Does SimpleTuner support quantization of the T5-XXL encoder for the SD 3.5 models?

[NirmalF7]

basically just the title, have tried setting it in the config.json file with text encoder precision but keep running into errors. Thanks for any help.

[bghira]

yes, but you'll have to elaborate on the errors received. some precision levels do not work, eg. fp8, but int8 will.

[NirmalF7]

These are the configs { "--model_type": "lora", "--model_family": "sd3", "--checkpointing_steps": 100, "--checkpoints_total_limit": 60, "--learning_rate": 3e-06, "--pretrained_model_name_or_path": "stabilityai/stable-diffusion-3.5-medium", "--max_train_steps": 100, "--num_train_epochs": 0, "--push_to_hub": false, "--push_checkpoints_to_hub": false, "--hub_model_id": "sd35-training", "--resolution": 1024, "--resolution_type": "pixel", "--minimum_image_size": 1024, "--instance_prompt": "abcx", "--validation_prompt": "abcx, car on motorway", "--validation_guidance": 7.5, "--validation_guidance_rescale": 0.0, "--validation_steps": 200, "--validation_num_inference_steps": 30, "--validation_negative_prompt": "blurry, cropped, ugly, unrealistic", "--validation_seed": 42, "--validation_resolution": "1024x1024", "--train_batch_size": 1, "--gradient_accumulation_steps": 1, "--lr_scheduler": "cosine", "--lr_warmup_steps": 20, "--caption_dropout_probability": 0.1, "--metadata_update_interval": 65, "--vae_batch_size": 1, "--delete_unwanted_images": false, "--delete_problematic_images": false, "--training_scheduler_timestep_spacing": "trailing", "--inference_scheduler_timestep_spacing": "trailing", "--snr_gamma": 5, "--enable_xformers_memory_efficient_attention": false, "--gradient_checkpointing": true, "--allow_tf32": true, "--optimizer": "adamw_bf16", "--use_ema": false, "--ema_decay": 0.999, "--seed": 42, "--mixed_precision": "bf16", "--lora_rank": 128, "--lora_alpha": 128, "--lora_type": "standard", "--base_model_precision": "int8-quanto", "--text_encoder_1_precision": "int8-quanto", "--text_encoder_2_precision": "int8-quanto", "--text_encoder_3_precision": "int8-quanto", "--quantize_via":"cpu", "--validation_disable_unconditional":true, "--attention_mechanism": "diffusers"

which lead to this error

RuntimeError: Given groups=1, weight of size [1536, 16, 2, 2], expected input[1, 4, 96, 168] to have 16 channels, but got 4 channels instead

[bghira]

i think CLIP models can't be quantised at all

[NirmalF7]

yep that would make sense. Weirdly enough I am getting the error even with "--base_model_precision": "int8-quanto", "--text_encoder_1_precision": "no_change", "--text_encoder_2_precision": "no_change", "--text_encoder_3_precision": "no_change",
First step of training runs fine then errors at the beginning of the second step.

[bghira]

sounds like the issue is actually that your VAE cache probably contains SDXL entries

[NirmalF7]

This and only quantizing the T5 encoder worked thank you!