sgl-project/sglang · error · ComponentCheckpointUnsupportedError

{component_name!r} already declares checkpoint quantization;

Error message

{component_name!r} already declares checkpoint quantization; drop the explicit online quantization override

What it means

ComponentCheckpointUnsupportedError raised when an explicit online quantization override is supplied AND the checkpoint already carries a serialized quantization config. Applying online quantization on top of an already-quantized checkpoint is contradictory, so the loader asks you to drop the override.

Source

Thrown at python/sglang/multimodal_gen/runtime/loader/component_loaders/text_encoder_loader.py:289

    _delegate_standard_bnb4_to_transformers(
        component_config,
        component_name,
    )
    try:
        quant_config = _get_encoder_quant_config(
            component_config,
            component_model_path,
            component_weights_path,
            model_cls,
        )
    except (KeyError, NotImplementedError, TypeError, ValueError) as error:
        raise ComponentCheckpointUnsupportedError(
            f"Cannot configure checkpoint quantization for {component_name!r}: {error}"
        ) from error
    model_config.quant_config = quant_config
    if explicit_quantization is not None:
        if quant_config is not None:
            raise ComponentCheckpointUnsupportedError(
                f"{component_name!r} already declares checkpoint quantization; "
                "drop the explicit online quantization override"
            )
        if explicit_quantization not in _ONLINE_ENCODER_QUANTIZATIONS:
            raise ComponentCheckpointUnsupportedError(
                f"Online quantization {explicit_quantization!r} is not supported "
                f"for native encoders; choose one of "
                f"{sorted(_ONLINE_ENCODER_QUANTIZATIONS)}"
            )
        from sglang.multimodal_gen.runtime.layers.quantization import (
            get_quantization_config,
        )

        model_config.quant_config = get_quantization_config(explicit_quantization)(
            ignored_layers=ignored_layers
        )
        quant_config = model_config.quant_config
    if quant_config is None:

View on GitHub (pinned to 0132848349)

Solutions

1. Remove the explicit online quantization argument/flag for this component — the checkpoint's serialized config wins
2. If you intended online quantization, switch to an unquantized (BF16/FP16) checkpoint of the encoder
3. Exclude self-quantized components from global quantization flag propagation

Example fix

# before
load_customized(..., explicit_quantization="fp8")  # checkpoint is already fp8

# after
load_customized(..., explicit_quantization=None)

if explicit_quantization is not None and checkpoint_declares_quant(model_path):
    explicit_quantization = None  # checkpoint wins

When it happens

Trigger: Calling load_customized with explicit_quantization="fp8" (etc.) for a component whose config.json/weights already declare quantization, so _get_encoder_quant_config returned a non-None quant_config.

Common situations: Server started with a global --quantization fp8 flag while loading an already-FP8-serialized CLIP encoder; mixing PD/D deployment flags with pre-quantized component checkpoints.

Related errors

• {component_name!r} manages its own checkpoint quantization a
• Online quantization {explicit_quantization!r} is not support
• Online quantization for {component_name!r} requires an in-tr
• The SRT encoder checkpoint adapter supports only serialized
• Serialized quantized component weights cannot use a stacked

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/af8cde52840085b1. Report an issue: GitHub.