vllm-contract

vllm/vllm/v1/core/sched/output.py · ef9339e · L245–279 · 固定 commit 上游

这是固定版本源码的只读摘录；可在线阅读，也可通过上游链接查看原文件。

245class SchedulerOutput:
246    # list of the requests that are scheduled for the first time.
247    # We cache the request's data in each worker process, so that we don't
248    # need to re-send it every scheduling step.
249    scheduled_new_reqs: list[NewRequestData]
250    # list of the requests that have been scheduled before.
251    # Since the request's data is already cached in the worker processes,
252    # we only send the diff to minimize the communication cost.
253    scheduled_cached_reqs: CachedRequestData
254
255    # req_id -> num_scheduled_tokens
256    # Number of tokens scheduled for each request.
257    num_scheduled_tokens: dict[str, int]
258    # Total number of tokens scheduled for all requests.
259    # Equal to sum(num_scheduled_tokens.values())
260    total_num_scheduled_tokens: int
261    # req_id -> spec_token_ids
262    # If a request does not have any spec decode tokens, it will not be
263    # included in the dictionary.
264    scheduled_spec_decode_tokens: dict[str, list[int]]
265    # req_id -> encoder input indices that need processing.
266    # E.g., if a request has [0, 1], it could mean the vision encoder needs
267    # to process that the request's 0-th and 1-th images in the current step.
268    scheduled_encoder_inputs: dict[str, list[int]]
269
270    # Request IDs that are finished in between the previous and the current
271    # steps. This is used to notify the workers about the finished requests
272    # so that they can free the cached states for those requests.
273    finished_req_ids: set[str]
274    # list of mm_hash strings associated with the encoder outputs to be
275    # freed from the encoder cache.
276    free_encoder_mm_hashes: list[str]
277
278    # Request-local CPU payload updates: None releases a consumed input; a
279    # payload restores it for re-encoding after KV-load rollback. GPU embeddings