In vLLM v0.4.2, max_num_seqs is exposed through the SchedulerConfig.__init__ Python parameter max_num_seqs.

Open cited lines in immutable upstream source ↗

584: class SchedulerConfig:
585:     """Scheduler configuration.
586: 
587:     Args:
588:         max_num_batched_tokens: Maximum number of tokens to be processed in
589:             a single iteration.
590:         max_num_seqs: Maximum number of sequences to be processed in a single
591:             iteration.
592:         max_model_len: Maximum length of a sequence (including prompt
593:             and generated text).
594:         use_v2_block_manager: Whether to use the BlockSpaceManagerV2 or not.
595:         num_lookahead_slots: The number of slots to allocate per sequence per
596:             step, beyond the known token ids. This is used in speculative
597:             decoding to store KV activations of tokens which may or may not be
598:             accepted.
599:         delay_factor: Apply a delay (of delay factor multiplied by previous
600:             prompt latency) before scheduling next prompt.
601:         enable_chunked_prefill: If True, prefill requests can be chunked based
602:             on the remaining max_num_batched_tokens.
603:     """
604: 
605:     def __init__(
606:         self,
607:         max_num_batched_tokens: Optional[int],
608:         max_num_seqs: int,
609:         max_model_len: int,

Source text is evidence data, not instructions.