sgl-project/sglang · error · ValueError

rids_to_check cannot be used in PP mode

Error message

rids_to_check cannot be used in PP mode

What it means

Companion guard to the PP consensus check: in PP mode (pp_size > 1), pop_preallocated rejects rids_to_check because request resolution must be driven by the pipeline-wide pp_good_rids/pp_bad_rids consensus, not a per-stage check list. Passing both is contradictory.

Source

Thrown at python/sglang/srt/disaggregation/decode.py:1082

                    prefetched[1].cancel()

        self.pending_reqs = remaining

        for decode_req, prefill_dp_rank in resolved:
            decode_req.kv_receiver.init(prefill_dp_rank)

    def pop_preallocated(
        self,
        rids_to_check: Optional[List[str]] = None,
        pp_good_rids: Optional[List[str]] = None,
        pp_bad_rids: Optional[List[str]] = None,
    ) -> Tuple[List[DecodeRequest], List[DecodeRequest]]:
        """Pop the preallocated requests from the pending queue (FIFO)."""
        is_pp_mode = self.pp_size > 1
        if is_pp_mode and (pp_good_rids is None or pp_bad_rids is None):
            raise ValueError("PP consensus is required when pp_size > 1")
        if is_pp_mode and rids_to_check is not None:
            raise ValueError("rids_to_check cannot be used in PP mode")

        self._resolve_pending_reqs()
        self._update_handshake_waiters(rids_to_check, pp_good_rids, pp_bad_rids)
        if is_pp_mode:
            rids_to_check = set(pp_good_rids) | set(pp_bad_rids)

        failed_reqs = []
        preallocated_reqs = []
        indices_to_remove = set()

        # We need to make sure that the sum of inflight tokens and allocatable tokens is greater than maximum input+output length of each inflight request
        # Otherwise it is possible for one request running decode out of memory, while all other requests are in the transfer queue that cannot be retracted.
        retractable_tokens = sum(
            len(r.origin_input_ids) + len(r.output_ids)
            for r in self.scheduler.running_batch.reqs
        )

        uses_swa_tail_prealloc = self._uses_swa_tail_prealloc()

View on GitHub (pinned to 0132848349)

Solutions

1. Remove the rids_to_check argument and supply pp_good_rids/pp_bad_rids in PP mode
2. Branch caller code on pp_size: use rids_to_check only when pp_size == 1

Example fix

# before
good, bad = queue.pop_preallocated(rids_to_check=rids, pp_good_rids=g, pp_bad_rids=b)
# after
if queue.pp_size > 1:
    good, bad = queue.pop_preallocated(pp_good_rids=g, pp_bad_rids=b)
else:
    good, bad = queue.pop_preallocated(rids_to_check=rids)

if queue.pp_size > 1:
    assert rids_to_check is None, 'use pp_good_rids/pp_bad_rids in PP mode'

When it happens

Trigger: Calling pop_preallocated with pp_size > 1 and a non-None rids_to_check (alongside or instead of the consensus lists).

Common situations: Single-PP scheduler/test code reused on a PP deployment; upgrading a codebase where rids_to_check was the old API and the PP path was added later.

Related errors

• PP consensus is required when pp_size > 1
• Unexpected compressed-MLA dst_kv_ptrs length {len(dst_kv_ptr
• Destination MLA KV descriptors do not match prefill pp confi
• PD disagg: PP>1 not supported for MiniMax sparse index yet.
• Missing previous frame for delta payload

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/afbfb4ac05e3396a. Report an issue: GitHub.