Stability-AI/generative-models · error

We do not support vanilla attention anymore, as it is too ex

Error message

We do not support vanilla attention anymore, as it is too expensive. Sorry.

What it means

When attn_mode is 'softmax' but PyTorch SDP (SDP_IS_AVAILABLE False, torch < 2.0) is unavailable, the code warns that vanilla (naive) attention is no longer supported because it is too expensive, and then asserts False — terminating with AssertionError if xformers is also missing (the assert's message directs to installing xformers).

Source

Thrown at sgm/modules/attention.py:486

        context_dim=None,
        gated_ff=True,
        checkpoint=True,
        disable_self_attn=False,
        attn_mode="softmax",
        sdp_backend=None,
    ):
        super().__init__()
        assert attn_mode in self.ATTENTION_MODES
        if attn_mode != "softmax" and not XFORMERS_IS_AVAILABLE:
            logpy.warn(
                f"Attention mode '{attn_mode}' is not available. Falling "
                f"back to native attention. This is not a problem in "
                f"Pytorch >= 2.0. FYI, you are running with PyTorch "
                f"version {torch.__version__}."
            )
            attn_mode = "softmax"
        elif attn_mode == "softmax" and not SDP_IS_AVAILABLE:
            logpy.warn(
                "We do not support vanilla attention anymore, as it is too "
                "expensive. Sorry."
            )
            if not XFORMERS_IS_AVAILABLE:
                assert (
                    False
                ), "Please install xformers via e.g. 'pip install xformers==0.0.16'"
            else:
                logpy.info("Falling back to xformers efficient attention.")
                attn_mode = "softmax-xformers"
        attn_cls = self.ATTENTION_MODES[attn_mode]
        if version.parse(torch.__version__) >= version.parse("2.0.0"):
            assert sdp_backend is None or isinstance(sdp_backend, SDPBackend)
        else:
            assert sdp_backend is None
        self.disable_self_attn = disable_self_attn
        self.attn1 = attn_cls(
            query_dim=dim,

View on GitHub (pinned to e8cd657656)

Solutions

1. Upgrade PyTorch to >= 2.0 so SDP attention is available
2. Install xformers so the assert's fallback (memory-efficient attention) succeeds: pip install xformers
3. If neither is possible, patch attention.py to allow a naive-attention implementation (not recommended — very slow/high memory)

Example fix

// before
pip list  # torch 1.13.0, no xformers -> AssertionError in attention.py
// after
pip install "torch>=2.0" xformers

import torch
sdp = hasattr(torch.nn.functional, "scaled_dot_product_attention")
try:
    import xformers.ops
    xf = True
except ImportError:
    xf = False
assert sdp or xf, "need torch>=2.0 SDP or xformers before building the model"

def any_attention_backend() -> bool:
    import torch
    if hasattr(torch.nn.functional, "scaled_dot_product_attention"):
        return True
    try:
        import xformers.ops
        return True
    except ImportError:
        return False

try:
    model = instantiate_from_config(config)
except AssertionError:
    raise RuntimeError("No attention backend: install torch>=2.0 or xformers")

When it happens

Trigger: Running with PyTorch < 2.0 (no SDP) AND xformers not installed, while instantiating attention with the default attn_mode='softmax'.

Common situations: Legacy torch 1.x environments without xformers attempting to run the SD model; stripped-down deployments lacking both backends; CPU-only images where xformers is hard to install.

Related errors

• Attention mode '{attn_mode}' is not available. Falling back
• input has {x.ndim} dims but target_dims is {target_dims}, wh
• Did not find parameters for pattern {pattern_}
• No SDP backend available, likely because you are running in
• no module 'xformers'. Processing without...

AI-assisted analysis of Stability-AI/generative-models@e8cd657656 (2026-08-29). Data as JSON: /api/errors/99cfba35c45548ad. Report an issue: GitHub.