sgl-project/sglang · error · ValueError

Cannot determine processor class for {model_path}

Error message

Cannot determine processor class for {model_path}

What it means

When auto-loading the processor class fails, sglang falls back to _build_processor_manually using the processor_class/auto_map entry from preprocessor_config.json; that reference is missing too, so no processor class can be constructed for the model.

Source

Thrown at python/sglang/srt/utils/hf_transformers/processor.py:162

    auto_map = getattr(config, "auto_map", None) or {}
    proc_ref = auto_map.get("AutoProcessor")
    if not proc_ref:
        try:
            pp_file = _resolve_local_or_cached_file(
                model_path, "preprocessor_config.json", revision
            )
            with open(pp_file) as f:
                pp_auto_map = json.load(f).get("auto_map", {})
            proc_ref = pp_auto_map.get("AutoProcessor")
        except (OSError, json.JSONDecodeError, ValueError) as e:
            logger.warning(
                "_build_processor_manually: could not read preprocessor_config.json "
                "for %s: %s",
                model_path,
                e,
            )
    if not proc_ref:
        raise ValueError(f"Cannot determine processor class for {model_path}")

    proc_cls = get_class_from_dynamic_module(
        proc_ref, model_path, code_revision=revision
    )

    # Load sub-components individually (these succeed)
    tokenizer = AutoTokenizer.from_pretrained(
        model_path, trust_remote_code=trust_remote_code, revision=revision
    )
    init_kwargs = {"tokenizer": tokenizer}

    if "image_processor" in getattr(proc_cls, "attributes", []):
        try:
            init_kwargs["image_processor"] = AutoImageProcessor.from_pretrained(
                model_path, trust_remote_code=trust_remote_code, revision=revision
            )
        except (ImportError, OSError, ValueError) as e:
            raise RuntimeError(

View on GitHub (pinned to 0132848349)

Solutions

1. Ensure preprocessor_config.json contains a processor_class or auto_map entry pointing at the processor implementation
2. Re-download the model to restore missing files
3. Pass trust_remote_code=True if the model ships remote code

import json; pc = json.load(open(Path(model)/'preprocessor_config.json'))
proc_ref = pc.get('auto_map',{}).get('AutoProcessor') or pc.get('processor_class')
if not proc_ref: expect manual-build failure

try:
    get_processor(model)
except ValueError as e:
    if 'Cannot determine processor class' in str(e): re-download model or use transformers AutoProcessor directly

When it happens

Trigger: Loading a multimodal model whose repo lacks both a resolvable auto_map/processor_class in preprocessor_config.json and standard processor files.

Common situations: Partial uploads, custom models missing processor metadata, or trust_remote_code models whose remote processor code path is absent.

Related errors

• Cannot find corresponding multimodal processor registered in
• Required `vision_config.model_type` is not found in hf_confi
• cos/sin shape does not cover image tokens and head_dim
• Unsupported model type: {model_type}
• Unsupported image type: {type(image)}

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/d535c5ca0542d28d. Report an issue: GitHub.