache.check_prefetch_progress(req.rid)

1743
1744
            req.init_next_round_input(self.tree_cache)
            res = adder.add_one_req(req, has_chunked_req=(self.chunked_req is not None))
1745

1746
1747
            if res != AddReqResult.CONTINUE:
                if res == AddReqResult.NO_TOKEN:
1748
1749
                    if self.enable_hierarchical_cache:
                        # Set batch_is_full after making sure there are requests that can be served
Lianmin Zheng's avatar
Lianmin Zheng committed
1750
1751
                        self.running_batch.batch_is_full = len(
                            adder.can_run_list
1752
                        ) > 0 or (not self.running_batch.is_empty())
1753
                    else:
Lianmin Zheng's avatar
Lianmin Zheng committed
1754
                        self.running_batch.batch_is_full = True
1755
1756
                break

Lianmin Zheng's avatar
Lianmin Zheng committed
1757
        # Update waiting queue
1758
        can_run_list: List[Req] = adder.can_run_list
Lianmin Zheng's avatar
Lianmin Zheng committed
1759
1760
        if len(can_run_list) == 0:
            return None
1761
1762
1763
1764

        if self.enable_metrics:
            # only record queue time when enable_metrics is True to avoid overhead
            for req in can_run_list:
1765
                req.queue_time_end = time.perf_counter()
1766

Lianmin Zheng's avatar
Lianmin Zheng committed
1767
1768
1769
        self.waiting_queue = [
            x for x in self.waiting_queue if x not in set(can_run_list)
        ]
1770

1771
1772
1773
        if adder.new_chunked_req is not None:
            assert self.chunked_req is None
            self.chunked_req = adder.new_chunked_req
1774

1775
1776
        if self.chunked_req:
            self.chunked_req.is_chunked += 1
Lianmin Zheng's avatar
Lianmin Zheng committed
1777

1778
        # Print stats
1779
        if self.current_scheduler_metrics_enabled():
1780
            self.log_prefill_stats(adder, can_run_list, running_bs)
1781

Lianmin Zheng's avatar
Lianmin Zheng committed
1782
        # Create a new batch
1783
1784
1785
        new_batch = ScheduleBatch.init_new(
            can_run_list,
            self.req_to_token_pool,
1786
            self.token_to_kv_pool_allocator,
1787
            self.tree_cache,
1788
            self.model_config,
1789
            self.enable_overlap,
1790
            self.spec_algorithm,
1791
            self.server_args.enable_custom_logit_processor,
1792
            chunked_req=self.chunked_req,
1793
        )
1794
1795
        if self.enable_hierarchical_cache:
            # todo (zhiqiang): disable cuda graph execution if hicache loading triggered
1796
1797
1798
            new_batch.hicache_consumer_index = (
                self.tree_cache.ready_to_load_host_cache()
            )
1799

1800
        new_batch.prepare_for_extend()
1801

Lianmin Zheng's avatar
Lianmin Zheng committed
1802
        # Mixed-style chunked prefill
1803
1804
        if (
            self.is_mixed_chunk
Lianmin Zheng's avatar
Lianmin Zheng committed
1805
            and not self.running_batch.is_empty()
1806
1807
1808
            and not (new_batch.return_logprob or self.running_batch.return_logprob)
        ):
            # TODO (lianmin): support return_logprob + mixed chunked prefill
1809
1810
            self.running_batch.filter_batch()
            if not self.running_batch.is_empty():
1811
                self.running_batch.prepare_for_decode()
1812
1813
                new_batch.mix_with_running(self.running_batch)
                new_batch.decoding_reqs = self.running_batch.reqs
Lianmin Zheng's avatar
Lianmin Zheng committed
1814
1815
1816
            self.running_batch = ScheduleBatch(
                reqs=[], batch_is_full=self.running_batch.batch_is_full
            )
Lianmin Zheng's avatar
Lianmin Zheng committed
1817
1818
        else:
            new_batch.decoding_reqs = None
Lianmin Zheng's avatar
Lianmin Zheng committed
1819
1820
1821

        return new_batch

Lianmin Zheng's avatar
Lianmin Zheng committed
1822
    def update_running_batch(self, batch: ScheduleBatch) -> Optional[ScheduleBatch]:
1823
        """Update the current running decoding batch."""
Lianmin Zheng's avatar
Lianmin Zheng committed
1824
        initial_bs = batch.batch_size()
Lianmin Zheng's avatar
Lianmin Zheng committed
1825

1826
1827
        batch.filter_batch()
        if batch.is_empty():
Lianmin Zheng's avatar
Lianmin Zheng committed
1828
1829
            batch.batch_is_full = False
            return batch
1830

Lianmin Zheng's avatar
Lianmin Zheng committed
1831
        # Check if decode out of memory
1832
        if not batch.check_decode_mem(self.decode_mem_cache_buf_multiplier) or (
1833
            TEST_RETRACT and batch.batch_size() > 10
1834
        ):
Lianmin Zheng's avatar
Lianmin Zheng committed
1835
1836
            old_ratio = self.new_token_ratio

1837
            retracted_reqs, new_token_ratio = batch.retract_decode(self.server_args)
1838
            num_retracted_reqs = len(retracted_reqs)
Lianmin Zheng's avatar
Lianmin Zheng committed
1839
            self.new_token_ratio = new_token_ratio
1840

Lianmin Zheng's avatar
Lianmin Zheng committed
1841
            logger.info(
1842
                "KV cache pool is full. Retract requests. "
1843
                f"#retracted_reqs: {num_retracted_reqs}, "
Lianmin Zheng's avatar
Lianmin Zheng committed
1844
1845
                f"#new_token_ratio: {old_ratio:.4f} -> {self.new_token_ratio:.4f}"
            )
1846

1847
            self._extend_requests_to_queue(retracted_reqs, is_retracted=True)
1848
            self.total_retracted_reqs += num_retracted_reqs
Lianmin Zheng's avatar
Lianmin Zheng committed
1849
1850
        else:
            self.new_token_ratio = max(
1851
                self.new_token_ratio - self.new_token_ratio_decay,
Lianmin Zheng's avatar
Lianmin Zheng committed
1852
1853
1854
                self.min_new_token_ratio,
            )

Lianmin Zheng's avatar
Lianmin Zheng committed
1855
        if batch.batch_size() < initial_bs:
Lianmin Zheng's avatar
Lianmin Zheng committed
1856
            batch.batch_is_full = False
Lianmin Zheng's avatar
Lianmin Zheng committed
1857
1858

        # Update batch tensors
1859
        batch.prepare_for_decode()
Lianmin Zheng's avatar
Lianmin Zheng committed
1860
        return batch
Lianmin Zheng's avatar
Lianmin Zheng committed
1861

1862
1863
1864
    def run_batch(
        self, batch: ScheduleBatch
    ) -> Union[GenerationBatchResult, EmbeddingBatchResult]:
1865
        """Run a batch."""
Lianmin Zheng's avatar
Lianmin Zheng committed
1866
1867
        self.forward_ct += 1

1868
1869
        # Whether to run the profiler
        self._profile_batch_predicate(batch)
1870
1871
1872
1873
        if self.forward_sleep_time is not None:
            logger.info(f"Scheduler.run_batch sleep {self.forward_sleep_time}s")
            time.sleep(self.forward_sleep_time)

1874
        # Run forward
1875
        if self.is_generation:
1876
1877
            if self.spec_algorithm.is_none():
                model_worker_batch = batch.get_model_worker_batch()
1878
1879
1880
1881
1882

                # update the consumer index of hicache to the running batch
                self.tp_worker.set_hicache_consumer(
                    model_worker_batch.hicache_consumer_index
                )
1883
                if self.pp_group.is_last_rank:
1884
                    logits_output, next_token_ids, can_run_cuda_graph = (
1885
1886
1887
                        self.tp_worker.forward_batch_generation(model_worker_batch)
                    )
                else:
1888
                    pp_hidden_states_proxy_tensors, _, can_run_cuda_graph = (
1889
1890
                        self.tp_worker.forward_batch_generation(model_worker_batch)
                    )
1891
                bid = model_worker_batch.bid
Lianmin Zheng's avatar
Lianmin Zheng committed
1892
            else:
1893
1894
1895
                (
                    logits_output,
                    next_token_ids,
1896
                    bid,
1897
                    num_accepted_tokens,
1898
                    can_run_cuda_graph,
1899
                ) = self.draft_worker.forward_batch_speculative_generation(batch)
1900
1901
1902
                bs = batch.batch_size()
                self.spec_num_total_accepted_tokens += num_accepted_tokens + bs
                self.spec_num_total_forward_ct += bs
1903
                self.num_generated_tokens += num_accepted_tokens
1904
1905
1906

            if self.pp_group.is_last_rank:
                batch.output_ids = next_token_ids
1907

1908
1909
1910
            # These 2 values are needed for processing the output, but the values can be
            # modified by overlap schedule. So we have to copy them here so that
            # we can use the correct values in output processing.
1911
            if batch.return_logprob or self.spec_algorithm.is_eagle():
1912
                extend_input_len_per_req = [req.extend_input_len for req in batch.reqs]
1913
1914
1915
            else:
                extend_input_len_per_req = None
            if batch.return_logprob:
1916
1917
1918
1919
1920
1921
                extend_logprob_start_len_per_req = [
                    req.extend_logprob_start_len for req in batch.reqs
                ]
            else:
                extend_logprob_start_len_per_req = None

1922
            ret = GenerationBatchResult(
1923
1924
1925
1926
1927
1928
1929
                logits_output=logits_output if self.pp_group.is_last_rank else None,
                pp_hidden_states_proxy_tensors=(
                    pp_hidden_states_proxy_tensors
                    if not self.pp_group.is_last_rank
                    else None
                ),
                next_token_ids=next_token_ids if self.pp_group.is_last_rank else None,
1930
1931
                extend_input_len_per_req=extend_input_len_per_req,
                extend_logprob_start_len_per_req=extend_logprob_start_len_per_req,
1932
                bid=bid,
1933
                can_run_cuda_graph=can_run_cuda_graph,
1934
            )
Lianmin Zheng's avatar
Lianmin Zheng committed
1935
1936
1937
        else:  # embedding or reward model
            model_worker_batch = batch.get_model_worker_batch()
            embeddings = self.tp_worker.forward_batch_embedding(model_worker_batch)
1938
1939
1940
            ret = EmbeddingBatchResult(
                embeddings=embeddings, bid=model_worker_batch.bid
            )
1941
        return ret
Chayenne's avatar
Chayenne committed
1942

1943
1944
1945
1946
    def process_batch_result(
        self,
        batch: ScheduleBatch,
        result: Union[GenerationBatchResult, EmbeddingBatchResult],
1947
        launch_done: Optional[threading.Event] = None,
1948
    ):
Lianmin Zheng's avatar
Lianmin Zheng committed
1949
        if batch.forward_mode.is_decode():
1950
            self.process_batch_result_decode(batch, result, launch_done)
1951
        elif batch.forward_mode.is_extend():
1952
            self.process_batch_result_prefill(batch, result, launch_done)
1953
1954
        elif batch.forward_mode.is_idle():
            if self.enable_overlap:
1955
                self.tp_worker.resolve_last_batch_result(launch_done)
1956
                self.set_next_batch_sampling_info_done(batch)
1957
        elif batch.forward_mode.is_dummy_first():
1958
            self.set_next_batch_sampling_info_done(batch)
Lianmin Zheng's avatar
Lianmin Zheng committed
1959

1960
1961
1962
1963
1964
1965
1966
        if self.return_health_check_ct:
            # Return some signal for the health check.
            # This is used to prevent the health check signal being blocked by long context prefill.
            # However, one minor issue is that this code path does not check the status of detokenizer manager.
            self.return_health_check_ct -= 1
            self.send_to_tokenizer.send_pyobj(HealthCheckOutput())

1967
1968
    def prepare_mlp_sync_batch(self, local_batch: ScheduleBatch):
        return self.prepare_mlp_sync_batch_raw(
1969
1970
1971
            local_batch,
            dp_size=self.server_args.dp_size,
            attn_tp_size=self.attn_tp_size,
1972
            tp_group=self.tp_group,
1973
1974
1975
1976
            get_idle_batch=self.get_idle_batch,
            disable_cuda_graph=self.server_args.disable_cuda_graph,
            spec_algorithm=self.spec_algorithm,
            speculative_num_draft_tokens=self.server_args.speculative_num_draft_tokens,
1977
1978
1979
            enable_two_batch_overlap=self.server_args.enable_two_batch_overlap,
            enable_deepep_moe=self.server_args.enable_deepep_moe,
            deepep_mode=DeepEPMode[self.server_args.deepep_mode],
1980
            require_mlp_tp_gather=require_mlp_tp_gather(self.server_args),
1981
            disable_overlap_schedule=self.server_args.disable_overlap_schedule,
1982
1983
1984
        )

    @staticmethod
1985
    def prepare_mlp_sync_batch_raw(
1986
1987
1988
        local_batch: ScheduleBatch,
        dp_size,
        attn_tp_size: int,
1989
        tp_group,
1990
1991
1992
1993
        get_idle_batch,
        disable_cuda_graph: bool,
        spec_algorithm,
        speculative_num_draft_tokens,
1994
1995
1996
        enable_two_batch_overlap: bool,
        enable_deepep_moe: bool,
        deepep_mode: DeepEPMode,
1997
        require_mlp_tp_gather: bool,
1998
        disable_overlap_schedule: bool,
1999
    ):
2000
2001
2002
        # Check if other DP workers have running batches
        if local_batch is None:
            num_tokens = 0
2003
            num_tokens_for_logprob = 0
2004
2005
        elif local_batch.forward_mode.is_decode():
            num_tokens = local_batch.batch_size()
2006
            num_tokens_for_logprob = num_tokens
2007
2008
        else:
            num_tokens = local_batch.extend_num_tokens
2009
            num_tokens_for_logprob = sum(
Lianmin Zheng's avatar
Lianmin Zheng committed
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
                [
                    # We should have at least 1 token for sample in every case.
                    max(extend_len - logprob_start_len, 1)
                    for logprob_start_len, extend_len in zip(
                        local_batch.extend_logprob_start_lens, local_batch.extend_lens
                    )
                ]
            )

        if local_batch is None or local_batch.forward_mode.is_decode_or_idle():
            can_cuda_graph = 1
        else:
            can_cuda_graph = 0

        is_extend_in_batch = (
            local_batch.forward_mode.is_extend() if local_batch else False
        )
2027
2028

        tbo_preparer = TboDPAttentionPreparer()
2029
2030
2031
2032
2033
2034
        if disable_overlap_schedule:
            group = tp_group.device_group
            device = tp_group.device
        else:
            group = tp_group.cpu_group
            device = "cpu"
2035

Lianmin Zheng's avatar
Lianmin Zheng committed
2036
2037
2038
2039
        local_info = torch.tensor(
            [
                num_tokens,
                can_cuda_graph,
2040
                num_tokens_for_logprob,
Lianmin Zheng's avatar
Lianmin Zheng committed
2041
                is_extend_in_batch,
2042
2043
2044
2045
2046
2047
                *tbo_preparer.prepare_all_gather(
                    local_batch,
                    deepep_mode,
                    enable_deepep_moe,
                    enable_two_batch_overlap,
                ),
Lianmin Zheng's avatar
Lianmin Zheng committed
2048
2049
            ],
            dtype=torch.int64,
2050
            device=device,
Lianmin Zheng's avatar
Lianmin Zheng committed
2051
2052
        )
        global_info = torch.empty(
2053
            (dp_size, attn_tp_size, 6),
Lianmin Zheng's avatar
Lianmin Zheng committed
2054
            dtype=torch.int64,
2055
            device=device,
Lianmin Zheng's avatar
Lianmin Zheng committed
2056
        )
2057
        torch.distributed.all_gather_into_tensor(
Lianmin Zheng's avatar
Lianmin Zheng committed
2058
2059
            global_info.flatten(),
            local_info,
2060
            group=group,
2061
        )
Lianmin Zheng's avatar
Lianmin Zheng committed
2062
2063
2064
2065
        global_num_tokens = global_info[:, 0, 0].tolist()
        can_cuda_graph = min(global_info[:, 0, 1].tolist())
        global_num_tokens_for_logprob = global_info[:, 0, 2].tolist()
        is_extend_in_batch = global_info[:, 0, 3].tolist()
2066

2067
2068
2069
2070
        tbo_split_seq_index, global_forward_mode = tbo_preparer.compute_output(
            global_info[:, :, 4:6]
        )

Lianmin Zheng's avatar
Lianmin Zheng committed
2071
        if local_batch is None and max(global_num_tokens) > 0:
2072
            local_batch = get_idle_batch()
2073
2074

        if local_batch is not None:
2075
            # TODO: handle the case when moe_dense_tp_size != 1
2076
            if not require_mlp_tp_gather:
2077
2078
2079
2080
2081
2082
2083
                local_batch.global_num_tokens = [num_tokens]
                local_batch.global_num_tokens_for_logprob = [num_tokens_for_logprob]
            else:
                local_batch.global_num_tokens = global_num_tokens
                local_batch.global_num_tokens_for_logprob = (
                    global_num_tokens_for_logprob
                )
2084
            local_batch.is_extend_in_batch = any(is_extend_in_batch)
2085
2086
            local_batch.tbo_split_seq_index = tbo_split_seq_index
            local_batch.global_forward_mode = global_forward_mode
2087

2088
            # Check forward mode for cuda graph
2089
            if not disable_cuda_graph:
Lianmin Zheng's avatar
Lianmin Zheng committed
2090
                local_batch.can_run_dp_cuda_graph = can_cuda_graph
2091

2092
        return local_batch
2093
2094
2095
2096
2097

    def get_idle_batch(self):
        idle_batch = ScheduleBatch.init_new(
            [],
            self.req_to_token_pool,
2098
            self.token_to_kv_pool_allocator,
2099
2100
2101
            self.tree_cache,
            self.model_config,
            self.enable_overlap,
2102
            self.spec_algorithm,
2103
            self.server_args.enable_custom_logit_processor,
2104
2105
2106
2107
        )
        idle_batch.prepare_for_idle()
        return idle_batch

2108
2109
    def move_ready_grammar_requests(self):
        """Move requests whose grammar objects are ready from grammar_queue to waiting_queue."""
2110

2111
        num_ready_reqs = 0
2112
        num_timeout_reqs = 0
2113
2114
        for req in self.grammar_queue:
            try:
2115
2116
2117
                if req.finished():  # It is aborted by AbortReq
                    num_ready_reqs += 1
                    continue
2118
                req.grammar = req.grammar.result(timeout=0.03)
2119
2120
2121
2122
2123
                self.grammar_backend.set_cache(req.grammar_key, req.grammar.copy())
                if req.grammar is INVALID_GRAMMAR_OBJ:
                    req.set_finish_with_abort(
                        f"Invalid grammar request: {req.grammar_key=}"
                    )
2124
2125
                num_ready_reqs += 1
            except futures._base.TimeoutError:
2126
                req.grammar_wait_ct += 1
2127
2128
                # NOTE(lianmin): this timeout is the waiting time of the above line. It is
                # not the waiting time from it enters the grammar queue.
2129
                if req.grammar_wait_ct > GRAMMAR_TIMEOUT / 0.03:
2130
                    num_timeout_reqs = 1
2131
2132
                break

2133
        if self.server_args.enable_dp_attention:
2134
2135
            tp_size = self.attn_tp_size
            tp_group = self.attn_tp_cpu_group
2136
        else:
2137
2138
2139
2140
2141
            tp_size = self.tp_size
            tp_group = self.tp_cpu_group

        if tp_size > 1:
            # Sync across TP ranks to make sure they have the same number of ready requests
2142
            tensor = torch.tensor([num_ready_reqs, num_timeout_reqs], dtype=torch.int32)
2143
2144
2145
            torch.distributed.all_reduce(
                tensor, op=torch.distributed.ReduceOp.MAX, group=tp_group
            )
2146
            num_ready_reqs_max, num_timeout_reqs_max = tensor.tolist()
2147

2148
            for i in range(num_ready_reqs, num_ready_reqs_max):
2149
                req = self.grammar_queue[i]
2150
2151
                if req.finished():  # It is aborted by AbortReq
                    continue
2152
                req.grammar = req.grammar.result()
2153
2154
2155
2156
2157
2158
2159
2160
                self.grammar_backend.set_cache(req.grammar_key, req.grammar.copy())
                if req.grammar is INVALID_GRAMMAR_OBJ:
                    req.set_finish_with_abort(
                        f"Invalid grammar request: {req.grammar_key=}"
                    )
        else:
            num_ready_reqs_max = num_ready_reqs
            num_timeout_reqs_max = num_timeout_reqs
2161

2162
2163
2164
2165
2166
2167
2168
        for i in range(num_ready_reqs, num_ready_reqs + num_timeout_reqs_max):
            req = self.grammar_queue[i]
            req.grammar.cancel()
            error_msg = f"Grammar preprocessing timed out for {req.grammar_key=}"
            req.set_finish_with_abort(error_msg)
            self.grammar_backend.set_cache(req.grammar_key, INVALID_GRAMMAR_OBJ)
        num_ready_reqs = num_ready_reqs_max + num_timeout_reqs_max
2169

2170
        self._extend_requests_to_queue(self.grammar_queue[:num_ready_reqs])
2171
2172
        self.grammar_queue = self.grammar_queue[num_ready_reqs:]

2173
2174
2175
2176
2177
2178
2179
    def set_next_batch_sampling_info_done(self, batch: ScheduleBatch):
        if batch.next_batch_sampling_info:
            if batch.next_batch_sampling_info.grammars is not None:
                batch.next_batch_sampling_info.update_regex_vocab_mask()
                self.current_stream.synchronize()
            batch.next_batch_sampling_info.sampling_info_done.set()

2180
2181
2182
    def watchdog_thread(self):
        """A watch dog thread that will try to kill the server itself if one forward batch takes too long."""
        self.watchdog_last_forward_ct = 0
2183
        self.watchdog_last_time = time.perf_counter()
2184
2185

        while True:
2186
            current = time.perf_counter()
2187
2188
2189
2190
2191
2192
2193
2194
2195
            if self.cur_batch is not None:
                if self.watchdog_last_forward_ct == self.forward_ct:
                    if current > self.watchdog_last_time + self.watchdog_timeout:
                        break
                else:
                    self.watchdog_last_forward_ct = self.forward_ct
                    self.watchdog_last_time = current
            time.sleep(self.watchdog_timeout // 2)

Lianmin Zheng's avatar
Lianmin Zheng committed
2196
2197
        if not disable_request_logging():
            # Print batch size and memory pool info to check whether there are de-sync issues.
Hanming Lu's avatar
Hanming Lu committed
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
            if self.is_hybrid:
                (
                    _,
                    _,
                    _,
                    _,
                    full_available_size,
                    full_evictable_size,
                    swa_available_size,
                    swa_evictable_size,
                ) = self._get_swa_token_info()
                info_msg = (
                    f"{full_available_size=}, "
                    f"{full_evictable_size=}, "
                    f"{swa_available_size=}, "
                    f"{swa_evictable_size=}, "
                )
            else:
                _, _, available_size, evictable_size = self._get_token_info()
                info_msg = f"{available_size=}, " f"{evictable_size=}, "
Lianmin Zheng's avatar
Lianmin Zheng committed
2218
2219
2220
            logger.error(
                f"{self.cur_batch.batch_size()=}, "
                f"{self.cur_batch.reqs=}, "
Hanming Lu's avatar
Hanming Lu committed
2221
                f"{info_msg}"
Lianmin Zheng's avatar
Lianmin Zheng committed
2222
2223
            )

2224
        pyspy_dump_schedulers()
Lianmin Zheng's avatar
Lianmin Zheng committed
2225
        logger.error(f"Watchdog timeout ({self.watchdog_timeout=})")
2226
2227
        print(file=sys.stderr, flush=True)
        print(file=sys.stdout, flush=True)
Lianmin Zheng's avatar
Lianmin Zheng committed
2228
2229

        # Wait for some time so that the parent process can print the error.
2230
2231
2232
        time.sleep(5)
        self.parent_process.send_signal(signal.SIGQUIT)

2233
2234
2235
    def flush_cache_wrapped(self, recv_req: FlushCacheReqInput):
        success = self.flush_cache()
        return FlushCacheReqOutput(success=success)
2236

2237
    def flush_cache(self):
2238
        """Flush the memory pool and cache."""
2239
2240
2241
2242
2243
        if (
            len(self.waiting_queue) == 0
            and self.running_batch.is_empty()
            and (self.pp_size == 1 or all(x.is_empty() for x in self.running_mbs))
        ):
2244
2245
            self.cur_batch = None
            self.last_batch = None
2246
            self.tree_cache.reset()
2247
            if self.grammar_backend:
Lianmin Zheng's avatar
Lianmin Zheng committed
2248
                self.grammar_backend.reset()
2249
            self.req_to_token_pool.clear()
2250
            self.token_to_kv_pool_allocator.clear()
2251
2252
2253

            if not self.spec_algorithm.is_none():
                self.draft_worker.model_runner.req_to_token_pool.clear()
2254
                self.draft_worker.model_runner.token_to_kv_pool_allocator.clear()
2255
2256
2257
2258
2259

            self.num_generated_tokens = 0
            self.forward_ct_decode = 0
            self.spec_num_total_accepted_tokens = 0
            self.spec_num_total_forward_ct = 0
2260
2261
            self.cum_spec_accept_length = 0
            self.cum_spec_accept_count = 0
2262
2263
2264
2265
2266
2267
2268
            torch.cuda.empty_cache()
            logger.info("Cache flushed successfully!")
            if_success = True
        else:
            logging.warning(
                f"Cache not flushed because there are pending requests. "
                f"#queue-req: {len(self.waiting_queue)}, "
Lianmin Zheng's avatar
Lianmin Zheng committed
2269
                f"#running-req: {len(self.running_batch.reqs)}"
2270
2271
2272
2273
            )
            if_success = False
        return if_success

Liangsheng Yin's avatar
Liangsheng Yin committed
2274
2275
    def get_load(self):
        # TODO(lsyin): use dynamically maintained num_waiting_tokens
Hanming Lu's avatar
Hanming Lu committed
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
        if self.is_hybrid:
            load_full = (
                self.full_tokens_per_layer
                - self.token_to_kv_pool_allocator.full_available_size()
                - self.tree_cache.full_evictable_size()
            )
            load_swa = (
                self.swa_tokens_per_layer
                - self.token_to_kv_pool_allocator.swa_available_size()
                - self.tree_cache.swa_evictable_size()
            )
            load = max(load_full, load_swa)
        else:
            load = (
                self.max_total_num_tokens
                - self.token_to_kv_pool_allocator.available_size()
                - self.tree_cache.evictable_size()
            )
Liangsheng Yin's avatar
Liangsheng Yin committed
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
        load += sum(len(req.origin_input_ids) for req in self.waiting_queue)
        if self.disaggregation_mode == DisaggregationMode.PREFILL:
            load += sum(
                len(req.origin_input_ids)
                for req in self.disagg_prefill_bootstrap_queue.queue
            )
        elif self.disaggregation_mode == DisaggregationMode.DECODE:
            load += sum(
                len(req.req.origin_input_ids)
                for req in self.disagg_decode_prealloc_queue.queue
            )

        return load

2308
2309
2310
    def get_internal_state(self, recv_req: GetInternalStateReq):
        ret = dict(global_server_args_dict)
        ret["last_gen_throughput"] = self.last_gen_throughput
2311
2312
2313
2314
2315
2316
2317
2318
2319
        ret["memory_usage"] = {
            "weight": round(
                self.tp_worker.worker.model_runner.weight_load_mem_usage, 2
            ),
            "kvcache": round(
                self.token_to_kv_pool_allocator.get_kvcache().mem_usage, 2
            ),
            "token_capacity": int(self.max_total_num_tokens),
        }
2320
2321
2322
2323
2324
2325

        if not _is_cpu:
            ret["memory_usage"]["cuda_graph"] = round(
                self.tp_worker.worker.model_runner.cuda_graph_mem_usage, 2
            )

2326
2327
2328
2329
2330
2331
        if not self.spec_algorithm.is_none() and self.cum_spec_accept_count > 0:
            ret["avg_spec_accept_length"] = (
                self.cum_spec_accept_length / self.cum_spec_accept_count
            )
        if RECORD_STEP_TIME:
            ret["step_time_dict"] = self.step_time_dict
Liangsheng Yin's avatar
Liangsheng Yin committed
2332
2333
2334
2335

        ret["load"] = self.get_load()

        return GetInternalStateReqOutput(internal_state=ret)
2336
2337
2338
2339
2340

    def set_internal_state(self, recv_req: SetInternalStateReq):
        server_args_dict = recv_req.server_args
        args_allow_update = set(
            [
2341
                "max_micro_batch_size",
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
                "speculative_accept_threshold_single",
                "speculative_accept_threshold_acc",
            ]
        )
        if_success = True
        for k, v in server_args_dict.items():
            if k not in args_allow_update:
                logging.warning(f"Updating {k} is not supported.")
                if_success = False
                break
2352
2353
2354
2355
2356
2357
2358
2359
            elif k == "max_micro_batch_size" and (
                v > self.max_running_requests // self.pp_size or v < 1
            ):
                logging.warning(
                    f"Updating {k} to {v} is rejected because it is out of the valid range [1, {self.max_running_requests // self.pp_size}]."
                )
                if_success = False
                break
2360
2361
2362
2363
2364
2365
2366
2367
2368
        if if_success:
            if not self.spec_algorithm.is_none() and self.cum_spec_accept_count > 0:
                avg_spec_accept_length = (
                    self.cum_spec_accept_length / self.cum_spec_accept_count
                )
                logger.info(f"{avg_spec_accept_length=}")
            self.cum_spec_accept_length = self.cum_spec_accept_count = 0
            for k, v in server_args_dict.items():
                global_server_args_dict[k] = v
2369
            logger.info(f"Global server args updated! {global_server_args_dict=}")
2370
2371
2372
2373
2374
        return SetInternalStateReqOutput(
            updated=True,
            server_args=global_server_args_dict,
        )

2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
    def handle_rpc_request(self, recv_req: RpcReqInput):
        # Handle RPC requests
        logger.info(
            f"handle_rpc_request: {recv_req.method}, param: {recv_req.parameters}"
        )

        success = True
        exec = None
        try:
            func = getattr(self, recv_req.method)
            func(recv_req.parameters)
        except Exception as e:
            success = False
            exec = e
            logger.error(f"Failed to call rpc {recv_req.method}: {str(e)}")

        barrier()
        return RpcReqOutput(success, "" if not exec else str(exec))

    def save_remote_model(self, params):
        url = params["url"]

2397
        worker = self.tp_worker.worker
2398
2399
2400
2401

        worker.model_runner.save_remote_model(url)

    def save_sharded_model(self, params):
2402
        worker = self.tp_worker.worker
2403
2404
2405
2406
2407
2408
2409

        worker.model_runner.save_sharded_model(
            path=params["path"],
            pattern=params["pattern"],
            max_size=params["max_size"],
        )

2410
2411
    def abort_request(self, recv_req: AbortReq):
        # Delete requests in the waiting queue
Lianmin Zheng's avatar
Lianmin Zheng committed
2412
        to_del = []
2413
        for i, req in enumerate(self.waiting_queue):
2414
            if recv_req.abort_all or req.rid.startswith(recv_req.rid):
Lianmin Zheng's avatar
Lianmin Zheng committed
2415
                to_del.append(i)
2416

Lianmin Zheng's avatar
Lianmin Zheng committed
2417
        # Sort in reverse order to avoid index issues when deleting
Lianmin Zheng's avatar
Lianmin Zheng committed
2418
        for i in reversed(to_del):
2419
2420
2421
            # Abort method 1: directly pop from the queue
            # This only works for requests that have not started anything.
            # We still need to send something back to TokenizerManager to clean up the state.
Lianmin Zheng's avatar
Lianmin Zheng committed
2422
            req = self.waiting_queue.pop(i)
Lianmin Zheng's avatar
Lianmin Zheng committed
2423
            self.send_to_tokenizer.send_pyobj(AbortReq(req.rid))
2424
            logger.debug(f"Abort queued request. {req.rid=}")
2425

2426
2427
2428
2429
2430
        # Delete the requests in the grammar queue
        for req in self.grammar_queue:
            # Abort method 2: call `set_finish_with_abort`
            # The request will still run one prefill forward pass.
            # In this case, we change the input_ids to be only one token to make this prefill cheap.
2431
            if recv_req.abort_all or req.rid.startswith(recv_req.rid):
2432
                logger.debug(f"Abort grammar queue request. {req.rid=}")
2433
2434
                if req.grammar:
                    req.grammar.cancel()
2435
2436
                req.set_finish_with_abort("Aborted by AbortReq.")

2437
        # Delete requests in the running batch
Lianmin Zheng's avatar
Lianmin Zheng committed
2438
2439
2440
2441
2442
2443
        if self.cur_batch is self.running_batch or self.cur_batch is None:
            reqs = self.running_batch.reqs
        else:
            reqs = self.running_batch.reqs + self.cur_batch.reqs

        for req in reqs:
2444
2445
2446
            if not req.finished() and (
                recv_req.abort_all or req.rid.startswith(recv_req.rid)
            ):
2447
2448
2449
                # Abort method 3: set `to_abort=True`
                # The request will still run one decode forward pass.
                # Then we reuse all existing code to clean up the KV cache allocation.
Lianmin Zheng's avatar
Lianmin Zheng committed
2450
2451
                logger.debug(f"Abort running request. {req.rid=}")
                req.to_abort = True
2452

2453
2454
2455
    def _pause_engine(self) -> Tuple[List[Req], int]:
        raise NotImplementedError()

Chayenne's avatar
Chayenne committed
2456
2457
2458
    def update_weights_from_disk(self, recv_req: UpdateWeightFromDiskReqInput):
        """In-place update of the weights from disk."""
        success, message = self.tp_worker.update_weights_from_disk(recv_req)
2459
        if success:
Stefan He's avatar
Stefan He committed
2460
2461
            flush_cache_success = self.flush_cache()
            assert flush_cache_success, "Cache flush failed after updating weights"
2462
2463
        else:
            logger.error(message)
2464
        return UpdateWeightFromDiskReqOutput(success, message, 0)
2465

2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
    def load_lora_adapter(
        self, recv_req: LoadLoRAAdapterReqInput
    ) -> LoadLoRAAdapterReqOutput:
        """In-place loading a new lora adapter from disk or huggingface."""

        result = self.tp_worker.load_lora_adapter(recv_req)
        return result

    def unload_lora_adapter(
        self, recv_req: UnloadLoRAAdapterReqInput
    ) -> UnloadLoRAAdapterReqOutput:
        """Unload the lora adapter."""

        result = self.tp_worker.unload_lora_adapter(recv_req)
        return result

2482
2483
2484
    def init_weights_update_group(self, recv_req: InitWeightsUpdateGroupReqInput):
        """Initialize the online model parameter update group."""
        success, message = self.tp_worker.init_weights_update_group(recv_req)
2485
        return InitWeightsUpdateGroupReqOutput(success, message)
2486
2487

    def update_weights_from_distributed(
2488
2489
2490
        self,
        recv_req: UpdateWeightsFromDistributedReqInput,
    ) -> Tuple[bool, str]:
2491
2492
2493
        """Update the online model parameter."""
        success, message = self.tp_worker.update_weights_from_distributed(recv_req)
        if success:
2494
2495
2496
            if recv_req.flush_cache:
                flush_cache_success = self.flush_cache()
                assert flush_cache_success, "Cache flush failed after updating weights"
2497
2498
        else:
            logger.error(message)
2499
        return UpdateWeightsFromDistributedReqOutput(success, message)
2500

2501
2502
2503
2504
2505
    def update_weights_from_tensor(self, recv_req: UpdateWeightsFromTensorReqInput):
        """Update the online model parameter from tensors."""
        success, message = self.tp_worker.update_weights_from_tensor(recv_req)
        # TODO extract common code b/t update_weights_from_distributed and update_weights_from_tensor later
        if success:
2506
            if recv_req.flush_cache:
Stefan He's avatar
Stefan He committed
2507
2508
                flush_cache_success = self.flush_cache()
                assert flush_cache_success, "Cache flush failed after updating weights"
2509
2510
        else:
            logger.error(message)
2511
        barrier(group=self.tp_cpu_group)
2512
        return UpdateWeightsFromTensorReqOutput(success, message)
2513

2514
2515
    def get_weights_by_name(self, recv_req: GetWeightsByNameReqInput):
        parameter = self.tp_worker.get_weights_by_name(recv_req)
2516
        return GetWeightsByNameReqOutput(parameter)
2517

2518
    def release_memory_occupation(self, recv_req: ReleaseMemoryOccupationReqInput):
2519
2520
        tags = recv_req.tags

2521
        if tags is None or len(tags) == 0:
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
            tags = [GPU_MEMORY_TYPE_WEIGHTS, GPU_MEMORY_TYPE_KV_CACHE]

        if GPU_MEMORY_TYPE_KV_CACHE in tags:
            self.memory_saver_adapter.pause(GPU_MEMORY_TYPE_KV_CACHE)
            self.flush_cache()

        if GPU_MEMORY_TYPE_WEIGHTS in tags:
            self.stashed_model_static_state = _export_static_state(
                self.tp_worker.worker.model_runner.model
            )
2532
            torch.distributed.barrier(self.tp_cpu_group)
2533
2534
            self.memory_saver_adapter.pause(GPU_MEMORY_TYPE_WEIGHTS)

2535
        return ReleaseMemoryOccupationReqOutput()
2536

2537
    def resume_memory_occupation(self, recv_req: ResumeMemoryOccupationReqInput):
2538
        tags = recv_req.tags
2539

2540
2541
2542
2543
2544
        if tags is None or len(tags) == 0:
            tags = [GPU_MEMORY_TYPE_WEIGHTS, GPU_MEMORY_TYPE_KV_CACHE]

        if GPU_MEMORY_TYPE_WEIGHTS in tags:
            self.memory_saver_adapter.resume(GPU_MEMORY_TYPE_WEIGHTS)
2545
            torch.distributed.barrier(self.tp_cpu_group)
2546
2547
2548
2549
2550
2551
2552
2553
2554
            _import_static_state(
                self.tp_worker.worker.model_runner.model,
                self.stashed_model_static_state,
            )
            del self.stashed_model_static_state

        if GPU_MEMORY_TYPE_KV_CACHE in tags:
            self.memory_saver_adapter.resume(GPU_MEMORY_TYPE_KV_CACHE)

2555
2556
        return ResumeMemoryOccupationReqOutput()

2557
2558
2559
2560
2561
2562
2563
    def slow_down(self, recv_req: SlowDownReqInput):
        t = recv_req.forward_sleep_time
        if t is not None and t <= 0:
            t = None
        self.forward_sleep_time = t
        return SlowDownReqOutput()

2564
    def profile(self, recv_req: ProfileReq):
2565
        if recv_req.type == ProfileReqType.START_PROFILE:
2566
            if recv_req.profile_by_stage or recv_req.start_step:
2567
2568
                return self.init_profile(
                    recv_req.output_dir,
2569
                    recv_req.start_step,
2570
2571
2572
2573
2574
                    recv_req.num_steps,
                    recv_req.activities,
                    recv_req.with_stack,
                    recv_req.record_shapes,
                    recv_req.profile_by_stage,
2575
                    recv_req.profile_id,
2576
2577
2578
2579
                )
            else:
                self.init_profile(
                    recv_req.output_dir,
2580
                    recv_req.start_step,
2581
2582
2583
2584
2585
                    recv_req.num_steps,
                    recv_req.activities,
                    recv_req.with_stack,
                    recv_req.record_shapes,
                    recv_req.profile_by_stage,
2586
                    recv_req.profile_id,
2587
2588
                )
                return self.start_profile(True)
2589
        else:
2590
2591
            return self.stop_profile()

2592
    def init_profile(
2593
2594
        self,
        output_dir: Optional[str],
2595
        start_step: Optional[int],
2596
2597
        num_steps: Optional[int],
        activities: Optional[List[str]],
2598
2599
        with_stack: Optional[bool],
        record_shapes: Optional[bool],
2600
        profile_by_stage: bool,
2601
        profile_id: str,
2602
2603
    ) -> ProfileReqOutput:
        if self.profile_in_progress:
2604
2605
2606
2607
2608
            return ProfileReqOutput(
                success=False,
                message="Profiling is already in progress. Call /stop_profile first.",
            )

2609
2610
        self.profile_by_stage = profile_by_stage

2611
2612
2613
2614
2615
2616
        if output_dir is None:
            output_dir = os.getenv("SGLANG_TORCH_PROFILER_DIR", "/tmp")
        if activities is None:
            activities = ["CPU", "GPU"]

        self.torch_profiler_output_dir = output_dir
2617
2618
        self.torch_profiler_with_stack = with_stack
        self.torch_profiler_record_shapes = record_shapes
2619
        self.profiler_activities = activities
2620
        self.profile_id = profile_id
2621

2622
2623
2624
        if start_step:
            self.profiler_start_forward_ct = max(start_step, self.forward_ct + 1)

2625
2626
2627
2628
2629
2630
2631
        if num_steps:
            self.profile_steps = num_steps
            if self.profile_by_stage:
                self.profiler_target_prefill_ct = num_steps
                self.profiler_target_decode_ct = num_steps
                self.profiler_prefill_ct = 0
                self.profiler_decode_ct = 0
2632
2633
2634
2635
            elif start_step:
                self.profiler_target_forward_ct = (
                    self.profiler_start_forward_ct + num_steps
                )
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
            else:
                self.profiler_target_forward_ct = self.forward_ct + num_steps
            # The caller will be notified when reaching profiler_target_forward_ct
        else:
            self.profiler_target_forward_ct = None

        return ProfileReqOutput(success=True, message="Succeeded")

    def start_profile(
        self, stage: Optional[ForwardMode] = None
    ) -> ProfileReqOutput | None:
        stage_str = f" for {stage.__str__()}" if stage else ""
2648
        logger.info(
2649
            f"Profiling starts{stage_str}. Traces will be saved to: {self.torch_profiler_output_dir} (with profile id: {self.profile_id})",
2650
2651
        )

2652
2653
2654
2655
        activities = self.profiler_activities
        with_stack = self.torch_profiler_with_stack
        record_shapes = self.torch_profiler_record_shapes

2656
2657
2658
2659
2660
2661
2662
2663
        activity_map = {
            "CPU": torch.profiler.ProfilerActivity.CPU,
            "GPU": torch.profiler.ProfilerActivity.CUDA,
        }
        torchprof_activities = [
            activity_map[a] for a in activities if a in activity_map
        ]

2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
        if "RPD" in activities:
            from rpdTracerControl import rpdTracerControl

            rpdTracerControl.skipCreate()

            self.rpd_profile_path = os.path.join(
                self.torch_profiler_output_dir,
                "rpd-" + str(time.time()) + f"-TP-{self.tp_rank}" + ".trace.json.gz",
            )

            if self.tp_rank == 0:
                import sqlite3

                from rocpd.schema import RocpdSchema

                if os.path.exists("trace.rpd"):
                    os.unlink("trace.rpd")
                schema = RocpdSchema()
                connection = sqlite3.connect("trace.rpd")
                schema.writeSchema(connection)
                connection.commit()
                del connection
            torch.distributed.barrier(self.tp_cpu_group)

            self.rpd_profiler = rpdTracerControl()
            self.rpd_profiler.setPythonTrace(True)
            self.rpd_profiler.start()
            self.rpd_profiler.rangePush("", "rpd profile range", "")
            self.profile_in_progress = True
        elif torchprof_activities:
2694
2695
            self.torch_profiler = torch.profiler.profile(
                activities=torchprof_activities,
2696
2697
                with_stack=with_stack if with_stack is not None else True,
                record_shapes=record_shapes if record_shapes is not None else False,
2698
2699
            )
            self.torch_profiler.start()
2700
            self.profile_in_progress = True
2701
2702
2703

        if "MEM" in activities:
            torch.cuda.memory._record_memory_history(max_entries=100000)
2704
            self.profile_in_progress = True
2705

2706
2707
        if "CUDA_PROFILER" in activities:
            torch.cuda.cudart().cudaProfilerStart()
2708
            self.profile_in_progress = True
2709

2710
        return ProfileReqOutput(success=True, message="Succeeded")
2711

2712
2713
2714
2715
    def stop_profile(
        self, stage: Optional[ForwardMode] = None
    ) -> ProfileReqOutput | None:
        if not self.profile_in_progress:
2716
2717
2718
2719
            return ProfileReqOutput(
                success=False,
                message="Profiling is not in progress. Call /start_profile first.",
            )
2720

2721
2722
2723
        if not Path(self.torch_profiler_output_dir).exists():
            Path(self.torch_profiler_output_dir).mkdir(parents=True, exist_ok=True)

2724
2725
        stage_suffix = f"-{stage.__str__()}" if stage else ""
        logger.info("Stop profiling" + stage_suffix + "...")
2726
2727
2728
2729
2730
        if self.torch_profiler is not None:
            self.torch_profiler.stop()
            self.torch_profiler.export_chrome_trace(
                os.path.join(
                    self.torch_profiler_output_dir,
2731
                    self.profile_id
2732
2733
2734
                    + f"-TP-{self.tp_rank}"
                    + stage_suffix
                    + ".trace.json.gz",
2735
2736
                )
            )
2737
2738
2739
2740
2741
2742
            torch.distributed.barrier(self.tp_cpu_group)

        if self.rpd_profiler is not None:
            self.rpd_profiler.rangePop()
            self.rpd_profiler.stop()
            self.rpd_profiler.flush()
2743

2744
2745
2746
2747
2748
2749
2750
2751
2752
            torch.distributed.barrier(self.tp_cpu_group)
            if self.tp_rank == 0:
                from sglang.srt.utils import rpd_to_chrome_trace

                rpd_to_chrome_trace("trace.rpd", self.rpd_profile_path)
            self.rpd_profiler = None
            self.rpd_profiler_path = None

        if self.profiler_activities is not None and "MEM" in self.profiler_activities:
2753
            memory_profile_path = os.path.join(
2754
                self.torch_profiler_output_dir,
2755
2756
2757
2758
                str(time.time())
                + f"-TP-{self.tp_rank}-memory"
                + stage_suffix
                + ".pickle",
2759
2760
2761
2762
            )
            torch.cuda.memory._dump_snapshot(memory_profile_path)
            torch.cuda.memory._record_memory_history(enabled=None)

2763
2764
2765
        if "CUDA_PROFILER" in self.profiler_activities:
            torch.cuda.cudart().cudaProfilerStop()

2766
2767
2768
        logger.info(
            "Profiling done. Traces are saved to: %s",
            self.torch_profiler_output_dir,
2769
        )
2770
        self.torch_profiler = None
2771
        self.profile_in_progress = False
2772
        self.profiler_start_forward_ct = None
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794

        return ProfileReqOutput(success=True, message="Succeeded.")

    def _profile_batch_predicate(self, batch):
        if self.profile_by_stage:
            if batch.forward_mode.is_prefill():
                if self.profiler_prefill_ct == 0:
                    self.start_profile(batch.forward_mode)
                self.profiler_prefill_ct += 1
                if self.profiler_prefill_ct > self.profiler_target_prefill_ct:
                    if self.profile_in_progress:
                        self.stop_profile(stage=ForwardMode.EXTEND)
            elif batch.forward_mode.is_decode():
                if self.profiler_decode_ct == 0:
                    if self.profile_in_progress:
                        # force trace flush
                        self.stop_profile(ForwardMode.EXTEND)
                    self.start_profile(batch.forward_mode)
                self.profiler_decode_ct += 1
                if self.profiler_decode_ct > self.profiler_target_decode_ct:
                    if self.profile_in_progress:
                        self.stop_profile(stage=ForwardMode.DECODE)
2795
2796
            elif batch.forward_mode.is_idle():
                pass
2797
            else:
2798
                raise RuntimeError(f"unsupported profile stage: {batch.forward_mode}")
2799
2800
2801
2802
2803
2804
2805
        else:
            # Check profiler
            if (
                self.profiler_target_forward_ct
                and self.profiler_target_forward_ct <= self.forward_ct
            ):
                self.stop_profile()
2806
2807
2808
2809
2810
            if (
                self.profiler_start_forward_ct
                and self.profiler_start_forward_ct == self.forward_ct
            ):
                self.start_profile()
2811

2812
2813
    def expert_distribution_handle(self, recv_req: ExpertDistributionReq):
        if recv_req == ExpertDistributionReq.START_RECORD:
2814
            get_global_expert_distribution_recorder().start_record()
2815
        elif recv_req == ExpertDistributionReq.STOP_RECORD:
2816
            get_global_expert_distribution_recorder().stop_record()
2817
        elif recv_req == ExpertDistributionReq.DUMP_RECORD:
2818
            get_global_expert_distribution_recorder().dump_record()
2819
2820
        else:
            raise ValueError("Unrecognized ExpertDistributionReq value")
2821
        return ExpertDistributionReqOutput()
2822

2823
    def open_session(self, recv_req: OpenSessionReqInput):
2824
2825
2826
2827
        # handle error
        session_id = recv_req.session_id
        if session_id in self.sessions:
            logger.warning(f"session id {session_id} already exist, cannot open.")
2828
            return OpenSessionReqOutput(session_id, False)
2829
        elif session_id is None:
2830
            logger.warning("session id is None, cannot open.")
2831
            return OpenSessionReqOutput(session_id, False)
2832
2833
2834
2835
        else:
            self.sessions[session_id] = Session(
                recv_req.capacity_of_str_len, session_id
            )
2836
            return OpenSessionReqOutput(session_id, True)
2837
2838
2839
2840
2841
2842
2843
2844
2845

    def close_session(self, recv_req: CloseSessionReqInput):
        # handle error
        session_id = recv_req.session_id
        if session_id not in self.sessions:
            logger.warning(f"session id {session_id} does not exist, cannot delete.")
        else:
            del self.sessions[session_id]

2846
2847
    def get_print_prefix(self):
        prefix = ""
2848
2849
        if self.attn_dp_rank is not None:
            prefix += f" DP{self.attn_dp_rank}"
2850
2851
2852
2853
2854
2855
        if self.server_args.tp_size > 1:
            prefix += f" TP{self.tp_rank}"
        if self.pp_size > 1:
            prefix += f" PP{self.pp_rank}"
        return prefix

2856
2857
2858
2859
2860
2861
2862
    def _publish_kv_events(self):
        if self.enable_kv_cache_events:
            events = self.tree_cache.take_events()
            if events:
                batch = KVEventBatch(ts=time.time(), events=events)
                self.kv_event_publisher.publish(batch)

2863

2864
2865
2866
2867
def is_health_check_generate_req(recv_req):
    return getattr(recv_req, "rid", "").startswith("HEALTH_CHECK")


2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
def _export_static_state(model):
    return dict(
        buffers=[
            (name, buffer.detach().clone()) for name, buffer in model.named_buffers()
        ]
    )


def _import_static_state(model, static_params):
    self_named_buffers = dict(model.named_buffers())
    for name, tensor in static_params["buffers"]:
        self_named_buffers[name][...] = tensor


2882
2883
2884
2885
2886
def run_scheduler_process(
    server_args: ServerArgs,
    port_args: PortArgs,
    gpu_id: int,
    tp_rank: int,
2887
    pp_rank: int,
2888
    dp_rank: Optional[int],
2889
    pipe_writer,
2890
):
2891
    # Generate the prefix
2892
2893
2894
2895
2896
2897
2898
    prefix = ""
    if dp_rank is not None:
        prefix += f" DP{dp_rank}"
    if server_args.tp_size > 1:
        prefix += f" TP{tp_rank}"
    if server_args.pp_size > 1:
        prefix += f" PP{pp_rank}"
2899

2900
    # Config the process
2901
    setproctitle.setproctitle(f"sglang::scheduler{prefix.replace(' ', '_')}")
2902
    faulthandler.enable()
2903
    kill_itself_when_parent_died()
2904
    parent_process = psutil.Process().parent()
2905

2906
2907
2908
    # [For Router] if env var "SGLANG_DP_RANK" exist, set dp_rank to the value of the env var
    if dp_rank is None and "SGLANG_DP_RANK" in os.environ:
        dp_rank = int(os.environ["SGLANG_DP_RANK"])
2909

Wang Ran (汪然)'s avatar
Wang Ran (汪然) committed
2910
    # Configure the logger
2911
    configure_logger(server_args, prefix=prefix)
2912
    suppress_other_loggers()
2913

2914
    # Set cpu affinity to this gpu process
2915
2916
2917
    if get_bool_env_var("SGLANG_SET_CPU_AFFINITY"):
        set_gpu_proc_affinity(server_args.tp_size, server_args.nnodes, gpu_id)

2918
    # Create a scheduler and run the event loop
2919
    try:
2920
        scheduler = Scheduler(server_args, port_args, gpu_id, tp_rank, pp_rank, dp_rank)
2921
        pipe_writer.send(
Mick's avatar
Mick committed
2922
2923
2924
2925
2926
            {
                "status": "ready",
                "max_total_num_tokens": scheduler.max_total_num_tokens,
                "max_req_input_len": scheduler.max_req_input_len,
            }
2927
        )
Byron Hsu's avatar
Byron Hsu committed
2928

2929
        disaggregation_mode: DisaggregationMode = scheduler.disaggregation_mode
Byron Hsu's avatar
Byron Hsu committed
2930
        if disaggregation_mode == DisaggregationMode.NULL:
2931
2932
2933
            if server_args.pp_size > 1:
                scheduler.event_loop_pp()
            elif scheduler.enable_overlap:
Byron Hsu's avatar
Byron Hsu committed
2934
2935
2936
2937
                scheduler.event_loop_overlap()
            else:
                scheduler.event_loop_normal()
        elif disaggregation_mode == DisaggregationMode.PREFILL:
2938
2939
2940
2941
            if scheduler.enable_overlap:
                scheduler.event_loop_overlap_disagg_prefill()
            else:
                scheduler.event_loop_normal_disagg_prefill()
2942

Byron Hsu's avatar
Byron Hsu committed
2943
        elif disaggregation_mode == DisaggregationMode.DECODE:
2944
2945
2946
2947
            if scheduler.enable_overlap:
                scheduler.event_loop_overlap_disagg_decode()
            else:
                scheduler.event_loop_normal_disagg_decode()
Byron Hsu's avatar
Byron Hsu committed
2948

2949
    except Exception:
2950
2951
2952
        traceback = get_exception_traceback()
        logger.error(f"Scheduler hit an exception: {traceback}")
        parent_process.send_signal(signal.SIGQUIT)