Flux2 klein

examples/dreambooth/README_flux2.md

@@ -1,14 +1,22 @@
-# DreamBooth training example for FLUX.2 [dev]
+# DreamBooth training example for FLUX.2 [dev] and FLUX 2 [klein]
 
 [DreamBooth](https://huggingface.co/papers/2208.12242) is a method to personalize image generation models given just a few (3~5) images of a subject/concept.
+[LoRA](https://huggingface.co/docs/peft/conceptual_guides/adapter#low-rank-adaptation-lora) is a popular parameter-efficient fine-tuning technique that allows you to achieve full-finetuning like performance but with a fraction of learnable parameters.
+
+The `train_dreambooth_lora_flux2.py`, `train_dreambooth_lora_flux2_klein.py` scripts shows how to implement the training procedure for [LoRAs](https://huggingface.co/blog/lora) and adapt it for [FLUX.2 [dev]](https://huggingface.co/black-forest-labs/FLUX.2-dev) and [FLUX 2 [klein]](https://huggingface.co/black-forest-labs/FLUX.2-klein).
 
-The `train_dreambooth_lora_flux2.py` script shows how to implement the training procedure for [LoRAs](https://huggingface.co/blog/lora) and adapt it for [FLUX.2 [dev]](https://github.com/black-forest-labs/flux2).
+> [!NOTE]
+> **Model Variants**
+>
+> We support two FLUX model families:
+> - **FLUX.2 [dev]**: The full-size model using Mistral Small 3.1 as the text encoder. Very capable but memory intensive.
+> - **FLUX 2 [klein]**: Available in 4B and 9B parameter variants, using Qwen VL as the text encoder. Much more memory efficient and suitable for consumer hardware.
 
 > [!NOTE]
 > **Memory consumption**
 >
-> Flux can be quite expensive to run on consumer hardware devices and as a result finetuning it comes with high memory requirements -
-> a LoRA with a rank of 16 can exceed XXGB of VRAM for training. below we provide some tips and tricks to reduce memory consumption during training.
+> FLUX.2 [dev] can be quite expensive to run on consumer hardware devices and as a result finetuning it comes with high memory requirements -
+> a LoRA with a rank of 16 can exceed XXGB of VRAM for training. FLUX 2 [klein] models (4B and 9B) are significantly more memory efficient alternatives. Below we provide some tips and tricks to reduce memory consumption during training.
 
 > For more tips & guidance on training on a resource-constrained device and general good practices please check out these great guides and trainers for FLUX: 
 > 1) [`@bghira`'s guide](https://github.com/bghira/SimpleTuner/blob/main/documentation/quickstart/FLUX2.md)
@@ -17,7 +25,7 @@ The `train_dreambooth_lora_flux2.py` script shows how to implement the training
 > [!NOTE]
 > **Gated model**
 >
-> As the model is gated, before using it with diffusers you first need to go to the [FLUX.2 [dev] Hugging Face page](https://huggingface.co/black-forest-labs/FLUX.2-dev), fill in the form and accept the gate. Once you are in, you need to log in so that your system knows you’ve accepted the gate. Use the command below to log in:
+> As the model is gated, before using it with diffusers you first need to go to the [FLUX.2 [dev] Hugging Face page](https://huggingface.co/black-forest-labs/FLUX.2-dev), fill in the form and accept the gate. Once you are in, you need to log in so that your system knows you've accepted the gate. Use the command below to log in:
 
 ```bash
 hf auth login
@@ -88,23 +96,32 @@ snapshot_download(
 
 This will also allow us to push the trained LoRA parameters to the Hugging Face Hub platform.
 
-As mentioned, Flux2 LoRA training is *very* memory intensive. Here are memory optimizations we can use (some still experimental) for a more memory efficient training:
+As mentioned, Flux2 LoRA training is *very* memory intensive (especially for FLUX.2 [dev]). Here are memory optimizations we can use (some still experimental) for a more memory efficient training:
 
 ## Memory Optimizations
 > [!NOTE] many of these techniques complement each other and can be used together to further reduce memory consumption. 
 > However some techniques may be mutually exclusive so be sure to check before launching a training run.
+
 ### Remote Text Encoder 
-Flux.2 uses  Mistral Small 3.1 as text encoder which is quite large and can take up a lot of memory. To mitigate this, we can use the `--remote_text_encoder` flag to enable remote computation of the prompt embeddings using the HuggingFace Inference API. 
+FLUX.2 [dev] uses Mistral Small 3.1 as text encoder which is quite large and can take up a lot of memory. To mitigate this, we can use the `--remote_text_encoder` flag to enable remote computation of the prompt embeddings using the HuggingFace Inference API. 
 This way, the text encoder model is not loaded into memory during training.
+
+> [!IMPORTANT]
+> **Remote text encoder is only supported for FLUX.2 [dev]**. FLUX 2 [klein] models use the Qwen VL text encoder and do not support remote text encoding.
+
 > [!NOTE] 
 > to enable remote text encoding you must either be logged in to your HuggingFace account (`hf auth login`) OR pass a token with `--hub_token`.
+
 ### FSDP Text Encoder 
-Flux.2 uses  Mistral Small 3.1 as text encoder which is quite large and can take up a lot of memory. To mitigate this, we can use the `--fsdp_text_encoder` flag to enable distributed computation of the prompt embeddings. 
+FLUX.2 [dev] uses Mistral Small 3.1 as text encoder which is quite large and can take up a lot of memory. To mitigate this, we can use the `--fsdp_text_encoder` flag to enable distributed computation of the prompt embeddings. 
 This way, it distributes the memory cost across multiple nodes.
+
 ### CPU Offloading 
 To offload parts of the model to CPU memory, you can use `--offload` flag. This will offload the vae and text encoder to CPU memory and only move them to GPU when needed.
+
 ### Latent Caching 
 Pre-encode the training images with the vae, and then delete it to free up some memory. To enable `latent_caching` simply pass `--cache_latents`.
+
 ### QLoRA: Low Precision Training with Quantization
 Perform low precision training using 8-bit or 4-bit quantization to reduce memory usage. You can use the following flags:
 - **FP8 training** with `torchao`: 
@@ -114,22 +131,29 @@ enable FP8 training by passing `--do_fp8_training`.
 - **NF4 training** with `bitsandbytes`: 
 Alternatively, you can use 8-bit or 4-bit quantization with `bitsandbytes` by passing:
 `--bnb_quantization_config_path` to enable 4-bit NF4 quantization.
+
 ### Gradient Checkpointing and Accumulation
 * `--gradient accumulation` refers to the number of updates steps to accumulate before performing a backward/update pass.
 by passing a value > 1 you can reduce the amount of backward/update passes and hence also memory reqs.
 * with `--gradient checkpointing` we can save memory by not storing all intermediate activations during the forward pass.
 Instead, only a subset of these activations (the checkpoints) are stored and the rest is recomputed as needed during the backward pass. Note that this comes at the expanse of a slower backward pass.
+
 ### 8-bit-Adam Optimizer
 When training with `AdamW`(doesn't apply to `prodigy`) You can pass `--use_8bit_adam` to reduce the memory requirements of training. 
 Make sure to install `bitsandbytes` if you want to do so.
+
 ### Image Resolution
 An easy way to mitigate some of the memory requirements is through `--resolution`. `--resolution` refers to the resolution for input images, all the images in the train/validation dataset are resized to this.
 Note that by default, images are resized to resolution of 512, but it's good to keep in mind in case you're accustomed to training on higher resolutions.
+
 ### Precision of saved LoRA layers
 By default, trained transformer layers are saved in the precision dtype in which training was performed. E.g. when training in mixed precision is enabled with `--mixed_precision="bf16"`, final finetuned layers will be saved in `torch.bfloat16` as well. 
 This reduces memory requirements significantly w/o a significant quality loss. Note that if you do wish to save the final layers in float32 at the expanse of more memory usage, you can do so by passing `--upcast_before_saving`.
 
+## Training Examples
 
+### FLUX.2 [dev] Training
+To perform DreamBooth with LoRA on FLUX.2 [dev], run:
 ```bash
 export MODEL_NAME="black-forest-labs/FLUX.2-dev"
 export INSTANCE_DIR="dog"
@@ -161,13 +185,84 @@ accelerate launch train_dreambooth_lora_flux2.py \
   --push_to_hub
 ```
 
+### FLUX 2 [klein] Training
+
+FLUX 2 [klein] models are more memory efficient alternatives available in 4B and 9B parameter variants. They use the Qwen VL text encoder instead of Mistral Small 3.1.
+
+> [!NOTE]
+> The `--remote_text_encoder` flag is **not supported** for FLUX 2 [klein] models. The Qwen VL text encoder must be loaded locally, but offloading is still supported.
+
+**FLUX 2 [klein] 4B:**
+
+```bash
+export MODEL_NAME="black-forest-labs/FLUX.2-klein-4B"
+export INSTANCE_DIR="dog"
+export OUTPUT_DIR="trained-flux2-klein-4b"
+
+accelerate launch train_dreambooth_lora_flux2_klein.py \
+  --pretrained_model_name_or_path=$MODEL_NAME  \
+  --instance_data_dir=$INSTANCE_DIR \
+  --output_dir=$OUTPUT_DIR \
+  --do_fp8_training \
+  --gradient_checkpointing \
+  --cache_latents \
+  --instance_prompt="a photo of sks dog" \
+  --resolution=1024 \
+  --train_batch_size=1 \
+  --guidance_scale=1 \
+  --use_8bit_adam \
+  --gradient_accumulation_steps=4 \
+  --optimizer="adamW" \
+  --learning_rate=1e-4 \
+  --report_to="wandb" \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=100 \
+  --max_train_steps=500 \
+  --validation_prompt="A photo of sks dog in a bucket" \
+  --validation_epochs=25 \
+  --seed="0" \
+  --push_to_hub
+```
+
+**FLUX 2 [klein] 9B:**
+
+```bash
+export MODEL_NAME="black-forest-labs/FLUX.2-klein-9B"
+export INSTANCE_DIR="dog"
+export OUTPUT_DIR="trained-flux2-klein-9b"
+
+accelerate launch train_dreambooth_lora_flux2_klein.py \
+  --pretrained_model_name_or_path=$MODEL_NAME  \
+  --instance_data_dir=$INSTANCE_DIR \
+  --output_dir=$OUTPUT_DIR \
+  --do_fp8_training \
+  --gradient_checkpointing \
+  --cache_latents \
+  --instance_prompt="a photo of sks dog" \
+  --resolution=1024 \
+  --train_batch_size=1 \
+  --guidance_scale=1 \
+  --use_8bit_adam \
+  --gradient_accumulation_steps=4 \
+  --optimizer="adamW" \
+  --learning_rate=1e-4 \
+  --report_to="wandb" \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=100 \
+  --max_train_steps=500 \
+  --validation_prompt="A photo of sks dog in a bucket" \
+  --validation_epochs=25 \
+  --seed="0" \
+  --push_to_hub
+```
+
 To better track our training experiments, we're using the following flags in the command above:
 
 * `report_to="wandb` will ensure the training runs are tracked on [Weights and Biases](https://wandb.ai/site). To use it, be sure to install `wandb` with `pip install wandb`. Don't forget to call `wandb login <your_api_key>` before training if you haven't done it before.
 * `validation_prompt` and `validation_epochs` to allow the script to do a few validation inference runs. This allows us to qualitatively check if the training is progressing as expected.
 
 > [!NOTE]
-> If you want to train using long prompts with the T5 text encoder, you can use `--max_sequence_length` to set the token limit. The default is 77, but it can be increased to as high as 512. Note that this will use more resources and may slow down the training in some cases.
+> If you want to train using long prompts, you can use `--max_sequence_length` to set the token limit. Note that this will use more resources and may slow down the training in some cases.
 
 ### FSDP on the transformer
 By setting the accelerate configuration with FSDP, the transformer block will be wrapped automatically. E.g. set the configuration to:
@@ -189,12 +284,6 @@ fsdp_config:
   fsdp_cpu_ram_efficient_loading: false
 ```
 
-## LoRA + DreamBooth
-
-[LoRA](https://huggingface.co/docs/peft/conceptual_guides/adapter#low-rank-adaptation-lora) is a popular parameter-efficient fine-tuning technique that allows you to achieve full-finetuning like performance but with a fraction of learnable parameters.
-
-Note also that we use PEFT library as backend for LoRA training, make sure to have `peft>=0.6.0` installed in your environment.
-
 ### Prodigy Optimizer
 Prodigy is an adaptive optimizer that dynamically adjusts the learning rate learned parameters based on past gradients, allowing for more efficient convergence. 
 By using prodigy we can "eliminate" the need for manual learning rate tuning. read more [here](https://huggingface.co/blog/sdxl_lora_advanced_script#adaptive-optimizers).
@@ -206,8 +295,6 @@ to use prodigy, first make sure to install the prodigyopt library: `pip install
 > [!TIP]
 > When using prodigy it's generally good practice to set- `--learning_rate=1.0`
 
-To perform DreamBooth with LoRA, run:
-
 ```bash
 export MODEL_NAME="black-forest-labs/FLUX.2-dev"
 export INSTANCE_DIR="dog"
@@ -271,13 +358,10 @@ the exact modules for LoRA training. Here are some examples of target modules yo
 > keep in mind that while training more layers can improve quality and expressiveness, it also increases the size of the output LoRA weights.
 
 
-
 ## Training Image-to-Image
 
 Flux.2 lets us perform image editing as well as image generation. We provide a simple script for image-to-image(I2I) LoRA fine-tuning in [train_dreambooth_lora_flux2_img2img.py](./train_dreambooth_lora_flux2_img2img.py) for both T2I and I2I. The optimizations discussed above apply this script, too.
 
-**important**
-
 **Important**
 To make sure you can successfully run the latest version of the image-to-image example script, we highly recommend installing from source, specifically from the commit mentioned below. To do this, execute the following steps in a new virtual environment:
 
@@ -334,5 +418,6 @@ we've added aspect ratio bucketing support which allows training on images with
 To enable aspect ratio bucketing, pass `--aspect_ratio_buckets` argument with a semicolon-separated list of height,width pairs, such as:
 
 `--aspect_ratio_buckets="672,1568;688,1504;720,1456;752,1392;800,1328;832,1248;880,1184;944,1104;1024,1024;1104,944;1184,880;1248,832;1328,800;1392,752;1456,720;1504,688;1568,672"
-`
-Since Flux.2 finetuning is still an experimental phase, we encourage you to explore different settings and share your insights! 🤗
+
+
+Since Flux.2 finetuning is still an experimental phase, we encourage you to explore different settings and share your insights! 🤗