sgl-project/sglang · error · ValueError

{name} must be a floating point tensor

Error message

{name} must be a floating point tensor

What it means

The unit-timestep validator requires a floating-point dtype tensor (float32/float16/bfloat16). Integer-dtype timesteps — the diffusers convention (e.g. 700, 250 as long tensors) — are rejected because the flow-matching math treats timestep as a continuous value in [0,1].

Source

Thrown at python/sglang/multimodal_gen/runtime/models/schedulers/scheduling_minimax_h3_euler_ancestral.py:19

# SPDX-License-Identifier: Apache-2.0
from __future__ import annotations

import math
from typing import Any

import torch


def _require_finite_tensor(tensor: torch.Tensor, name: str) -> None:
    if not bool(torch.isfinite(tensor).all().item()):
        raise ValueError(f"{name} must be finite")


def _validate_unit_timestep(timestep: torch.Tensor, name: str) -> None:
    if not isinstance(timestep, torch.Tensor):
        raise ValueError(f"{name} must be a torch.Tensor")
    if not torch.is_floating_point(timestep):
        raise ValueError(f"{name} must be a floating point tensor")
    _require_finite_tensor(timestep, name)
    out_of_range = (timestep < 0) | (timestep > 1)
    if bool(out_of_range.any().item()):
        raise ValueError(f"{name} must be in [0, 1]")


def _validate_sigma(value: float, name: str) -> float:
    sigma = float(value)
    if not math.isfinite(sigma):
        raise ValueError(f"{name} must be finite")
    if sigma < 0.0:
        raise ValueError(f"{name} must be non-negative")
    return sigma


def _validate_timestep_sigma_pair(
    timestep: torch.Tensor,
    sigma_curr: float,

View on GitHub (pinned to 0132848349)

Solutions

1. Convert the discrete timestep to the unit scale the scheduler expects: t_unit = timestep.float() / num_train_timesteps (e.g. /1000), yielding a float in [0,1]
2. Or use the scheduler's own generated schedule (sigma/timestep pairs) instead of hand-built integer timesteps

Example fix

# before
x0 = minimax_h3_rf_v_to_x0(xt, v, torch.tensor(700))  # ValueError: not floating point

# after
t_unit = torch.tensor(700, dtype=torch.float32) / 1000.0  # 0.7
x0 = minimax_h3_rf_v_to_x0(xt, v, t_unit)

assert isinstance(timestep, torch.Tensor) and torch.is_floating_point(timestep)

def is_float_unit_timestep(t) -> bool:
    return isinstance(t, torch.Tensor) and torch.is_floating_point(t)

When it happens

Trigger: Passing timestep as an int/long tensor such as torch.tensor(700) or a timesteps schedule of dtype torch.int64 from a standard diffusion pipeline.

Common situations: Reusing a discrete diffusion timestep schedule (ints in [1,1000]) with the minimax_h3 normalized flow scheduler without converting to unit-scale floats.

Related errors

• {name} must be a torch.Tensor
• unsupported dtype for causal Conv3D cat/pad: {x.dtype}
• Validate failed: unsupported dtype: {t.dtype}
• {output_batch.error}
• action policy returned no output

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/d2ab428a2dfc2775. Report an issue: GitHub.