No SDP backend available, likely because you are running in

Q: What causes "No SDP backend available, likely because you are running in pytorch versions < 2.0. In fact, you are using PyTorch {torch.__version__}. You might want to consider upgrading." in Stability-AI/generative-models?

Upgrade PyTorch to >= 2.0: pip install --upgrade torch

Error message

No SDP backend available, likely because you are running in pytorch versions < 2.0. In fact, you are using PyTorch {torch.__version__}. You might want to consider upgrading.

What it means

attention.py imports torch.nn.functional.sdp_kernel / scaled_dot_product_attention guarded by a version check. If torch.nn does not expose SDP (PyTorch < 2.0), it logs this warning, sets SDP_IS_AVAILABLE=False, and substitutes a nullcontext — meaning attention runs through slower fallback paths.

Source

Thrown at sgm/modules/attention.py:43

        SDPBackend.FLASH_ATTENTION: {
            "enable_math": False,
            "enable_flash": True,
            "enable_mem_efficient": False,
        },
        SDPBackend.EFFICIENT_ATTENTION: {
            "enable_math": False,
            "enable_flash": False,
            "enable_mem_efficient": True,
        },
        None: {"enable_math": True, "enable_flash": True, "enable_mem_efficient": True},
    }
else:
    from contextlib import nullcontext

    SDP_IS_AVAILABLE = False
    sdp_kernel = nullcontext
    BACKEND_MAP = {}
    logpy.warn(
        f"No SDP backend available, likely because you are running in pytorch "
        f"versions < 2.0. In fact, you are using PyTorch {torch.__version__}. "
        f"You might want to consider upgrading."
    )

try:
    import xformers
    import xformers.ops

    XFORMERS_IS_AVAILABLE = True
except:
    XFORMERS_IS_AVAILABLE = False
    logpy.warn("no module 'xformers'. Processing without...")

# from .diffusionmodules.util import mixed_checkpoint as checkpoint


def exists(val):

View on GitHub (pinned to e8cd657656)

Solutions

Upgrade PyTorch to >= 2.0: pip install --upgrade torch
If you must stay on 1.x, accept the warning and ensure xformers is installed so attention still runs efficiently
Pin your environment (requirements.txt) to torch>=2.0 for this codebase

Example fix

// before
torch 1.13.1
// after
pip install "torch>=2.0.0"

Defensive patterns

Strategy: validation

Validate before calling

import torch
assert tuple(map(int, torch.__version__.split("+")[0].split(".")[:2])) >= (2, 0), \
    f"need torch>=2.0, got {torch.__version__}"

Type guard

def sdp_available() -> bool:
    return hasattr(torch.nn.functional, "scaled_dot_product_attention")

Try / catch

try:
    import sgm.modules.attention  # may warn at import
except Exception:
    pass
if not sdp_available():
    print("install torch>=2.0 for SDP attention")

Prevention

Pin torch>=2.0 in requirements
Log torch.__version__ at startup
Smoke-test imports in CI on the target torch build

When it happens

Trigger: Importing sgm.modules.attention with PyTorch < 2.0 installed (or a build without SDP support); the message is emitted at import time and includes the detected torch.__version__.

Common situations: Old environments pinned to torch 1.13 or earlier; CPU-only or custom torch builds lacking SDP kernels; forgetting to upgrade torch after cloning newer generative-models code.

Related errors

AI-assisted analysis of Stability-AI/generative-models@e8cd657656 (2026-08-29). Data as JSON: /api/errors/941465fb9839f16a. Report an issue: GitHub.