sgl-project/sglang · critical · ValueError

Kimi expert object is not contiguous at index {index}

Error message

Kimi expert object is not contiguous at index {index}

What it means

For each (layer, expert) object, the three role blobs must be physically contiguous: each role's offset must equal the object's base offset plus the role's cumulative role_offsets. A mismatch means the pack interleaves objects or has gaps, which the mmap-based loader cannot handle.

Source

Thrown at python/sglang/srt/layers/moe/expert_pack.py:856

                expected_nbytes = role_nbytes[physical_role]
                if (
                    nbytes != expected_nbytes
                    or offset % int(pack_manifest["alignment"])
                    or offset < previous_end
                    or offset + nbytes > self.path.stat().st_size
                ):
                    raise ValueError(
                        f"Kimi expert-pack range mismatch at index {index}"
                    )
                previous_end = offset + nbytes
                object_key = (layer, expert)
                if physical_role_id == 0:
                    self.object_offsets[object_key] = offset
                expected_offset = (
                    self.object_offsets[object_key] + role_offsets[physical_role]
                )
                if offset != expected_offset:
                    raise ValueError(
                        f"Kimi expert object is not contiguous at index {index}"
                    )
                generation_bytes = hashlib.sha256(
                    f"{pack_manifest['index_sha256']}:{layer}:{expert}".encode("ascii")
                ).digest()
                generation = int.from_bytes(generation_bytes[:8], "little") or 1
                logical_role_id = ROLE_NAMES.index(physical_role)
                logical_shape = tuple(
                    int(value) for value in roles[physical_role]["logical_shape"]
                )
                self.entries[(layer, expert, logical_role_id)] = ExpertPackEntry(
                    layer=layer,
                    expert=expert,
                    role_id=logical_role_id,
                    dtype_id=int(roles[physical_role]["dtype_id"]),
                    dtype=str(roles[physical_role]["dtype"]),
                    tensor_name=name,
                    source_slice_offset=0,

View on GitHub (pinned to 0132848349)

Solutions

1. Re-pack so each expert's up/gate/down blobs are adjacent in the fixed physical role order
2. If writing a custom packer, compute each role offset as object_base + role_offsets[role]

null

try:
    pack = KimiExpertPack(path, manifest)
except ValueError as e:
    if 'not contiguous' in str(e):
        log.error('pack layout groups by role, not per-expert; re-pack required')
    raise

When it happens

Trigger: A pack that stores tensors grouped by role (all ups, then all gates) instead of grouped per expert; or gaps/inserted padding between an expert's role blobs.

Common situations: See trigger scenarios.

Related errors

• Kimi active routed MoE layers must be exactly 1..92
• Kimi manifest expert-pack path does not match pack_path
• Kimi expert-pack size does not match its manifest
• Kimi expert-pack physical role order is unsupported
• Kimi expert-pack role sizes do not match object bytes

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/6c75edced53aa376. Report an issue: GitHub.