sgl-project/sglang · error · ValueError

{name} must be finite

Error message

{name} must be finite

What it means

The minimax_h3 Euler ancestral scheduler validates that every element of xt/v/timestep tensors is finite before doing math. NaN or Inf anywhere in the tensor raises this ValueError, guarding against silently propagating NaNs through the flow-matching update.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/schedulers/scheduling_minimax_h3_euler_ancestral.py:12

# 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:

View on GitHub (pinned to 0132848349)

Solutions

1. Inspect model outputs each step with torch.isfinite(...).all() to find the first step that produces NaN/Inf
2. Reduce guidance_scale (or eta/churn settings) that cause denoising divergence
3. Run the model in fp32/bf16 instead of fp16 to avoid overflow
4. Verify checkpoint and embeddings load cleanly (no NaNs at t=0)

Example fix

# before
x0 = minimax_h3_rf_v_to_x0(xt, v, timestep)  # ValueError: v must be finite

# after
if not torch.isfinite(v).all():
    v = torch.nan_to_num(v, nan=0.0, posinf=0.0, neginf=0.0)
x0 = minimax_h3_rf_v_to_x0(xt, v, timestep)

assert torch.isfinite(xt).all() and torch.isfinite(v).all(), "non-finite inputs to scheduler"

def all_finite(*ts: torch.Tensor) -> bool:
    return all(bool(torch.isfinite(t).all().item()) for t in ts)

try:
    x0 = minimax_h3_rf_v_to_x0(xt, v, t)
except ValueError as e:
    if "must be finite" in str(e):
        v = torch.nan_to_num(v)
        x0 = minimax_h3_rf_v_to_x0(xt, v, t)
    else:
        raise

When it happens

Trigger: Passing a noisy sample xt, a model velocity output v, or a timestep tensor containing NaN/Inf to minimax_h3_rf_v_to_x0, minimax_h3_euler_eta0_step, or the validators. Usually the NaN originates from the model forward pass (diverged training-free CFG, fp16 overflow, bad guidance scale).

Common situations: fp16/bf16 numerical overflow in the denoising UNet/DiT producing NaN velocities; extremely high guidance_scale or eta causing divergence; corrupted checkpoints or NaN-inducing embeddings.

Related errors

• denoising_strength must be positive
• [pred_noise_to_pred_video] Invalid timestep shape: {timestep
• Validate failed: unsupported dtype: {t.dtype}
• Validate failed: unsupported tensor shape: {t.shape}.
• {output_batch.error}

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/188a29f34528103f. Report an issue: GitHub.