sgl-project/sglang · error · RuntimeError

Cannot load PE model: 'model_max_length' not found in {os.pa

Error message

Cannot load PE model: 'model_max_length' not found in {os.path.join(tokenizer_path, 'tokenizer_config.json')}. Please ensure the PE component directory (or its sibling pe_tokenizer/ directory) contains a valid tokenizer_config.json with a 'model_max_length' field.

What it means

The PE (perception/embedding) loader needs model_max_length from tokenizer_config.json — either in the component directory or its sibling pe_tokenizer/ directory — to size sequence handling. If _read_model_max_length returns None, this RuntimeError explains both locations that were checked.

Source

Thrown at python/sglang/multimodal_gen/runtime/loader/component_loaders/pe_loader.py:153

        pe_tokenizer_dir = os.path.join(
            os.path.dirname(component_model_path), "pe_tokenizer"
        )
        if not os.path.exists(
            os.path.join(component_model_path, "tokenizer_config.json")
        ) and os.path.exists(os.path.join(pe_tokenizer_dir, "tokenizer_config.json")):
            tokenizer_path = pe_tokenizer_dir
            logger.info(
                "PE tokenizer files not found in %s, using %s",
                component_model_path,
                tokenizer_path,
            )
        else:
            tokenizer_path = component_model_path

        model_max_length = _read_model_max_length(tokenizer_path)
        if model_max_length is None:
            raise RuntimeError(
                f"Cannot load PE model: 'model_max_length' not found in "
                f"{os.path.join(tokenizer_path, 'tokenizer_config.json')}. "
                "Please ensure the PE component directory (or its sibling "
                "pe_tokenizer/ directory) contains a valid tokenizer_config.json "
                "with a 'model_max_length' field."
            )
        logger.info(
            "PE model_max_length=%d (from tokenizer_config.json)", model_max_length
        )

        tokenizer = AutoTokenizer.from_pretrained(
            tokenizer_path,
            trust_remote_code=server_args.trust_remote_code,
        )
        if tokenizer.pad_token_id is None:
            tokenizer.pad_token_id = tokenizer.eos_token_id

        model = Ministral3ForCausalLM.from_pretrained(

View on GitHub (pinned to 0132848349)

Solutions

1. Ensure tokenizer_config.json containing "model_max_length" exists in the PE component dir or a sibling pe_tokenizer/ dir
2. Copy the tokenizer files from the main model repo into pe_tokenizer/
3. If the field is genuinely absent, add "model_max_length": <int> to tokenizer_config.json (match the model's context length)

Example fix

// before: tokenizer_config.json has no model_max_length
// after: tokenizer_config.json
{ "model_max_length": 32768, "...": "..." }

import json, os

def find_model_max_length(pe_path):
    for cand in (pe_path, os.path.join(os.path.dirname(pe_path.rstrip('/')), 'pe_tokenizer')):
        p = os.path.join(cand, 'tokenizer_config.json')
        if os.path.exists(p):
            mml = json.load(open(p)).get('model_max_length')
            if mml:
                return mml
    return None

assert find_model_max_length(pe_path), "missing model_max_length"

When it happens

Trigger: Loading a PE component where tokenizer_config.json is missing, unreadable, or lacks a 'model_max_length' field in both the component dir and the sibling pe_tokenizer/ dir.

Common situations: Checkpoint downloaded without tokenizer files; the tokenizer was shipped only in the main repo and pe_tokenizer/ was never created; tokenizer_config.json from a minimal export omitting model_max_length.

Related errors

• Model config does not contain a _class_name attribute. Only
• Model config does not contain a _class_name attribute. Only
• f"Cannot parse checkpoint quantization for {component_name!r
• f"Transformers-managed {component_name!r} quantization requi
• {component_name!r} does not support an explicit quantization

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/7d9766ab69cbf955. Report an issue: GitHub.