'NoneType' object has no attribute 'gradient_checkpointing_enable'

[pzzmyc]

I used the --train_text_encoder parameter while training SDXL lora, got an error AttributeError: 'NoneType' object has no attribute 'gradient_checkpointing_enable'.full args is here: args:snr_gamma:5.0 args:use_soft_min_snr:False args:soft_min_snr_sigma_data:None args:model_type:lora args:pixart_sigma:False args:sd3:False args:sd3_uses_diffusion:False args:weighting_scheme:logit_normal args:logit_mean:0.0 args:logit_std:1.0 args:mode_scale:1.29 args:lora_type:None args:lora_init_type:default args:lora_rank:16 args:lora_alpha:16 args:lora_dropout:0.1 args:controlnet:False args:controlnet_model_name_or_path:None args:pretrained_model_name_or_path:/media/hhy/A983D940A4049592/base/stable-diffusion-xl-base-1.0 args:pretrained_vae_model_name_or_path:madebyollin/sdxl-vae-fp16-fix args:pretrained_t5_model_name_or_path:None args:prediction_type:epsilon args:snr_weight:1.0 args:training_scheduler_timestep_spacing:trailing args:inference_scheduler_timestep_spacing:trailing args:refiner_training:False args:refiner_training_invert_schedule:False args:refiner_training_strength:0.2 args:timestep_bias_strategy:none args:timestep_bias_multiplier:1.0 args:timestep_bias_begin:0 args:timestep_bias_end:1000 args:timestep_bias_portion:0.25 args:disable_segmented_timestep_sampling:False args:rescale_betas_zero_snr:False args:vae_dtype:bf16 args:vae_batch_size:4 args:vae_cache_scan_behaviour:recreate args:vae_cache_preprocess:False args:aspect_bucket_disable_rebuild:False args:keep_vae_loaded:False args:skip_file_discovery: args:revision:None args:variant:None args:preserve_data_backend_cache:False args:use_dora:False args:override_dataset_config:False args:cache_dir_text:cache args:cache_dir_vae: args:data_backend_config:/media/hhy/A983D940A4049592/base/dataset.json args:write_batch_size:64 args:enable_multiprocessing:False args:aspect_bucket_worker_count:12 args:cache_dir:/media/hhy/A983D940A4049592/base/savedmodels/cache args:cache_clear_validation_prompts:False args:caption_strategy:filename args:parquet_caption_column:None args:parquet_filename_column:None args:instance_prompt:None args:output_dir:/media/hhy/A983D940A4049592/base/savedmodels args:seed:42 args:seed_for_each_device:True args:resolution:1024.0 args:resolution_type:pixel args:aspect_bucket_rounding:None args:aspect_bucket_alignment:64 args:minimum_image_size:1024.0 args:maximum_image_size:None args:target_downsample_size:None args:train_text_encoder:True args:tokenizer_max_length:None args:train_batch_size:1 args:num_train_epochs:10 args:max_train_steps:None args:checkpointing_steps:1000 args:checkpoints_total_limit:2 args:resume_from_checkpoint:latest args:gradient_accumulation_steps:4 args:gradient_checkpointing:True args:learning_rate:1e-06 args:text_encoder_lr:1e-06 args:lr_scale:False args:lr_scheduler:constant args:lr_warmup_steps:1000 args:lr_num_cycles:1 args:lr_power:0.8 args:use_ema:False args:ema_decay:0.995 args:non_ema_revision:None args:offload_param_path:None args:use_8bit_adam:False args:use_adafactor_optimizer:False args:adafactor_relative_step:False args:use_prodigy_optimizer:False args:prodigy_beta3:None args:prodigy_decouple:True args:prodigy_use_bias_correction:True args:prodigy_safeguard_warmup:True args:prodigy_learning_rate:0.5 args:prodigy_weight_decay:0.01 args:prodigy_epsilon:1e-08 args:use_dadapt_optimizer:False args:dadaptation_learning_rate:1.0 args:adam_beta1:0.9 args:adam_beta2:0.999 args:adam_weight_decay:0.01 args:adam_epsilon:1e-08 args:adam_bfloat16:True args:max_grad_norm:2.0 args:push_to_hub:False args:push_checkpoints_to_hub:False args:hub_model_id:None args:logging_dir:logs args:validation_torch_compile:False args:validation_torch_compile_mode:max-autotune args:allow_tf32:True args:validation_using_datasets:None args:webhook_config:None args:report_to:wandb args:tracker_run_name:simpletuner-sdxl args:tracker_project_name:sdxl-training args:validation_prompt:ethnographic photography of teddy bear at a picnic args:validation_prompt_library:False args:user_prompt_library:None args:validation_negative_prompt:blurry, cropped, ugly args:num_validation_images:1 args:validation_steps:10000000000000000000 args:num_eval_images:4 args:eval_dataset_id:None args:validation_num_inference_steps:30 args:validation_resolution:1024 args:validation_noise_scheduler:euler args:validation_disable_unconditional:False args:disable_compel:False args:enable_watermark:False args:mixed_precision:bf16 args:local_rank:0 args:enable_xformers_memory_efficient_attention:True args:set_grads_to_none:True args:noise_offset:0.1 args:noise_offset_probability:0.25 args:validation_guidance:7.5 args:validation_guidance_rescale:0.0 args:validation_randomize:False args:validation_seed:42 args:fully_unload_text_encoder:False args:freeze_encoder_before:12 args:freeze_encoder_after:17 args:freeze_encoder_strategy:after args:freeze_unet_strategy:none args:unet_attention_slice:False args:print_filenames:False args:print_sampler_statistics:False args:metadata_update_interval:65 args:debug_aspect_buckets:False args:debug_dataset_loader:False args:freeze_encoder:True args:save_text_encoder:True args:text_encoder_limit:25 args:prepend_instance_prompt:False args:only_instance_prompt:False args:data_aesthetic_score:7.0 args:sdxl_refiner_uses_full_range:False args:caption_dropout_probability:0.1 args:delete_unwanted_images:False args:delete_problematic_images:False args:offset_noise:False args:lr_end:4e-7 args:i_know_what_i_am_doing:False

[bghira]

the text encoder training really isn't tested much, and i don't have much time to maintain that feature. if you'd like to submit a pull request to resolve the issue, i can review it for inclusion

[bghira]

in fact this one looks like it might just have been from adding stable diffusion 3 support. still, for any further issues it might take a while to resolve if they relate to eg. model quality issues.

[pzzmyc]

Pull request creation failed. Validation failed: must be a collaborator