open-mmlab/mmdetection · error · NotImplementedError

If you want to reduce GPU memory usage,

Error message

If you want to reduce GPU memory usage,                     please install fairscale by executing the                     following command: pip install fairscale.

What it means

Raised when DetrTransformerEncoder has num_cp > 0 but fairscale is not installed. The encoder layers are wrapped with fairscale's checkpoint_wrapper to trade compute for GPU memory; without the package this is impossible.

Source

Thrown at mmdet/models/layers/transformer/detr_layers.py:54

                 init_cfg: OptConfigType = None) -> None:

        super().__init__(init_cfg=init_cfg)
        self.num_layers = num_layers
        self.layer_cfg = layer_cfg
        self.num_cp = num_cp
        assert self.num_cp <= self.num_layers
        self._init_layers()

    def _init_layers(self) -> None:
        """Initialize encoder layers."""
        self.layers = ModuleList([
            DetrTransformerEncoderLayer(**self.layer_cfg)
            for _ in range(self.num_layers)
        ])

        if self.num_cp > 0:
            if checkpoint_wrapper is None:
                raise NotImplementedError(
                    'If you want to reduce GPU memory usage, \
                    please install fairscale by executing the \
                    following command: pip install fairscale.')
            for i in range(self.num_cp):
                self.layers[i] = checkpoint_wrapper(self.layers[i])

        self.embed_dims = self.layers[0].embed_dims

    def forward(self, query: Tensor, query_pos: Tensor,
                key_padding_mask: Tensor, **kwargs) -> Tensor:
        """Forward function of encoder.

        Args:
            query (Tensor): Input queries of encoder, has shape
                (bs, num_queries, dim).
            query_pos (Tensor): The positional embeddings of the queries, has
                shape (bs, num_queries, dim).
            key_padding_mask (Tensor): The `key_padding_mask` of `self_attn`

View on GitHub (pinned to cfd5d3a985)

Solutions

1. pip install fairscale
2. Or remove/zero num_cp in the encoder config

Example fix

# before
num_cp=6, fairscale not installed
# after
pip install fairscale  # or set num_cp=0

if cfg.get('num_cp', 0) > 0:
    import fairscale  # raises ImportError early with a clear message

When it happens

Trigger: Config with with_cp=True or num_cp>0 on a DETR transformer encoder while fairscale is absent from the environment.

Common situations: Using memory-saving configs on machines without fairscale; upgrading mmdet versions where the checkpointing path changed.

Related errors

• If you want to reduce GPU memory usage,
• If you want to reduce GPU memory usage,
• imagecorruptions is not installed
• albumentations is not installed
• DETR do not build sampler.

AI-assisted analysis of open-mmlab/mmdetection@cfd5d3a985 (2026-08-27). Data as JSON: /api/errors/82172860e387b3a4. Report an issue: GitHub.