Docker 镜像无法正常使用 xformers

```python
trainer | 18:56:52-195655 INFO Found 1 legal dataset
trainer | 18:56:52-214717 INFO Training started with config file /
trainer | 训练开始，使用配置文件:
trainer | /app/lora-scripts/config/autosave/20250414-185652.toml
trainer | 18:56:52-215951 INFO Using GPU(s) / 使用 GPU: ['0']
trainer | 18:56:52-216845 INFO Task f2a905f9-333d-4f27-90b7-221dbd89f316 created
trainer | [WARNING | __main__ ]: The following values were not passed to `accelerate launch` and had defaults used instead:
trainer | `--num_processes` was set to a value of `1`
trainer | `--num_machines` was set to a value of `1`
trainer | `--mixed_precision` was set to a value of `'no'`
trainer | `--dynamo_backend` was set to a value of `'no'`
trainer | To avoid this warning pass in values for each of the problematic parameters or run `accelerate config`.
trainer | [WARNING | xformers ]: WARNING[XFORMERS]: xFormers can't load C++/CUDA extensions. xFormers was built for:
trainer | PyTorch 2.4.0+cu121 with CUDA 1201 (you have 2.4.0a0+3bcc3cddb5.nv24.07)
trainer | Python 3.10.14 (you have 3.10.12)
trainer | Please reinstall xformers (see https://github.com/facebookresearch/xformers#installing-xformers)
trainer | Memory-efficient attention, SwiGLU, sparse and more won't be available.
trainer | Set XFORMERS_MORE_DETAILS=1 for more details
trainer | 2025-04-14 18:56:58 INFO prepare tokenizers sdxl_train_util.py:138
trainer | 2025-04-14 18:57:00 INFO update token length: 255 sdxl_train_util.py:163
trainer | INFO Using DreamBooth method. train_network.py:172
trainer | INFO prepare images. train_util.py:1686
trainer | INFO found directory train_util.py:1633
trainer | /app/lora-scripts/resources/reso
trainer | urces/images/03/1_zkz contains
trainer | 135 image files
trainer | 2025-04-14 18:57:01 INFO 135 train images with repeating. train_util.py:1727
trainer | INFO 0 reg images. train_util.py:1730
trainer | WARNING no regularization images / train_util.py:1735
trainer | 正則化画像が見つかりませんでした
trainer | INFO [Dataset 0] config_util.py:572
trainer | batch_size: 2
trainer | resolution: (1024, 1024)
trainer | enable_bucket: True
trainer | network_multiplier: 1.0
trainer | min_bucket_reso: 256
trainer | max_bucket_reso: 1024
trainer | bucket_reso_steps: 64
trainer | bucket_no_upscale: True
trainer |
trainer | [Subset 0 of Dataset 0]
trainer | image_dir:
trainer | "/app/lora-scripts/resources/res
trainer | ources/images/03/1_zkz"
trainer | image_count: 135
trainer | num_repeats: 1
trainer | shuffle_caption: False
trainer | keep_tokens: 0
trainer | keep_tokens_separator:
trainer | caption_separator: ,
trainer | secondary_separator: None
trainer | enable_wildcard: False
trainer | caption_dropout_rate: 0.0
trainer | caption_dropout_every_n_epoc
trainer | hes: 0
trainer | caption_tag_dropout_rate:
trainer | 0.0
trainer | caption_prefix: None
trainer | caption_suffix: None
trainer | color_aug: False
trainer | flip_aug: False
trainer | face_crop_aug_range: None
trainer | random_crop: False
trainer | token_warmup_min: 1,
trainer | token_warmup_step: 0,
trainer | alpha_mask: False,
trainer | is_reg: False
trainer | class_tokens: zkz
trainer | caption_extension: .txt
trainer |
trainer |
trainer | INFO [Dataset 0] config_util.py:578
trainer | INFO loading image sizes. train_util.py:901
100%|██████████| 135/135 [00:00<00:00, 498.63it/s]
trainer | 2025-04-14 18:57:02 INFO make buckets train_util.py:907
trainer | WARNING min_bucket_reso and train_util.py:924
trainer | max_bucket_reso are ignored if
trainer | bucket_no_upscale is set, because
trainer | bucket reso is defined by image
trainer | size automatically /
trainer | bucket_no_upscaleが指定された場合
trainer | は、bucketの解像度は画像サイズか
trainer | ら自動計算されるため、min_bucket_
trainer | resoとmax_bucket_resoは無視されま
trainer | す
trainer | INFO number of images (including train_util.py:953
trainer | repeats) /
trainer | 各bucketの画像枚数（繰り返し回数
trainer | を含む）
trainer | INFO bucket 0: resolution (1024, train_util.py:958
trainer | 1024), count: 16
trainer | INFO bucket 1: resolution (1088, 832), train_util.py:958
trainer | count: 119
trainer | INFO mean ar error (without repeats): train_util.py:963
trainer | 0.017536103743000367
trainer | INFO preparing accelerator train_network.py:225
trainer | accelerator device: cuda
trainer | INFO loading model for process 0/1 sdxl_train_util.py:33
trainer | INFO load StableDiffusion sdxl_train_util.py:74
trainer | checkpoint:
trainer | /app/lora-scripts/models/chec
trainer | kpoints/SDXL/sdxlNijiSeven_sd
trainer | xlNijiSeven.safetensors
trainer | 2025-04-14 18:57:09 INFO building U-Net sdxl_model_util.py:198
trainer | INFO loading U-Net from sdxl_model_util.py:202
trainer | checkpoint
trainer | 2025-04-14 18:58:28 INFO U-Net:
trainer | INFO building text encoders sdxl_model_util.py:211
trainer | INFO loading text encoders from sdxl_model_util.py:264
trainer | checkpoint
trainer | 2025-04-14 18:58:32 INFO text encoder 1:
trainer | 2025-04-14 18:58:55 INFO text encoder 2:
trainer | INFO building VAE sdxl_model_util.py:285
trainer | INFO loading VAE from checkpoint sdxl_model_util.py:290
trainer | INFO VAE:
trainer | INFO Enable xformers for U-Net train_util.py:2832
trainer | Traceback (most recent call last):
trainer | File "/app/lora-scripts/./scripts/stable/sdxl_train_network.py", line 184, in
trainer | trainer.train(args)
trainer | File "/app/lora-scripts/scripts/stable/train_network.py", line 242, in train
trainer | vae.set_use_memory_efficient_attention_xformers(args.xformers)
trainer | File "/usr/local/lib/python3.10/dist-packages/diffusers/models/modeling_utils.py", line 262, in set_use_memory_efficient_attention_xformers
trainer | fn_recursive_set_mem_eff(module)
trainer | File "/usr/local/lib/python3.10/dist-packages/diffusers/models/modeling_utils.py", line 258, in fn_recursive_set_mem_eff
trainer | fn_recursive_set_mem_eff(child)
trainer | File "/usr/local/lib/python3.10/dist-packages/diffusers/models/modeling_utils.py", line 258, in fn_recursive_set_mem_eff
trainer | fn_recursive_set_mem_eff(child)
trainer | File "/usr/local/lib/python3.10/dist-packages/diffusers/models/modeling_utils.py", line 258, in fn_recursive_set_mem_eff
trainer | fn_recursive_set_mem_eff(child)
trainer | File "/usr/local/lib/python3.10/dist-packages/diffusers/models/modeling_utils.py", line 255, in fn_recursive_set_mem_eff
trainer | module.set_use_memory_efficient_attention_xformers(valid, attention_op)
trainer | File "/usr/local/lib/python3.10/dist-packages/diffusers/models/attention_processor.py", line 273, in set_use_memory_efficient_attention_xformers
trainer | raise e
trainer | File "/usr/local/lib/python3.10/dist-packages/diffusers/models/attention_processor.py", line 267, in set_use_memory_efficient_attention_xformers
trainer | _ = xformers.ops.memory_efficient_attention(
trainer | File "/usr/local/lib/python3.10/dist-packages/xformers/ops/fmha/__init__.py", line 276, in memory_efficient_attention
trainer | return _memory_efficient_attention(
trainer | File "/usr/local/lib/python3.10/dist-packages/xformers/ops/fmha/__init__.py", line 395, in _memory_efficient_attention
trainer | return _memory_efficient_attention_forward(
trainer | File "/usr/local/lib/python3.10/dist-packages/xformers/ops/fmha/__init__.py", line 414, in _memory_efficient_attention_forward
trainer | op = _dispatch_fw(inp, False)
trainer | File "/usr/local/lib/python3.10/dist-packages/xformers/ops/fmha/dispatch.py", line 119, in _dispatch_fw
trainer | return _run_priority_list(
trainer | File "/usr/local/lib/python3.10/dist-packages/xformers/ops/fmha/dispatch.py", line 55, in _run_priority_list
trainer | raise NotImplementedError(msg)
trainer | NotImplementedError: No operator found for `memory_efficient_attention_forward` with inputs:
trainer | query : shape=(1, 2, 1, 40) (torch.float32)
trainer | key : shape=(1, 2, 1, 40) (torch.float32)
trainer | value : shape=(1, 2, 1, 40) (torch.float32)
trainer | attn_bias :
trainer | p : 0.0
trainer | `decoderF` is not supported because:
trainer | xFormers wasn't build with CUDA support
trainer | attn_bias type is
trainer | operator wasn't built - see `python -m xformers.info` for more info
trainer | `flshattF@2.5.6-pt` is not supported because:
trainer | xFormers wasn't build with CUDA support
trainer | dtype=torch.float32 (supported: {torch.float16, torch.bfloat16})
trainer | `cutlassF` is not supported because:
trainer | xFormers wasn't build with CUDA support
trainer | operator wasn't built - see `python -m xformers.info` for more info
trainer | `smallkF` is not supported because:
trainer | max(query.shape[-1] != value.shape[-1]) > 32
trainer | xFormers wasn't build with CUDA support
trainer | operator wasn't built - see `python -m xformers.info` for more info
trainer | unsupported embed per head: 40
trainer | 18:58:57-701463 ERROR Training failed / 训练失败
```