sgl-project/sglang · critical · ValueError

ffn_hidden_size must be positive.

Error message

ffn_hidden_size must be positive.

What it means

The FFN hidden dimension must be positive because it is TP-sharded (each rank holds ffn_hidden_size/tp_size columns/rows). The constructor rejects non-positive values before allocating MLP weights.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/dits/minimax_h3.py:1813

            and not envs.SGLANG_CACHE_DIT_ENABLED
            and not hasattr(self, "_sglang_cache_dit_adapter")
            and not is_layerwise_offloaded_module(self)
            and all(type(block) is MiniMaxH3DiTBlock for block in self.blocks)
        )

    def _validate_tp_config(
        self, *, arch: MiniMaxH3DiTArchConfig, tp_size: int
    ) -> None:
        if tp_size <= 0:
            raise ValueError("TP size must be positive.")
        if arch.num_attention_heads <= 0:
            raise ValueError("num_attention_heads must be positive.")
        if arch.hidden_size <= 0:
            raise ValueError("hidden_size must be positive.")
        if arch.attention_head_dim <= 0:
            raise ValueError("attention_head_dim must be positive.")
        if arch.ffn_hidden_size <= 0:
            raise ValueError("ffn_hidden_size must be positive.")
        for name, value in (
            ("num_attention_heads", arch.num_attention_heads),
            ("hidden_size", arch.hidden_size),
            ("ffn_hidden_size", arch.ffn_hidden_size),
            ("time_embed_hidden_size", arch.time_embed_hidden_size),
            ("adaln_out_features", arch.adaln_out_features),
            ("final_adaln_out_features", arch.final_adaln_out_features),
            ("video_patch_output_dim", arch.latents_dim * math.prod(arch.patch_size)),
            ("audio_patch_output_dim", arch.audio_latents_dim),
        ):
            if value % tp_size:
                raise ValueError(
                    f"MiniMax H3 {name}={value} must be divisible by "
                    f"TP size {tp_size}."
                )

    @staticmethod
    def _validate_sequence_parallel_config(

View on GitHub (pinned to 0132848349)

Solutions

1. Set ffn_hidden_size to the checkpoint's MLP width (commonly a multiple like 4x hidden_size)
2. Check the config deserialization: confirm the JSON key matches the dataclass field
3. Add a pre-flight assert on all positive-size fields before construction

Example fix

# before
arch = MiniMaxH3DiTArchConfig(..., ffn_hidden_size=0)
# after
arch = MiniMaxH3DiTArchConfig(..., ffn_hidden_size=4 * hidden_size)

assert arch.ffn_hidden_size > 0

def ffn_ok(arch) -> bool:
    return getattr(arch, "ffn_hidden_size", 0) > 0

When it happens

Trigger: Building the model with an arch config where ffn_hidden_size is 0, negative, or missing (defaulted to 0), e.g. from a partially converted checkpoint.

Common situations: Config field renamed between versions (intermediate_size vs ffn_hidden_size); hand-written configs omitting the MLP width; YAML/JSON with a null value coerced to 0.

Related errors

• num_attention_heads must be positive.
• hidden_size must be positive.
• attention_head_dim must be positive.
• TP size must be positive.
• Invalid threshold_type for topk: {threshold_type}. Choose 'q

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/3e42fdb0f996649e. Report an issue: GitHub.