sgl-project/sglang · error · ValueError

task {self.task!r} does not allow condition role={role!r} ty

Error message

task {self.task!r} does not allow condition role={role!r} type={condition_type!r}

What it means

MiniMaxH3TaskProfile.rule_for looks up a condition rule by (role, condition_type) pair within the task's declared condition_rules. If no rule matches, the task profile explicitly disallows that condition combination, and the ValueError names the task, role, and type. This is called during condition validation and plan resolution.

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/minimax_h3/task_profiles.py:132

    # For ref2va video, target may omit duration_seconds
    # when an audio reference is present — duration derives from the
    # reference audio probe and is then checked against the shared 4-15s range.
    duration_from_audio_reference: bool = False
    # Some deployments keep video references disabled while retaining the
    # routing rule for projection/contract compatibility.
    video_reference_supported: bool = True
    min_condition_count: int | None = None
    max_condition_count: int | None = None
    # Resolution policy for target.aspect_ratio="auto".  A concrete ratio
    # resolves immediately; None keeps geometry deferred to the named source.
    auto_aspect_ratio: str | None = None
    auto_geometry_source: str | None = None

    def rule_for(self, *, role: str, condition_type: str) -> MiniMaxH3ConditionRule:
        for rule in self.condition_rules:
            if rule.role == role and rule.condition_type == condition_type:
                return rule
        raise ValueError(
            f"task {self.task!r} does not allow condition "
            f"role={role!r} type={condition_type!r}"
        )


_BASE_COMPONENTS = (
    "processor",
    "text_encoder",
    "tokenizer",
    "transformer",
    "video_vae",
    "audio_vae",
)

_BRANCHES = tuple(dict(branch) for branch in MINIMAX_H3_DEFAULT_BRANCHES)
MINIMAX_H3_TASK_PROFILES: dict[str, MiniMaxH3TaskProfile] = {
    MINIMAX_H3_TASK_T2VA: MiniMaxH3TaskProfile(
        task=MINIMAX_H3_TASK_T2VA,

View on GitHub (pinned to 0132848349)

Solutions

1. Check the condition_rules on the task's MiniMaxH3TaskProfile to see allowed role/type pairs
2. Match the task to the conditions you send: t2va takes text only, fl2va takes keyframe images, ref2va takes a reference video
3. Fix the role or condition_type spelling in the request

Example fix

# before
resolve_plan(task="t2va", conditions=[Condition(role="target", condition_type="image", ...)])

# after
resolve_plan(task="fl2va", conditions=[Condition(role="target", condition_type="image", ...)])

profile = minimax_h3_task_profile(task)
allowed = {(r.role, r.condition_type) for r in profile.condition_rules}
for c in conditions:
    if (c.role, c.condition_type) not in allowed:
        raise ValueError(f"task {task} disallows {c.role}/{c.condition_type}")

def condition_allowed(task: str, role: str, condition_type: str) -> bool:
    profile = minimax_h3_task_profile(task)
    return any(r.role == role and r.condition_type == condition_type for r in profile.condition_rules)

catch ValueError from minimax_h3_resolve_plan and re-raise with the task's allowed role/type table

When it happens

Trigger: Calling minimax_h3_resolve_plan (or _validate_conditions) with a condition whose role/condition_type pair isn't declared for the task, e.g. role='reference' type='video' on a t2va profile (text-only task), or a mismatched pairing like role='target' type='video' where only type='image' is declared.

Common situations: Sending an image or video condition to a t2va (text-to-video) request; mixing up role names ('source' vs 'target'); adding a new condition type without extending the task profile's condition_rules.

Understand the failure class

Background: Schema validation failed / invalid input schema: payload rejected because its shape doesn't match the expected schema — this error's family across 28 libraries.

Related errors

• MiniMax-H3 adaln_t_table must have shape [N, D] with N >= 2,
• MiniMax H3 AdaLN cache has invalid timestep plans
• TP size must be positive.
• num_attention_heads must be positive.
• hidden_size must be positive.

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/6bd6d951778b4450. Report an issue: GitHub.