Skip to content

vllm.v1.core.sched.async_scheduler

logger module-attribute 

logger = init_logger(__name__)

AsyncScheduler

Bases: Scheduler

Source code in vllm/v1/core/sched/async_scheduler.py 
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
class AsyncScheduler(Scheduler):
    def _update_after_schedule(
        self,
        scheduler_output: SchedulerOutput,
    ) -> None:
        super()._update_after_schedule(scheduler_output)
        pending_structured_output_tokens = False
        spec_decode_tokens = scheduler_output.scheduled_spec_decode_tokens
        for req_id in scheduler_output.num_scheduled_tokens:
            request = self.requests[req_id]
            pending_structured_output_tokens |= (
                request.use_structured_output and request.num_output_placeholders > 0
            )
            cur_num_spec_tokens = len(spec_decode_tokens.get(req_id, ()))
            if (
                request.num_computed_tokens
                == request.num_tokens
                + request.num_output_placeholders
                + cur_num_spec_tokens
            ):
                # The request will generate a new token plus num_spec_tokens
                # in this scheduling step.
                request.num_output_placeholders += 1 + cur_num_spec_tokens
                # Add placeholders for the new tokens in spec_token_ids.
                # Wwe will update the actual spec token ids in the worker process.
                request.spec_token_ids = [-1] * self.num_spec_tokens

        scheduler_output.pending_structured_output_tokens = (
            pending_structured_output_tokens
        )

    def _update_request_with_output(
        self,
        request: Request,
        new_token_ids: list[int],
    ) -> tuple[list[int], bool]:
        status_before_update = request.status
        new_token_ids, stopped = super()._update_request_with_output(
            request, new_token_ids
        )

        # Update the number of output placeholders.
        request.num_output_placeholders -= len(new_token_ids)
        assert request.num_output_placeholders >= 0

        # Cache the new tokens. Preempted requests should be skipped.
        if status_before_update == RequestStatus.RUNNING:
            self.kv_cache_manager.cache_blocks(
                request, request.num_computed_tokens - request.num_output_placeholders
            )
        return new_token_ids, stopped

_update_after_schedule 

_update_after_schedule(
    scheduler_output: SchedulerOutput,
) -> None
Source code in vllm/v1/core/sched/async_scheduler.py 
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
def _update_after_schedule(
    self,
    scheduler_output: SchedulerOutput,
) -> None:
    super()._update_after_schedule(scheduler_output)
    pending_structured_output_tokens = False
    spec_decode_tokens = scheduler_output.scheduled_spec_decode_tokens
    for req_id in scheduler_output.num_scheduled_tokens:
        request = self.requests[req_id]
        pending_structured_output_tokens |= (
            request.use_structured_output and request.num_output_placeholders > 0
        )
        cur_num_spec_tokens = len(spec_decode_tokens.get(req_id, ()))
        if (
            request.num_computed_tokens
            == request.num_tokens
            + request.num_output_placeholders
            + cur_num_spec_tokens
        ):
            # The request will generate a new token plus num_spec_tokens
            # in this scheduling step.
            request.num_output_placeholders += 1 + cur_num_spec_tokens
            # Add placeholders for the new tokens in spec_token_ids.
            # Wwe will update the actual spec token ids in the worker process.
            request.spec_token_ids = [-1] * self.num_spec_tokens

    scheduler_output.pending_structured_output_tokens = (
        pending_structured_output_tokens
    )

_update_request_with_output 

_update_request_with_output(
    request: Request, new_token_ids: list[int]
) -> tuple[list[int], bool]
Source code in vllm/v1/core/sched/async_scheduler.py 
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
def _update_request_with_output(
    self,
    request: Request,
    new_token_ids: list[int],
) -> tuple[list[int], bool]:
    status_before_update = request.status
    new_token_ids, stopped = super()._update_request_with_output(
        request, new_token_ids
    )

    # Update the number of output placeholders.
    request.num_output_placeholders -= len(new_token_ids)
    assert request.num_output_placeholders >= 0

    # Cache the new tokens. Preempted requests should be skipped.
    if status_before_update == RequestStatus.RUNNING:
        self.kv_cache_manager.cache_blocks(
            request, request.num_computed_tokens - request.num_output_placeholders
        )
    return new_token_ids, stopped