Closed JEFFSEVENTHSENSE closed 1 month ago
kohya-ss
commented
1 month ago This error occurs in AE, so schnell doesn't seem to be related. It could be that the bucket_reso_steps value in the dataset config is too small or the max_bucket_reso value is too large. Could you please share the .toml of the dataset config?
JEFFSEVENTHSENSE
commented
1 month ago INFO [Dataset 0] config_util.py:570
batch_size: 4
resolution: (256, 256)
enable_bucket: False
network_multiplier: 1.0
[Subset 0 of Dataset 0]
image_dir: "/home/dluser/development/Jeff/test"
image_count: 11
num_repeats: 1
shuffle_caption: False
keep_tokens: 2
keep_tokens_separator:
caption_separator: ,
secondary_separator: None
enable_wildcard: False
caption_dropout_rate: 0.0
caption_dropout_every_n_epoches: 0
caption_tag_dropout_rate: 0.0
caption_prefix: None
caption_suffix: None
color_aug: False
flip_aug: False
face_crop_aug_range: None
random_crop: False
token_warmup_min: 1,
token_warmup_step: 0,
alpha_mask: False,
is_reg: False
class_tokens: J3FF
caption_extension: .txt I am currently running the script with | NVIDIA-SMI 510.47.03 Driver Version: 510.47.03 CUDA Version: 11.6 | torch 1.13.1+cu116 torchaudio 0.13.1+cu116 torchmetrics 1.4.2 torchvision 0.14.1+cu116
The above settings cant be changed due to some reasons
kohya-ss
commented
1 month ago resolution: (256, 256) seems to be too small for FLUX.1. Could you please try 512,512?
JEFFSEVENTHSENSE
commented
1 month ago thanks for the reply , i have re clone the whole repository and is currently only facing issues as mentioned in https://github.com/kohya-ss/sd-scripts/issues/1679
CUDA_VISIBLE_DEVICES=7 accelerate launch --mixed_precision bf16 --num_cpu_threads_per_process 1 flux_train_network.py --pretrained_model_name_or_path flux1-schnell.safetensors --clip_l sd3/clip_l.safetensors --t5xxl sd3/t5xxl_fp16.safetensors --ae ae.safetensors --save_model_as safetensors --sdpa --persistent_data_loader_workers --max_data_loader_n_workers 2 --seed 42 --gradient_checkpointing --mixed_precision bf16 --save_precision bf16 --network_module networks.lora_flux --network_dim 4 --optimizer_type adamw8bit --learning_rate 1e-4 --cache_text_encoder_outputs --highvram --max_train_epochs 4 --save_every_n_epochs 1 --dataset_config dataset_1024_bs2.toml --output_dir /home/dluser/development/Jeff/LoRA --output_name flux-lora-jeff --timestep_sampling shift --discrete_flow_shift 3.1582 --model_prediction_type raw --guidance_scale 1.0
this is the script i am running on but i am facing issue at this point when it is training:
File "/home/sd-scripts/train_network.py", line 1096, in train latents = self.encode_images_to_latents(args, accelerator, vae, batch["images"].to(vae_dtype)) File "/home/sd-scripts/flux_train_network.py", line 326, in encode_images_to_latents return vae.encode(images) File "/home/sd-scripts/library/flux_models.py", line 347, in encode z = self.reg(self.encoder(x)) File "/home/.virtualenvs/flux/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1194, in _call_impl return forward_call(*input, *kwargs) File "/home/sd-scripts/library/flux_models.py", line 210, in forward h = self.mid.attn_1(h) File "/home/.virtualenvs/flux/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1194, in _call_impl return forward_call(input, **kwargs) File "/home/sd-scripts/library/flux_models.py", line 88, in forward return x + self.proj_out(self.attention(x)) File "/home/sd-scripts/library/flux_models.py", line 83, in attention
RuntimeError: Expected (head_size % 8 == 0) && (head_size <= 128) to be true, but got false. (Could this error message be improved? If so, please report an enhancement request to PyTorch.)
This error occurs in