sgl-project/sglang · error · ValueError

MiniMax H3 text payload positive.text_len must match the hid

Error message

MiniMax H3 text payload positive.text_len must match the hidden-state sequence dimension

What it means

The payload's positive.text_len must be an int equal to hidden_states.shape[0] (the sequence dimension). This check keeps prompt_seq_lens consistent with the embeddings actually published to batch.prompt_embeds; any mismatch means the payload is internally inconsistent.

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/minimax_h3/stages/text_encoding.py:231

        H3 keeps token tags and presentation metadata in ``Req.extra``, but
        the shared TextEncodingStage contract still owns ``prompt_embeds``.
        Publishing the same tensor there preserves native verification,
        grouped-request deduplication, and downstream memory accounting
        without duplicating the embedding storage.
        """
        payload = batch.extra.get(MINIMAX_H3_TEXT_EMBEDDINGS_EXTRA_KEY)
        positive = payload.get("positive") if isinstance(payload, dict) else None
        hidden_states = (
            positive.get("hidden_states") if isinstance(positive, dict) else None
        )
        text_len = positive.get("text_len") if isinstance(positive, dict) else None
        if not isinstance(hidden_states, torch.Tensor) or hidden_states.ndim < 2:
            raise ValueError(
                "MiniMax H3 text payload must contain positive.hidden_states "
                "with at least two dimensions"
            )
        if not isinstance(text_len, int) or text_len != int(hidden_states.shape[0]):
            raise ValueError(
                "MiniMax H3 text payload positive.text_len must match the "
                "hidden-state sequence dimension"
            )
        batch.prompt_embeds = [hidden_states]
        batch.prompt_seq_lens = [[text_len]]

    def _encode_from_plan(
        self,
        batch: Req,
        plan,
        *,
        include_video_token_mask: bool = False,
    ) -> None:
        """Encode the positive Qwen3VL presentation into layer-50 states.

        MiniMax H3 only supports the CFG-distilled model path, so every task
        emits exactly one positive embedding payload. ComponentManager owns
        residency/offload, while every folded-TP rank enters the encoder

View on GitHub (pinned to 0132848349)

Solutions

1. Recompute text_len from the final tensor: int(hidden_states.shape[0]) and set it in the payload
2. Convert tensor/float text_len values to int before publishing
3. Avoid mutating hidden_states after the encoder computed text_len

Example fix

// before
payload["positive"]["text_len"] = old_len  # before truncation
// after
hs = payload["positive"]["hidden_states"]
payload["positive"]["text_len"] = int(hs.shape[0])

hs = payload["positive"]["hidden_states"]
payload["positive"]["text_len"] = int(hs.shape[0])  # keep consistent before publish

def text_len_matches(payload) -> bool:
    pos = payload.get("positive", {})
    tl, hs = pos.get("text_len"), pos.get("hidden_states")
    return isinstance(tl, int) and isinstance(hs, torch.Tensor) and tl == hs.shape[0]

When it happens

Trigger: _publish_native_text_conditioning finds text_len missing (None), a non-int (e.g. tensor or float), or an int different from hidden_states.shape[0] — e.g. embeddings truncated/padded after text_len was computed.

Common situations: Post-processing embeddings (slicing, pooling, padding) without updating text_len; encoders returning text_len as a 0-d tensor instead of int; chunked encodes appending tokens after length was recorded.

Related errors

• prompt must be non-empty
• MiniMax H3 text encode produced no native payload
• MiniMax H3 text encode failed on rank {owner}: {owner_error}
• MiniMax H3 text payload must contain positive.hidden_states
• MiniMaxH3TextEncodingStage direct encode requires a text_enc

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/02bcd076690fe2c9. Report an issue: GitHub.