sgl-project/sglang · error · ValueError

MiniMax H3 AdaLN cache is only compatible with unquantized w

Error message

MiniMax H3 AdaLN cache is only compatible with unquantized weights

What it means

The MiniMax H3 AdaLN cache (precomputed AdaLN conditioning tables loaded from adaln_cache_path or adaln_weight_files) is only valid for unquantized weights. Passing both a quantization config (quant_config) and an AdaLN cache is rejected because quantized kernels would consume the cached fp32 conditioning inconsistently.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/dits/minimax_h3.py:1879

                "the packed sequence alignment."
            )

    def __init__(
        self,
        config: MiniMaxH3DiTConfig,
        hf_config: dict[str, Any],
        quant_config: QuantizationConfig | None = None,
        adaln_cache_path: str | None = None,
        adaln_cache_model_variant: str | None = None,
        adaln_weight_files: list[str] | None = None,
        adaln_plan_width: int = MINIMAX_H3_ADALN_MAX_PLAN_WIDTH,
    ) -> None:
        super().__init__(config=config, hf_config=hf_config)
        arch = self.config
        if (
            adaln_cache_path is not None or adaln_weight_files is not None
        ) and quant_config is not None:
            raise ValueError(
                "MiniMax H3 AdaLN cache is only compatible with unquantized weights"
            )
        if arch.adaln_curve_grid is not None and (
            adaln_cache_path is not None or adaln_weight_files is not None
        ):
            raise ValueError(
                "MiniMax H3 pruned curve checkpoints cannot use a separate "
                "AdaLN cache"
            )
        self._adaln_precomputed = (
            adaln_cache_path is not None or adaln_weight_files is not None
        )
        self.arch = arch
        if arch.checkpoint_uses_diffusers_layout:
            self.preprocess_loaded_state_dict = _diffusers_h3_checkpoint
        self.hidden_size = arch.hidden_size
        self.num_attention_heads = arch.num_attention_heads
        self.num_channels_latents = arch.latents_dim

View on GitHub (pinned to 0132848349)

Solutions

1. Drop the AdaLN cache args and let AdaLN compute on the fly with the quantized checkpoint
2. Or use the unquantized checkpoint (quant_config=None) if you need the AdaLN cache's startup/time savings
3. Regenerate workflows: keep cache and quantization mutually exclusive in launch scripts

Example fix

# before
model = MiniMaxH3DiT(..., adaln_cache_path=cache, quant_config=fp8_cfg)
# after
model = MiniMaxH3DiT(..., quant_config=fp8_cfg)  # no adaln cache

if quant_config is not None:
    adaln_cache_path = None
    adaln_weight_files = None

def cache_quant_compatible(cache: Optional[str], qc) -> bool:
    return cache is None or qc is None

When it happens

Trigger: Constructing the model with adaln_cache_path (or adaln_weight_files) set while also passing a non-None quant_config (e.g. fp8/int8 quantized checkpoint).

Common situations: Reusing an AdaLN cache generated for the bf16 checkpoint after switching the server to a quantized model; enabling --quantization together with --adaln-cache flags.

Related errors

• Checkpoint quantization is encoded in per-layer metadata; do
• MiniMax H3 pruned curve checkpoints cannot use a separate Ad
• scale_shift_table must have shape [9, D]
• SGLang diffusion currently supports AutoRound auto_gptq chec
• AutoRound fused module {target!r} has inconsistent shard con

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/e75a12d2a1e3224d. Report an issue: GitHub.