BrokenPipeError: [Errno 32] Broken pipe

我们用4张32G运行train_flux_lora.py训练FLUX lora时报错：
[2025-06-17 20:04:09,900] [INFO] [real_accelerator.py:254:get_accelerator] Setting ds_accelerator to cuda (auto detect)
[2025-06-17 20:04:10,569] [INFO] [real_accelerator.py:254:get_accelerator] Setting ds_accelerator to cuda (auto detect)
[2025-06-17 20:04:11,024] [INFO] [real_accelerator.py:254:get_accelerator] Setting ds_accelerator to cuda (auto detect)
[2025-06-17 20:04:11,812] [INFO] [logging.py:107:log_dist] [Rank -1] [TorchCheckpointEngine] Initialized with serialization = False
initializing deepspeed distributed: GLOBAL_RANK: 2, MEMBER: 3/4
[2025-06-17 20:04:12,101] [INFO] [logging.py:107:log_dist] [Rank -1] [TorchCheckpointEngine] Initialized with serialization = False
initializing deepspeed distributed: GLOBAL_RANK: 3, MEMBER: 4/4
[2025-06-17 20:04:12,494] [INFO] [logging.py:107:log_dist] [Rank -1] [TorchCheckpointEngine] Initialized with serialization = False
initializing deepspeed distributed: GLOBAL_RANK: 1, MEMBER: 2/4
Exception ignored in: <_io.BufferedWriter name=54>
BrokenPipeError: [Errno 32] Broken pipe
Segmentation fault (core dumped)

参数：CUDA_VISIBLE_DEVICES="0,1,2,3" python examples/train/flux/train_flux_lora.py
--pretrained_text_encoder_path models/FLUX/FLUX.1-dev/text_encoder/model.safetensors
--pretrained_text_encoder_2_path models/FLUX/FLUX.1-dev/text_encoder_2
--pretrained_dit_path models/FLUX/FLUX.1-dev/flux1-dev.safetensors
--pretrained_vae_path models/FLUX/FLUX.1-dev/ae.safetensors
--dataset_path data/cel-shading
--output_path ./models
--max_epochs 1
--steps_per_epoch 100
--height 1024
--width 1024
--center_crop
--precision "bf16"
--learning_rate 1e-4
--lora_rank 16
--lora_alpha 16
--use_gradient_checkpointing
--align_to_opensource_format
--training_strategy deepspeed_stage_2

Start with examples/train/flux/train_flux_lora.py and reproduce the command using CUDA_VISIBLE_DEVICES across four GPUs. Inspect the DeepSpeed distributed initialization and TorchCheckpointEngine messages around the BrokenPipeError and segmentation fault; done means multi-GPU training starts without either failure.