sgl-project/sglang · error · ValueError

Cannot call `set_default_attn_processor` when attention proc

Error message

Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}

What it means

set_default_attn_processor picks the default processor class based on the current processors' class family (attention vs cross-attention). If the installed processors are neither in the known attention-processor set nor the cross-attention set (e.g. a fused or custom processor), it refuses to guess which default to restore.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/vaes/autoencoder.py:270

            fn_recursive_attn_processor(name, module, processor)

    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
    def set_default_attn_processor(self):
        """
        Disables custom attention processors and sets the default attention implementation.
        """
        if all(
            proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS
            for proc in self.attn_processors.values()
        ):
            processor = AttnAddedKVProcessor()
        elif all(
            proc.__class__ in CROSS_ATTENTION_PROCESSORS
            for proc in self.attn_processors.values()
        ):
            processor = AttnProcessor()
        else:
            raise ValueError(
                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
            )

        self.set_attn_processor(processor)

    def _encode(self, x: torch.Tensor) -> torch.Tensor:
        batch_size, num_channels, height, width = x.shape

        if self.use_tiling and (
            width > self.tile_sample_min_size or height > self.tile_sample_min_size
        ):
            return self._tiled_encode(x)

        enc = self.encoder(x)
        if self.quant_conv is not None:
            enc = self.quant_conv(enc)

        return enc

View on GitHub (pinned to 0132848349)

Solutions

1. Explicitly restore with set_attn_processor(self.original_attn_processors) (saved by fuse_qkv_projections) instead of set_default_attn_processor
2. Or explicitly set the exact processor class: model.set_attn_processor(AttnProcessor())
3. Upgrade/align the library version so the processor classes you installed are recognized

Example fix

# before
model.fuse_qkv_projections()
...
model.set_default_attn_processor()
# after
model.fuse_qkv_projections()
...
model.unfuse_qkv_projections()
# or: model.set_attn_processor(model.original_attn_processors)

known = (AttnProcessor,)
if not all(type(p) in known_set for p in model.attn_processors.values()):
    model.set_attn_processor(AttnProcessor())  # explicit reset instead of default

try:
    model.set_default_attn_processor()
except ValueError:
    model.set_attn_processor(AttnProcessor())

When it happens

Trigger: Calling set_default_attn_processor() after fuse_qkv_projections() installed fused processors, or after setting a custom/XFormers/Flash processor class not registered in either processor-class set.

Common situations: Trying to undo fusion or a custom processor by 'resetting to default'; using a processor from a newer diffusers version whose class isn't in the local CROSS_ATTENTION_PROCESSORS set; diffusers version change altering processor class hierarchy.

Related errors

• Cannot call `set_default_attn_processor` when attention proc
• A dict of processors was passed, but the number of processor
• `fuse_qkv_projections()` is not supported for models having
• A dict of processors was passed, but the number of processor
• num_heads must be divisible by num_epi_subtiles

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/762913339ecbe34a. Report an issue: GitHub.