sgl-project/sglang · critical · RuntimeError

{st.upper()} state index length mismatch: prefill={len(src_i

Error message

{st.upper()} state index length mismatch: prefill={len(src_indices)}, dst={len(dst_indices_local)}

What it means

Extra-state index lists (e.g. position/request-indexed components like C128_STATE) must match in length between prefill and decode; truncation would silently misalign rows and corrupt KV, so _requires_exact_state_index_match types raise instead of warn.

Source

Thrown at python/sglang/srt/disaggregation/mooncake/conn.py:1361

                ):
                    raise RuntimeError(
                        f"PD Disaggregation does NOT support PD different TP sizes for non-MLA {st.upper()} hybrid models yet."
                    )
                src_indices = list(indices)
                dst_indices_local = list(dst_indices)
                if (
                    st == StateType.C128_STATE
                    and len(src_indices) == 0
                    and len(dst_indices_local) == 0
                ):
                    continue
                if len(src_indices) != len(dst_indices_local):
                    # These components are position- or request-indexed:
                    # truncating silently misaligns rows and corrupts KV.
                    # Paged SWA/DSA tolerate a 1-page drift -> keep the
                    # lenient truncation below.
                    if self._requires_exact_state_index_match(st):
                        raise RuntimeError(
                            f"{st.upper()} state index length mismatch: "
                            f"prefill={len(src_indices)}, dst={len(dst_indices_local)}"
                        )
                    logger.warning(
                        f"len(prefill_state_indices) = {len(src_indices)}, len(dst_state_indices) = {len(dst_indices_local)}"
                    )
                    if len(src_indices) > len(dst_indices_local):
                        src_indices = src_indices[: len(dst_indices_local)]
                    else:
                        dst_indices_local = dst_indices_local[: len(src_indices)]
                rc = (
                    self._send_kvcache_generic(
                        mooncake_session_id=req.mooncake_session_id,
                        src_data_ptrs=src_data_ptrs,
                        dst_data_ptrs=dst_data_ptrs,
                        item_lens=src_item_lens,
                        prefill_data_indices=np.array(src_indices, dtype=np.int32),
                        dst_data_indices=np.array(dst_indices_local, dtype=np.int32),

View on GitHub (pinned to 0132848349)

Solutions

1. Check scheduler/allocator logs for why state index counts diverged (failed allocs, preemption on one side)
2. Restart both instances to reset allocator state
3. Ensure identical chunking/request scheduling config on both sides

catch RuntimeError mentioning 'index length mismatch'; drop the room's requests and alert — do not retry the same corrupted state

When it happens

Trigger: len(src_indices) != len(dst_indices_local) for an exactly-matched StateType during maybe_send_extra.

Common situations: Allocator drift producing different row counts per rank, scheduler divergence between PD sides, or earlier silent truncation bugs now surfaced.

Related errors

• Page size mismatch: prefill server has page_size={info.page_
• KV cache dtype mismatch: prefill server has kv_cache_dtype={
• --enable-unified-memory does not support different prefill /
• PD KV layout mismatch on the whole-envelope path: prefill ha
• Mamba state layouts differ between prefill and decode (src i

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/7bae6b4aa2805746. Report an issue: GitHub.