sgl-project/sglang · error · ValueError

Source contains multiple independent weight files; select on

Error message

Source contains multiple independent weight files; select one with an exact file URL or weight name. Candidates: {list(candidates)}

What it means

Multiple candidate weight files were found and no weight_name or exact file URL was provided, so the choice is ambiguous and the resolver requires an explicit selection.

Source

Thrown at python/sglang/multimodal_gen/runtime/weights/source.py:255

    )


def select_weight_file(
    inventory: WeightInventory, weight_name: str | None = None
) -> str:
    """Select weights deterministically; never guess among independent files."""
    candidates = tuple(
        path for path in inventory.files if path.lower().endswith(_WEIGHT_SUFFIXES)
    )
    if inventory.source.filename is not None:
        return inventory.files[0]
    if weight_name is not None:
        return _select_named_file(candidates, weight_name)
    if len(candidates) == 1:
        return candidates[0]
    if not candidates:
        raise FileNotFoundError("Source contains no recognized weight files")
    raise ValueError(
        "Source contains multiple independent weight files; select one with "
        f"an exact file URL or weight name. Candidates: {list(candidates)}"
    )


def resolve_weight(
    source: str,
    *,
    revision: str | None = None,
    weight_name: str | None = None,
) -> ResolvedWeight:
    """Resolve one weight file without downloading its tensor payload."""
    parsed_source = parse_weight_source(source, revision=revision)
    inventory = resolve_weight_inventory(parsed_source)
    selected_file = select_weight_file(inventory, weight_name)
    return ResolvedWeight(
        inventory=inventory,
        selected_file=selected_file,

View on GitHub (pinned to 0132848349)

Solutions

1. Pass weight_name for the desired basename
2. Use an exact file URL to pin one file
3. Restrict the source to the subfolder containing the desired file

Example fix

# before
resolve_weight("org/repo")
# after
resolve_weight("org/repo", weight_name="model_fp8.safetensors")

inv = resolve_weight_inventory(parse_weight_source(source))
if len(inv.files) > 1 and weight_name is None:
    raise SystemExit(f"multiple weights {inv.files}; pass weight_name")

try:
    f = select_weight_file(inv)
except ValueError as e:
    logger.error("%s", e)  # lists candidates
    raise

When it happens

Trigger: A repo or directory containing several independent recognized weight files (e.g. both a full checkpoint and a sharded/fp8 variant) and calling resolve_weight(source) with no selector.

Common situations: Mixed-format checkpoints in one folder; repos with bf16 and fp8 exports side by side.

Related errors

• Weight name {weight_name!r} matches multiple files: {list(ba
• Duplicate tensor names detected across safetensors files. Re
• RIFE weight file not found: {flownet_path} Expected layout:
• Invalid Hugging Face {field_name}: {path!r}
• Weight URL pins revision {url_revision!r}, which conflicts w

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/d5eabe20dc90b78c. Report an issue: GitHub.