sgl-project/sglang · error · ValueError

Cosmos3TokenizationStage requires a tokenizer; expected the

Error message

Cosmos3TokenizationStage requires a tokenizer; expected the Qwen2 tokenizer loaded from the checkpoint's text_tokenizer/ subfolder.

What it means

Cosmos3TokenizationStage needs the Qwen2 text tokenizer (loaded from the checkpoint's text_tokenizer/ subfolder) to build input ids with the chat template. Constructing the stage with tokenizer=None means the checkpoint layout is wrong or the loader skipped the text tokenizer, so __init__ fails fast.

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/cosmos3.py:275

        cond_indexes = getattr(batch.sampling_params, "condition_frame_indexes", None)
        if not cond_indexes:
            return [0, 1]
        return sorted(set(int(i) for i in cond_indexes))


class Cosmos3TokenizationStage(PipelineStage):
    """Tokenization stage for Cosmos3.

    Applies the Qwen2 chat template, appends a duration suffix, and writes
    ``text_ids`` / ``text_mask`` into ``batch.extra`` for the denoising stage.
    """

    parallelism_type = StageParallelismType.REPLICATED

    def __init__(self, tokenizer):
        super().__init__()
        if tokenizer is None:
            raise ValueError(
                "Cosmos3TokenizationStage requires a tokenizer; expected the "
                "Qwen2 tokenizer loaded from the checkpoint's text_tokenizer/ "
                "subfolder."
            )
        self.tokenizer = tokenizer

    def verify_input(self, batch: Req, server_args: ServerArgs) -> VerificationResult:
        result = VerificationResult()
        result.add_check("prompt", batch.prompt, V.string_or_list_strings)
        return result

    def _tokenize_prompt(
        self,
        text: str | list[str],
        max_sequence_length: int,
        device: torch.device,
        use_system_prompt: bool = False,
        system_prompt: str | None = None,

View on GitHub (pinned to 0132848349)

Solutions

1. Verify the checkpoint contains text_tokenizer/ with Qwen2 tokenizer files (tokenizer.json/tokenizer_config.json) and re-download/copy it if missing
2. Fix the loader call so it actually loads the text tokenizer from that subfolder and pass the result to the stage
3. If converting your own checkpoint, include the Qwen2 tokenizer files under text_tokenizer/

Example fix

# before
stage = Cosmos3TokenizationStage(tokenizer=load_model(path).text_tokenizer_or_none)
# after
tok = AutoTokenizer.from_pretrained(os.path.join(ckpt, "text_tokenizer"))
assert tok is not None and tok.pad_token_id is not None
stage = Cosmos3TokenizationStage(tokenizer=tok)

import os
from transformers import AutoTokenizer
sub = os.path.join(ckpt_dir, "text_tokenizer")
assert os.path.isdir(sub) and os.listdir(sub), f"missing {sub}"

def usable_tokenizer(tok) -> bool:
    return tok is not None and getattr(tok, "pad_token_id", None) is not None

When it happens

Trigger: Instantiating Cosmos3TokenizationStage(tokenizer=None) — typically because the checkpoint directory lacks text_tokenizer/, files failed to download, or the loader returned None for the text tokenizer.

Common situations: Incomplete checkpoint download/transfer missing the text_tokenizer subfolder; using a Cosmos3 checkpoint converted without the Qwen2 text encoder files; loader version change that renamed or stopped auto-loading the subfolder.

Related errors

• Unexpected return type from apply_chat_template: {type(resul
• Action normalization stats not found at {stats_path}.
• scale_shift_table must have shape [9, D]
• Cosmos3 action input accepts one image field; use a list or
• Cosmos3 observation image arrays must use uint8 dtype

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/d200bb8899e5bdd8. Report an issue: GitHub.