sgl-project/sglang · error · ValueError

Unsupported activation scheme {activation_scheme}

Error message

Unsupported activation scheme {activation_scheme}

What it means

The Fp8LinearConfig constructor validates activation_scheme against the ACTIVATION_SCHEMES whitelist ('static' and 'dynamic'). Any other string — including 'none', 'None' handled elsewhere, typos, or new scheme names — raises this ValueError at config construction time, before weights are created.

Source

Thrown at python/sglang/srt/layers/quantization/fp8.py:248

        is_checkpoint_fp8_serialized: bool = False,
        activation_scheme: str = "dynamic",
        ignored_layers: Optional[List[str]] = None,
        weight_block_size: List[int] = None,
        packed_modules_mapping: Optional[Dict[str, List[str]]] = None,
        use_mxfp8: bool = False,
        is_fp4_experts: bool = False,
        kv_cache_quant_algo: Optional[str] = None,
    ) -> None:
        super().__init__()
        # DSV4 mxfp4-packed (True) vs converted FP8 (False); injected by
        # model_loader from ModelConfig. Default False off the DSV4 path.
        self.is_fp4_experts = is_fp4_experts
        self.dequant_fp4_to_fp8 = False
        self.is_checkpoint_fp8_serialized = is_checkpoint_fp8_serialized
        if is_checkpoint_fp8_serialized:
            log_info_on_rank0(logger, "Detected fp8 checkpoint.")
        if activation_scheme not in ACTIVATION_SCHEMES:
            raise ValueError(f"Unsupported activation scheme {activation_scheme}")
        self.activation_scheme = activation_scheme
        self.ignored_layers = ignored_layers or []
        if ignored_layers_str := envs.SGLANG_FP8_IGNORED_LAYERS.get():
            self.ignored_layers.extend(
                [
                    layer.strip()
                    for layer in ignored_layers_str.split(",")
                    if layer.strip()
                ]
            )
        self.packed_modules_mapping = packed_modules_mapping or {}
        self.use_mxfp8 = use_mxfp8
        self.kv_cache_quant_algo = kv_cache_quant_algo
        if weight_block_size is not None:
            if not is_checkpoint_fp8_serialized:
                raise ValueError(
                    "The block-wise quantization only supports fp8-serialized checkpoint for now."
                )

View on GitHub (pinned to 0132848349)

Solutions

1. Set activation_scheme to "static" or "dynamic" in the model's config.json quantization_config
2. If the checkpoint truly uses a per-token scheme, map it to "dynamic" which covers per-token dynamic quantization
3. Check for trailing whitespace/case issues in the config value

Example fix

// before
"quantization_config": { "quant_method": "fp8", "activation_scheme": "static_per_token" }
// after
"quantization_config": { "quant_method": "fp8", "activation_scheme": "dynamic" }

from sglang.srt.layers.quantization.fp8 import Fp8LinearConfig  # ACTIVATION_SCHEMES lives in config module
scheme = qcfg.get("activation_scheme", "dynamic")
if scheme not in ("static", "dynamic"):
    qcfg["activation_scheme"] = "dynamic"  # or hard-fail with a clear message

def is_valid_activation_scheme(s: str) -> bool:
    return s in {"static", "dynamic"}

try:
    cfg = Fp8LinearConfig(...)
except ValueError as e:
    if "activation scheme" in str(e):
        cfg = Fp8LinearConfig(..., activation_scheme="dynamic")
    else:
        raise

When it happens

Trigger: Constructing Fp8LinearConfig(activation_scheme=...) with a value outside ACTIVATION_SCHEMES; usually triggered by parsing a checkpoint's quantization_config where activation_scheme is something like "static_per_token", "" (empty), or miscapitalized "Dynamic".

Common situations: Fine-tuned checkpoints saved with custom quant tools that write non-standard activation_scheme values; hand-edited config.json quantization sections; schemas that use "delayed" or other schemes SGLang hasn't mapped.

Understand the failure class

Background: "Invalid value" and "allowed values are" config errors: what your library rejected and how to fix it — this error's family across 41 libraries.

Related errors

• The hpc_ops MoE runner backend only supports FP8-quantized M
• The block-wise quantization only supports fp8-serialized che
• The quantization block size of weight must have 2 dimensions
• Found static activation scheme for checkpoint that was not s
• FP8 weight_block_size must contain two positive integers, go

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/64710ddf13640a8c. Report an issue: GitHub.