sgl-project/sglang · error · ValueError

Found static activation scheme for checkpoint that was not…

Error message

Found static activation scheme for checkpoint that was not serialized fp8.

What it means

When an FP8 MoE quant config declares activation_scheme="static", the loader expects per-expert input scales (w13/w2 input_scale tensors) which only exist in FP8-serialized checkpoints. If the checkpoint was not FP8-serialized, those scales cannot exist, so create_fp8_moe_weight_ raises this contradiction.

Solutions

Use a checkpoint that actually ships FP8 weights and static input scales
Change activation_scheme to "dynamic" (works with on-the-fly quantization)
Remove/fix the quantization_config if the model is not quantized

Example fix

// before
{"quant_method":"fp8","activation_scheme":"static"}  // on BF16 weights
// after
{"quant_method":"fp8","activation_scheme":"dynamic"}

Defensive patterns

Strategy: validation

Validate before calling

qcfg = model_config.quantization_config
if qcfg.get("activation_scheme") == "static" and not qcfg.get("is_checkpoint_fp8_serialized", False):
    raise SystemExit("static activation scheme requires an FP8-serialized checkpoint")

Type guard

def static_scheme_consistent(qcfg: dict) -> bool:
    return qcfg.get("activation_scheme") != "static" or bool(qcfg.get("is_checkpoint_fp8_serialized"))

Prevention

Validate static-scheme configs against checkpoint format before load
Prefer dynamic scheme for on-the-fly quantization

When it happens

Trigger: Creating MoE weights with quant_config.activation_scheme == "static" and quant_config.is_checkpoint_fp8_serialized == False — a config claiming static FP8 activation quantization over a non-FP8 checkpoint.

Common situations: Hand-written quantization_configs copied from FP8 models onto BF16 weights; configs produced by tools that always emit "static" regardless of serialization; misconverted checkpoints.

Related errors

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/3ed5a48a68a8c574. Report an issue: GitHub.

Appendix: source

Thrown at python/sglang/srt/layers/quantization/fp8.py:1384

        # If loading fp8 checkpoint, pass the weight loaders.
        # If loading an fp16 checkpoint, do not (we will quantize in
        #   process_weights_after_loading()
        if quant_config.is_checkpoint_fp8_serialized:
            set_weight_attrs(w13_weight_scale, extra_weight_attrs)
            set_weight_attrs(w2_weight_scale, extra_weight_attrs)

            if _is_hip and _use_hip_int4:
                extra_weight_attrs.update(
                    {"quant_method": FusedMoeWeightScaleSupported.CHANNEL.value}
                )
                set_weight_attrs(w13_weight_scale1, extra_weight_attrs)
                set_weight_attrs(w2_weight_scale1, extra_weight_attrs)

        # INPUT_SCALES
        if quant_config.activation_scheme == "static":
            if not quant_config.is_checkpoint_fp8_serialized:
                raise ValueError(
                    "Found static activation scheme for checkpoint that "
                    "was not serialized fp8."
                )

            w13_input_scale = torch.nn.Parameter(
                torch.ones(num_experts, dtype=torch.float32), requires_grad=False
            )
            layer.register_parameter("w13_input_scale", w13_input_scale)
            set_weight_attrs(w13_input_scale, extra_weight_attrs)

            w2_input_scale = torch.nn.Parameter(
                torch.ones(num_experts, dtype=torch.float32), requires_grad=False
            )
            layer.register_parameter("w2_input_scale", w2_input_scale)
            set_weight_attrs(w2_input_scale, extra_weight_attrs)

        else:
            layer.w13_input_scale = None

View on GitHub (pinned to 0132848349)