| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from collections.abc import Generator |
| 4 | from contextlib import contextmanager |
| 5 | from typing import TYPE_CHECKING |
| 6 | |
| 7 | import torch |
| 8 | |
| 9 | from vllm.config import VllmConfig |
| 10 | from vllm.distributed.ec_transfer import get_ec_transfer, has_ec_transfer |
| 11 | from vllm.distributed.ec_transfer.ec_connector.base import ECConnectorBase |
| 12 | from vllm.v1.outputs import ( |
| 13 | EMPTY_MODEL_RUNNER_OUTPUT, |
| 14 | ECConnectorOutput, |
| 15 | ModelRunnerOutput, |
| 16 | ) |
| 17 | |
| 18 | if TYPE_CHECKING: |
| 19 | from vllm.v1.core.sched.output import SchedulerOutput |
| 20 | from vllm.v1.worker.gpu.mm.encoder_cache import EncoderCache |
| 21 | |
| 22 | |
| 23 | class ECConnector: |
| 24 | """EC connector interface used by the V2 GPU model runner.""" |
| 25 | |
| 26 | @contextmanager |
| 27 | def maybe_get_output( |
| 28 | self, scheduler_output: "SchedulerOutput" |
| 29 | ) -> Generator[ECConnectorOutput | None, None, None]: |
| 30 | yield None |
| 31 | |
| 32 | def no_forward(self, scheduler_output: "SchedulerOutput") -> ModelRunnerOutput: |
| 33 | return EMPTY_MODEL_RUNNER_OUTPUT |
| 34 | |
| 35 | |
| 36 | class ActiveECConnector(ECConnector): |
| 37 | def __init__( |
| 38 | self, |
| 39 | vllm_config: VllmConfig, |
| 40 | encoder_cache: dict[str, torch.Tensor], |
| 41 | ) -> None: |
| 42 | self.encoder_cache = encoder_cache |
| 43 | self.ec_connector = get_ec_transfer() |
| 44 | assert isinstance(self.ec_connector, ECConnectorBase) |
| 45 | # Every producer offloads freshly computed encoder outputs, including |
| 46 | # an ec_both node that also reloads them. |
| 47 | self.save_new_caches = self.ec_connector.is_producer |
| 48 | |
| 49 | @contextmanager |
| 50 | def maybe_get_output( |
| 51 | self, scheduler_output: "SchedulerOutput" |
| 52 | ) -> Generator[ECConnectorOutput | None, None, None]: |
| 53 | if scheduler_output.ec_connector_metadata is None: |
| 54 | yield None |
| 55 | return |
| 56 | |
| 57 | output = ECConnectorOutput() |
| 58 | ec_connector = self.ec_connector |
| 59 | assert scheduler_output.ec_connector_metadata is not None |
| 60 | ec_connector.bind_connector_metadata(scheduler_output.ec_connector_metadata) |
| 61 | |
| 62 | if ec_connector.is_producer: |
| 63 | ec_connector.start_save_caches(encoder_cache=self.encoder_cache) |
| 64 | |
| 65 | if ec_connector.is_consumer: |
| 66 | ec_connector.start_load_caches(self.encoder_cache) |
| 67 | |
| 68 | cached_hashes = set(self.encoder_cache) if self.save_new_caches else None |
| 69 | try: |
| 70 | yield output |
| 71 | if cached_hashes is not None: |
| 72 | for mm_hash in self.encoder_cache.keys() - cached_hashes: |
| 73 | ec_connector.save_caches( |
| 74 | encoder_cache=self.encoder_cache, mm_hash=mm_hash |
| 75 | ) |
| 76 | finally: |
| 77 | output.finished_sending, output.finished_recving = ( |
| 78 | ec_connector.get_finished(scheduler_output.finished_req_ids) |
| 79 | ) |
| 80 | output.ec_connector_stats = ec_connector.get_ec_connector_stats() |
| 81 | output.ec_connector_worker_meta = ec_connector.build_connector_worker_meta() |
| 82 | ec_connector.clear_connector_metadata() |
| 83 | |
| 84 | def no_forward(self, scheduler_output: "SchedulerOutput") -> ModelRunnerOutput: |
| 85 | # EC send/recv even if no work to do. |
| 86 | with self.maybe_get_output(scheduler_output) as ec_connector_output: |
| 87 | pass |
| 88 | |
| 89 | return ModelRunnerOutput.with_ec_conn_output_only(ec_connector_output) |
| 90 | |
| 91 | |
| 92 | NO_OP_EC_CONNECTOR = ECConnector() |
| 93 | |
| 94 | |
| 95 | def get_ec_connector( |
| 96 | vllm_config: VllmConfig, encoder_cache: "EncoderCache | None" |
| 97 | ) -> ECConnector: |
| 98 | if ( |
| 99 | not has_ec_transfer() |
| 100 | or vllm_config.model_config.is_encoder_decoder |
| 101 | or encoder_cache is None |
| 102 | ): |
| 103 | return NO_OP_EC_CONNECTOR |
| 104 | |
| 105 | return ActiveECConnector(vllm_config, encoder_cache.encoder_outputs) |