SimpleTuner
SimpleTuner copied to clipboard
Published
4 months ago
•
bghira
bghira
Sometimes training process dies with Signals.SIGKILL 9
After setting up the repo with the FLUX quickstart guide, I ran a training session overnight with my RTX 4090 to find that it had died somewhere along the way. It seems to have died after processing bucket 1.0 , console output:
(.venv) vwing@vwing-Desktop:~/Documents/SimpleTuner$ bash train.sh
2024-08-05 21:09:57,307 [WARNING] (ArgsParser) The VAE model madebyollin/sdxl-vae-fp16-fix is not compatible. Please use a compatible VAE to eliminate this warning. The baked-in VAE will be used, instead.
2024-08-05 21:09:57,307 [INFO] (ArgsParser) VAE Model: black-forest-labs/FLUX.1-dev
2024-08-05 21:09:57,307 [INFO] (ArgsParser) Default VAE Cache location:
2024-08-05 21:09:57,307 [INFO] (ArgsParser) Text Cache location: cache
2024-08-05 21:09:57,307 [WARNING] (ArgsParser) Updating T5 XXL tokeniser max length to 512 for Flux.
2024-08-05 21:09:57,307 [WARNING] (ArgsParser) Gradient accumulation steps are enabled, but gradient precision is set to 'unmodified'. This may lead to numeric instability. Consider setting --gradient_precision=fp32.
2024-08-05 21:09:57,365 [INFO] (__main__) Enabling tf32 precision boost for NVIDIA devices due to --allow_tf32.
2024-08-05 21:09:57,365 [INFO] (__main__) Load tokenizers
You set `add_prefix_space`. The tokenizer needs to be converted from the slow tokenizers
2024-08-05 21:09:57,958 [INFO] (__main__) Loading OpenAI CLIP-L text encoder from black-forest-labs/FLUX.1-dev/text_encoder..
2024-08-05 21:09:58,215 [INFO] (__main__) Loading T5 XXL v1.1 text encoder from black-forest-labs/FLUX.1-dev/text_encoder_2..
Downloading shards: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 3514.29it/s]
Loading checkpoint shards: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:03<00:00, 1.59s/it]
2024-08-05 21:10:02,078 [INFO] (__main__) Load VAE: black-forest-labs/FLUX.1-dev
2024-08-05 21:10:02,315 [INFO] (__main__) Moving text encoder to GPU.
2024-08-05 21:10:02,517 [INFO] (__main__) Moving text encoder 2 to GPU.
2024-08-05 21:10:04,141 [INFO] (__main__) Initialising VAE in bf16 precision, you may specify a different value if preferred: bf16, fp32, default
2024-08-05 21:10:04,189 [INFO] (__main__) Loaded VAE into VRAM.
2024-08-05 21:10:04,217 [INFO] (DataBackendFactory) Loading data backend config from config/multidatabackend.json
2024-08-05 21:10:04,218 [INFO] (DataBackendFactory) Configuring text embed backend: text-embeds
model_index.json: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 536/536 [00:00<00:00, 7.37MB/s]
Loading pipeline components...: 0%| | 0/5 [00:00<?, ?it/s]Loaded scheduler as FlowMatchEulerDiscreteScheduler from `scheduler` subfolder of black-forest-labs/FLUX.1-dev.
Loading pipeline components...: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 5/5 [00:00<00:00, 790.51it/s]
2024-08-05 21:10:04,394 [INFO] (TextEmbeddingCache) (Rank: 0) (id=text-embeds) Listing all text embed cache entries
2024-08-05 21:10:04,394 [INFO] (DataBackendFactory) Pre-computing null embedding
2024-08-05 21:10:09,692 [INFO] (DataBackendFactory) Completed loading text embed services.
2024-08-05 21:10:09,692 [INFO] (DataBackendFactory) Configuring data backend: pseudo-camera-10k-flux
2024-08-05 21:10:09,692 [INFO] (DataBackendFactory) (id=pseudo-camera-10k-flux) Loading bucket manager.
2024-08-05 21:10:09,693 [INFO] (JsonMetadataBackend) Checking for cache file: datasets/pseudo-camera-10k/aspect_ratio_bucket_indices.json
2024-08-05 21:10:09,693 [WARNING] (JsonMetadataBackend) No cache file found, creating new one.
2024-08-05 21:10:09,694 [INFO] (DataBackendFactory) (id=pseudo-camera-10k-flux) Refreshing aspect buckets on main process.
2024-08-05 21:10:09,694 [INFO] (BaseMetadataBackend) Discovering new files...
2024-08-05 21:10:13,224 [INFO] (BaseMetadataBackend) Compressed 0 existing files from 0.
2024-08-05 21:11:59,472 [INFO] (BaseMetadataBackend) Image processing statistics: {'total_processed': 14102, 'skipped': {'already_exists': 0, 'metadata_missing': 0, 'not_found': 0, 'too_small': 0, 'other': 0}}
2024-08-05 21:11:59,513 [INFO] (BaseMetadataBackend) Enforcing minimum image size of 0.5. This could take a while for very-large datasets.
2024-08-05 21:11:59,553 [INFO] (BaseMetadataBackend) Completed aspect bucket update.
2024-08-05 21:11:59,569 [INFO] (DataBackendFactory) Configured backend: {'id': 'pseudo-camera-10k-flux', 'config': {'crop': True, 'crop_aspect': 'square', 'crop_aspect_buckets': None, 'crop_style': 'center', 'disable_validation': False, 'resolution': 1.0, 'resolution_type': 'area', 'caption_strategy': 'filename', 'instance_data_dir': 'datasets/pseudo-camera-10k', 'maximum_image_size': 1.0, 'target_downsample_size': 1.0, 'config_version': 2}, 'dataset_type': 'image', 'data_backend': <helpers.data_backend.local.LocalDataBackend object at 0x7fcb8a4a7df0>, 'instance_data_dir': 'datasets/pseudo-camera-10k', 'metadata_backend': <helpers.metadata.backends.json.JsonMetadataBackend object at 0x7fcb8a4a7cd0>}
(Rank: 0) | Bucket | Image Count (per-GPU)
------------------------------
(Rank: 0) | 1.0 | 14080
2024-08-05 21:11:59,570 [INFO] (DataBackendFactory) (id=pseudo-camera-10k-flux) Collecting captions.
2024-08-05 21:11:59,622 [INFO] (DataBackendFactory) (id=pseudo-camera-10k-flux) Initialise text embed pre-computation using the filename caption strategy. We have 14102 captions to process.
2024-08-05 21:25:46,726 [INFO] (DataBackendFactory) (id=pseudo-camera-10k-flux) Completed processing 14102 captions.
2024-08-05 21:25:46,727 [INFO] (DataBackendFactory) (id=pseudo-camera-10k-flux) Creating VAE latent cache.
2024-08-05 21:25:46,834 [INFO] (DataBackendFactory) (id=pseudo-camera-10k-flux) Discovering cache objects..
Processing bucket 1.0: 4%|██ | 504/14080 [12:10<5:56:13, 1.57s/it]
Processing bucket 1.0: 5%|██▉ | 710/14080 [17:22<5:41:32, 1.53s/it]
(id=pseudo-camera-10k-flux) Bucket 1.0 caching results: {'not_local': 0, 'already_cached': 4056, 'cached': 200, 'total': 14080}
2024-08-06 01:46:57,256 [INFO] (DataBackendFactory) Configured backend: {'id': 'pseudo-camera-10k-flux', 'config': {'crop': True, 'crop_aspect': 'square', 'crop_aspect_buckets': None, 'crop_style': 'center', 'disable_validation': False, 'resolution': 1.0, 'resolution_type': 'area', 'caption_strategy': 'filename', 'instance_data_dir': 'datasets/pseudo-camera-10k', 'maximum_image_size': 1.0, 'target_downsample_size': 1.0, 'config_version': 2, 'hash_filenames': True}, 'dataset_type': 'image', 'data_backend': <helpers.data_backend.local.LocalDataBackend object at 0x7fcb8a4a7df0>, 'instance_data_dir': 'datasets/pseudo-camera-10k', 'metadata_backend': <helpers.metadata.backends.json.JsonMetadataBackend object at 0x7fcb8a4a7cd0>, 'train_dataset': <helpers.multiaspect.dataset.MultiAspectDataset object at 0x7fcb8a4a5e10>, 'sampler': <helpers.multiaspect.sampler.MultiAspectSampler object at 0x7fcb8a4a56f0>, 'train_dataloader': <torch.utils.data.dataloader.DataLoader object at 0x7fcb8a4a68f0>, 'text_embed_cache': <helpers.caching.text_embeds.TextEmbeddingCache object at 0x7fcb8ab1e2c0>, 'vaecache': <helpers.caching.vae.VAECache object at 0x7fcb8981fd30>}
2024-08-06 01:46:57,451 [INFO] (validation) Precomputing the negative prompt embed for validations.
2024-08-06 01:46:57,608 [INFO] (__main__) Unloading text encoders, as they are not being trained.
2024-08-06 01:46:58,565 [INFO] (__main__) After nuking text encoders from orbit, we freed 9.11 GB of VRAM. The real memories were the friends we trained a model on along the way.
Fetching 3 files: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 11135.32it/s]
Traceback (most recent call last):
File "/home/vwing/Documents/SimpleTuner/.venv/bin/accelerate", line 8, in <module>
sys.exit(main())
File "/home/vwing/Documents/SimpleTuner/.venv/lib/python3.10/site-packages/accelerate/commands/accelerate_cli.py", line 48, in main
args.func(args)
File "/home/vwing/Documents/SimpleTuner/.venv/lib/python3.10/site-packages/accelerate/commands/launch.py", line 1097, in launch_command
simple_launcher(args)
File "/home/vwing/Documents/SimpleTuner/.venv/lib/python3.10/site-packages/accelerate/commands/launch.py", line 703, in simple_launcher
raise subprocess.CalledProcessError(returncode=process.returncode, cmd=cmd)
subprocess.CalledProcessError: Command '['/home/vwing/Documents/SimpleTuner/.venv/bin/python', 'train.py', '--model_type=lora', '--pretrained_model_name_or_path=black-forest-labs/FLUX.1-dev', '--enable_xformers_memory_efficient_attention', '--gradient_checkpointing', '--set_grads_to_none', '--gradient_accumulation_steps=4', '--resume_from_checkpoint=latest', '--snr_gamma=5', '--data_backend_config=config/multidatabackend.json', '--num_train_epochs=0', '--max_train_steps=30000', '--metadata_update_interval=65', '--adam_bfloat16', '--learning_rate=8e-7', '--lr_scheduler=sine', '--seed', '42', '--lr_warmup_steps=1000', '--output_dir=output/models', '--inference_scheduler_timestep_spacing=trailing', '--training_scheduler_timestep_spacing=trailing', '--report_to=wandb', '--allow_tf32', '--mixed_precision=bf16', '--base_model_precision=int8-quanto', '--lora_rank=4', '--flux', '--train_batch=10', '--max_workers=32', '--read_batch_size=25', '--write_batch_size=64', '--caption_dropout_probability=0.1', '--torch_num_threads=8', '--image_processing_batch_size=32', '--vae_batch_size=12', '--validation_prompt=ethnographic photography of teddy bear at a picnic', '--num_validation_images=1', '--validation_num_inference_steps=30', '--validation_seed=42', '--minimum_image_size=1024', '--resolution=1024', '--validation_resolution=1024x1024', '--resolution_type=pixel', '--checkpointing_steps=150', '--checkpoints_total_limit=2', '--validation_steps=100', '--tracker_run_name=simpletuner-sdxl', '--tracker_project_name=sdxl-training', '--validation_guidance=3', '--validation_guidance_rescale=0.0', '--validation_negative_prompt=blurry, cropped, ugly']' died with <Signals.SIGKILL: 9>.
Does this error mean that the training ran out of VRAM / RAM along the way? I'm not sure what to make of the error. I'm using WSL. I believe I am running the bugfix/quanto-lora-loading branch. Pip freeze:
absl-py==2.1.0
accelerate==0.31.0
aiohttp==3.9.3
aiosignal==1.3.1
appdirs==1.4.4
async-timeout==4.0.3
attrs==23.2.0
bitsandbytes==0.42.0
boto3==1.34.79
botocore==1.34.79
build==1.2.1
CacheControl==0.14.0
certifi==2024.2.2
cffi==1.16.0
charset-normalizer==3.3.2
cleo==2.1.0
click==8.1.7
clip-interrogator==0.6.0
colorama==0.4.6
compel==2.0.2
crashtest==0.4.1
cryptography==43.0.0
dadaptation==3.2
datasets==2.14.4
deepspeed==0.10.3
diffusers @ git+https://github.com/huggingface/diffusers@15924bc73bfd74c769f23c8d2636d6c7514163a0
dill==0.3.7
distlib==0.3.8
docker-pycreds==0.4.0
dulwich==0.21.7
fastjsonschema==2.20.0
filelock==3.13.3
frozenlist==1.4.1
fsspec==2024.3.1
ftfy==6.2.0
gitdb==4.0.11
GitPython==3.1.43
grpcio==1.62.1
hjson==3.1.0
huggingface-hub==0.23.2
idna==3.6
importlib_metadata==7.1.0
installer==0.7.0
iterutils==0.1.6
jaraco.classes==3.4.0
jeepney==0.8.0
Jinja2==3.1.3
jmespath==1.0.1
keyring==24.3.1
lightning-utilities==0.11.2
Markdown==3.6
MarkupSafe==2.1.5
more-itertools==10.3.0
mpmath==1.3.0
msgpack==1.0.8
multidict==6.0.5
multiprocess==0.70.15
networkx==3.2.1
ninja==1.11.1.1
numpy==1.26.0
nvidia-cublas-cu12==12.1.3.1
nvidia-cuda-cupti-cu12==12.1.105
nvidia-cuda-nvrtc-cu12==12.1.105
nvidia-cuda-runtime-cu12==12.1.105
nvidia-cudnn-cu12==8.9.2.26
nvidia-cufft-cu12==11.0.2.54
nvidia-curand-cu12==10.3.2.106
nvidia-cusolver-cu12==11.4.5.107
nvidia-cusparse-cu12==12.1.0.106
nvidia-nccl-cu12==2.20.5
nvidia-nvjitlink-cu12==12.5.40
nvidia-nvtx-cu12==12.1.105
open-clip-torch==2.24.0
opencv-python==4.9.0.80
optimum-quanto==0.2.2
packaging==24.0
pandas==2.2.1
peft==0.9.0
pexpect==4.9.0
pillow==10.3.0
pkginfo==1.11.1
platformdirs==4.2.2
poetry==1.8.3
poetry-core==1.9.0
poetry-plugin-export==1.8.0
prodigyopt==1.0
protobuf==4.25.3
psutil==5.9.8
ptyprocess==0.7.0
py-cpuinfo==9.0.0
pyarrow==15.0.2
pycparser==2.22
pydantic==1.10.15
pyparsing==3.1.2
pyproject_hooks==1.1.0
python-dateutil==2.9.0.post0
pytz==2024.1
PyYAML==6.0.1
rapidfuzz==3.9.5
regex==2023.12.25
requests==2.31.0
requests-toolbelt==1.0.0
s3transfer==0.10.1
safetensors==0.4.2
scipy==1.13.0
SecretStorage==3.3.3
sentencepiece==0.2.0
sentry-sdk==1.44.1
setproctitle==1.3.3
shellingham==1.5.4
six==1.16.0
smmap==5.0.1
sympy==1.12
tensorboard==2.16.2
tensorboard-data-server==0.7.2
timm==0.9.16
tokenizers==0.19.1
tomli==2.0.1
tomlkit==0.13.0
torch==2.3.0+cu121
torchaudio==2.3.0+cu121
torchmetrics==1.3.2
torchsde==0.2.6
torchvision==0.18.0+cu121
tqdm==4.66.2
trampoline==0.1.2
transformers==4.42.4
triton==2.3.0
triton-library==1.0.0rc3
trove-classifiers==2024.7.2
typing_extensions==4.11.0
tzdata==2024.1
urllib3==1.26.18
virtualenv==20.26.3
wandb==0.16.6
wcwidth==0.2.13
Werkzeug==3.0.2
xformers==0.0.26.post1
xxhash==3.4.1
yarl==1.9.4
zipp==3.18.1
config.env :
# Configure these values.
# 'lora' or 'full'
# lora - train a small network for a character or style, or both. quite versatile.
# full - requires lots of vram, trains very slowly, needs a lot of data and concepts.
export MODEL_TYPE='lora'
# Set this to 'true' if you are training a Stable Diffusion 3 checkpoint.
# Use MODEL_NAME="stabilityai/stable-diffusion-3-medium-diffusers"
export STABLE_DIFFUSION_3=false
# Similarly, this is to train PixArt Sigma (1K or 2K) models.
# Use MODEL_NAME="PixArt-alpha/PixArt-Sigma-XL-2-1024-MS"
export PIXART_SIGMA=false
# For old Stable Diffusion 1.x/2.x models, you'll enable this.
# Use MODEL_NAME="stabilityai/stable-diffusion-2-1"
export STABLE_DIFFUSION_LEGACY=false
# For Kwai-Kolors, enable KOLORS.
# Use MODEL_NAME="kwai-kolors/kolors-diffusers"
export KOLORS=false
# For Flux, if you have 8 GPUs and DeepSpeed configured.
# Use MODEL_NAME="black-forest-labs/FLUX.1-dev"
export FLUX=true
# ControlNet model training is only supported when MODEL_TYPE='full'
# See this document for more information: https://github.com/bghira/SimpleTuner/blob/main/documentation/CONTROLNET.md
# DeepFloyd, PixArt, and SD3 do not currently support ControlNet model training.
export CONTROLNET=false
# DoRA enhances the training style of LoRA, but it will run more slowly at the same rank.
# See: https://arxiv.org/abs/2402.09353
# See: https://github.com/huggingface/peft/pull/1474
export USE_DORA=false
# BitFit freeze strategy for the u-net causes everything but the biases to be frozen.
# This may help retain the full model's underlying capabilities. LoRA is currently not tested/known to work.
#if [[ "$MODEL_TYPE" == "full" ]]; then
# # When training a full model, we will rely on BitFit to keep the u-net intact.
# export USE_BITFIT=true
#elif [[ "$MODEL_TYPE" == "lora" ]]; then
# # LoRA can not use BitFit.
# export USE_BITFIT=false
#elif [[ "$MODEL_TYPE" == "deepfloyd-full" ]]; then
# export USE_BITFIT=true
#fi
# Restart where we left off. Change this to "checkpoint-1234" to start from a specific checkpoint.
export RESUME_CHECKPOINT="latest"
# How often to checkpoint. Depending on your learning rate, you may wish to change this.
# For the default settings with 10 gradient accumulations, more frequent checkpoints might be preferable at first.
export CHECKPOINTING_STEPS=150
# This is how many checkpoints we will keep. Two is safe, but three is safer.
export CHECKPOINTING_LIMIT=2
# This is decided as a relatively conservative 'constant' learning rate.
# Adjust higher or lower depending on how burnt your model becomes.
export LEARNING_RATE=8e-7 #@param {type:"number"}
# Using a Huggingface Hub model:
export MODEL_NAME="black-forest-labs/FLUX.1-dev"
# Using a local path to a huggingface hub model or saved checkpoint:
#export MODEL_NAME="/datasets/models/pipeline"
# Make DEBUG_EXTRA_ARGS empty to disable wandb.
export DEBUG_EXTRA_ARGS="--report_to=wandb"
export TRACKER_PROJECT_NAME="sdxl-training"
export TRACKER_RUN_NAME="simpletuner-sdxl"
# Max number of steps OR epochs can be used. Not both.
export MAX_NUM_STEPS=30000
# Will likely overtrain, but that's fine.
export NUM_EPOCHS=0
# A convenient prefix for all of your training paths.
# These may be absolute or relative paths. Here, we are using relative paths.
# The output will just be in a folder called "output/models" by default.
export DATALOADER_CONFIG="config/multidatabackend.json"
export OUTPUT_DIR="output/models"
# Set this to "true" to push your model to Hugging Face Hub.
export PUSH_TO_HUB="false"
# If PUSH_TO_HUB and PUSH_CHECKPOINTS are both enabled, every saved checkpoint will be pushed to Hugging Face Hub.
export PUSH_CHECKPOINTS="true"
# This will be the model name for your final hub upload, eg. "yourusername/yourmodelname"
# It defaults to the wandb project name, but you can override this here.
export HUB_MODEL_NAME=$TRACKER_PROJECT_NAME
# By default, images will be resized so their SMALLER EDGE is 1024 pixels, maintaining aspect ratio.
# Setting this value to 768px might result in more reasonable training data sizes for SDXL.
export RESOLUTION=1024
# If you want to have the training data resized by pixel area (Megapixels) rather than edge length,
# set this value to "area" instead of "pixel", and uncomment the next RESOLUTION declaration.
export RESOLUTION_TYPE="pixel"
#export RESOLUTION=1 # 1.0 Megapixel training sizes
# If RESOLUTION_TYPE="pixel", the minimum resolution specifies the smaller edge length, measured in pixels. Recommended: 1024.
# If RESOLUTION_TYPE="area", the minimum resolution specifies the total image area, measured in megapixels. Recommended: 1.
export MINIMUM_RESOLUTION=$RESOLUTION
# How many decimals to round aspect buckets to.
#export ASPECT_BUCKET_ROUNDING=2
# Use this to append an instance prompt to each caption, used for adding trigger words.
# This has not been tested in SDXL.
#export INSTANCE_PROMPT="lotr style "
# If you also supply a user prompt library or `--use_prompt_library`, this will be added to those lists.
export VALIDATION_PROMPT="ethnographic photography of teddy bear at a picnic"
export VALIDATION_GUIDANCE=3
# You'll want to set this to 0.7 if you are training a terminal SNR model.
export VALIDATION_GUIDANCE_RESCALE=0.0
# How frequently we will save and run a pipeline for validations.
export VALIDATION_STEPS=100
export VALIDATION_NUM_INFERENCE_STEPS=30
export VALIDATION_NEGATIVE_PROMPT="blurry, cropped, ugly"
export VALIDATION_SEED=42
export VALIDATION_RESOLUTION=1024x1024
# Adjust this for your GPU memory size. This, and resolution, are the biggest VRAM killers.
export TRAIN_BATCH_SIZE=10
# Accumulate your update gradient over many steps, to save VRAM while still having higher effective batch size:
# effective batch size = ($TRAIN_BATCH_SIZE * $GRADIENT_ACCUMULATION_STEPS).
export GRADIENT_ACCUMULATION_STEPS=4
# Use any standard scheduler type. constant, polynomial, constant_with_warmup
export LR_SCHEDULE="sine"
# A warmup period allows the model and the EMA weights more importantly to familiarise itself with the current quanta.
# For the cosine or sine type schedules, the warmup period defines the interval between peaks or valleys.
# Use a sine schedule to simulate a warmup period, or a Cosine period to simulate a polynomial start.
#export LR_WARMUP_STEPS=$((MAX_NUM_STEPS / 10))
export LR_WARMUP_STEPS=1000
# Caption dropout probability. Set to 0.1 for 10% of captions dropped out. Set to 0 to disable.
# You may wish to disable dropout if you want to limit your changes strictly to the prompts you show the model.
# You may wish to increase the rate of dropout if you want to more broadly adopt your changes across the model.
export CAPTION_DROPOUT_PROBABILITY=0.1
export METADATA_UPDATE_INTERVAL=65
# How many workers to use for VAE caching.
export MAX_WORKERS=32
# Read and write batch sizes for VAE caching.
export READ_BATCH_SIZE=25
export WRITE_BATCH_SIZE=64
# How many images to encode at once with the VAE. Can increase VRAM use.
export VAE_BATCH_SIZE=12
# How many images to process at once (resize, crop, transform) during VAE caching.
export IMAGE_PROCESSING_BATCH_SIZE=32
# When using large batch sizes, you'll need to increase the pool connection limit.
export AWS_MAX_POOL_CONNECTIONS=128
# For very large systems, setting this can reduce CPU overhead of torch spawning an unnecessarily large number of threads.
export TORCH_NUM_THREADS=8
# If this is set, any images that fail to open will be DELETED to avoid re-checking them every time.
export DELETE_ERRORED_IMAGES=0
# If this is set, any images that are too small for the minimum resolution size will be DELETED.
export DELETE_SMALL_IMAGES=0
# Bytedance recommends these be set to "trailing" so that inference and training behave in a more congruent manner.
# To follow the original SDXL training strategy, use "leading" instead, though results are generally worse.
export TRAINING_SCHEDULER_TIMESTEP_SPACING="trailing"
export INFERENCE_SCHEDULER_TIMESTEP_SPACING="trailing"
# Removing this option or unsetting it uses vanilla training. Setting it reweights the loss by the position of the timestep in the noise schedule.
# A value "5" is recommended by the researchers. A value of "20" is the least impact, and "1" is the most impact.
export MIN_SNR_GAMMA=5
# Set this to an explicit value of "false" to disable Xformers. Probably required for AMD users.
export USE_XFORMERS=true
# There's basically no reason to unset this. However, to disable it, use an explicit value of "false".
# This will save a lot of memory consumption when enabled.
export USE_GRADIENT_CHECKPOINTING=true
##
# Options below here may require a bit more complicated configuration, so they are not simple variables.
##
# TF32 is great on Ampere or Ada, not sure about earlier generations.
export ALLOW_TF32=true
# AdamW 8Bit is a robust and lightweight choice. Adafactor might reduce memory consumption, and Dadaptation is slow and experimental.
# AdamW is the default optimizer, but it uses a lot of memory and is slower than AdamW8Bit or Adafactor.
# When training a quantised base model, you can't use adamw_bf16. Instead, try adafactor or adamw.
# Choices: adamw, adamw8bit, adafactor, dadaptation, adamw_bf16
export OPTIMIZER="adamw_bf16"
# EMA is a strong regularisation method that uses a lot of extra VRAM to hold two copies of the weights.
# This is worthwhile on large training runs, but not so much for smaller training runs.
export USE_EMA=false
export EMA_DECAY=0.999
export TRAINER_EXTRA_ARGS="--base_model_precision=int8-quanto --lora_rank=4"
## For offset noise training:
# Not recommended for terminal SNR models.
#export TRAINER_EXTRA_ARGS="${TRAINER_EXTRA_ARGS} --text_encoder_1_precision=no_change --text_encoder_2_precision=no_change"
## For terminal SNR training:
#export TRAINER_EXTRA_ARGS="${TRAINER_EXTRA_ARGS} --prediction_type=v_prediction --rescale_betas_zero_snr"
#export TRAINER_EXTRA_ARGS="${TRAINER_EXTRA_ARGS} --training_scheduler_timestep_spacing=trailing --inference_scheduler_timestep_spacing=trailing"
## You may benefit from directing training toward a specific weighted subset of timesteps.
# In this example, we train the final 25% of the timestep schedule with a 3x bias.
#export TRAINER_EXTRA_ARGS="${TRAINER_EXTRA_ARGS} --timestep_bias_strategy=later --timestep_bias_portion=0.25 --timestep_bias_multiplier=3"
# In this example, we train the earliest 25% of the timestep schedule with a 5x bias.
#export TRAINER_EXTRA_ARGS="${TRAINER_EXTRA_ARGS} --timestep_bias_strategy=earlier --timestep_bias_portion=0.25 --timestep_bias_multiplier=5"
# Here, we designate that specifically, timesteps 200 to 500 should be prioritised.
#export TRAINER_EXTRA_ARGS="${TRAINER_EXTRA_ARGS} --timestep_bias_strategy=range --timestep_bias_begin=200 --timestep_bias_end=500 --timestep_bias_multiplier=3"
## For experimental min-SNR weighted loss training (5 is suggested value by the original researchers):
# Not recommended for terminal SNR models.
#export TRAINER_EXTRA_ARGS="${TRAINER_EXTRA_ARGS} --snr_gamma=5.0"
# For Wasabi S3 filesystem backend (experimental)
#export TRAINER_EXTRA_ARGS="${TRAINER_EXTRA_ARGS} --data_backend=aws --aws_bucket_name=test123"
#export TRAINER_EXTRA_ARGS="${TRAINER_EXTRA_ARGS} --aws_endpoint_url=https://s3.wasabisys.com"
#export TRAINER_EXTRA_ARGS="${TRAINER_EXTRA_ARGS} --aws_access_key=1234567890"
#export TRAINER_EXTRA_ARGS="${TRAINER_EXTRA_ARGS} --aws_secret_access_key=0987654321"
# Reproducible training. Set to -1 to disable.
export TRAINING_SEED=42
# Mixed precision is the best. You honestly might need to YOLO it in fp16 mode for Google Colab type setups.
export MIXED_PRECISION="bf16" # Might not be supported on all GPUs. fp32 will be needed for others.
export PURE_BF16=true
# This has to be changed if you're training with multiple GPUs.
export TRAINING_NUM_PROCESSES=1
export TRAINING_NUM_MACHINES=1
export ACCELERATE_EXTRA_ARGS="" # --multi_gpu or other similar flags for huggingface accelerate
# With Pytorch 2.1, you might have pretty good luck here.
# If you're using aspect bucketing however, each resolution change will recompile. Seriously, just don't do it.
# Well, then again... Pytorch 2.2 has support for dynamic shapes. Why not?
export TRAINING_DYNAMO_BACKEND='no' # or 'no' if you want to disable torch compile in case of performance issues or lack of support (eg. AMD)
export TOKENIZERS_PARALLELISM=false
multidatabackend.json :
[
{
"id": "pseudo-camera-10k-flux",
"type": "local",
"crop": true,
"crop_aspect": "square",
"crop_style": "center",
"resolution": 1.0,
"minimum_image_size": 0.5,
"maximum_image_size": 1.0,
"target_downsample_size": 1.0,
"resolution_type": "area",
"cache_dir_vae": "cache/vae/flux/pseudo-camera-10k",
"instance_data_dir": "datasets/pseudo-camera-10k",
"disabled": false,
"skip_file_discovery": "",
"caption_strategy": "filename",
"metadata_backend": "json"
},
{
"id": "text-embeds",
"type": "local",
"dataset_type": "text_embeds",
"default": true,
"cache_dir": "cache/text/flux/pseudo-camera-10k",
"disabled": false,
"write_batch_size": 128
}
]
I'd appreciate any and all help getting this running. Thank you.
aw man :[
that means your system ran out of actual system memory and did not have enough swap space to continue trying to load the transformer.
just after the text encoders unload is when we pull in the transformer.
unfortunately, there's not much to do - i don't think i can even catch this error and show a better message to the user, because it has run out of memory and we cannot allocate anymore.
SIGKILL-9 is an outside force that became very angry at us (the Linux kernel).
aw man :[
that means your system ran out of actual system memory and did not have enough swap space to continue trying to load the transformer.
just after the text encoders unload is when we pull in the transformer.
unfortunately, there's not much to do - i don't think i can even catch this error and show a better message to the user, because it has run out of memory and we cannot allocate anymore.
SIGKILL-9 is an outside force that became very angry at us (the Linux kernel).
Thank you for the information and swift response, @bghira. Another question, if I may. I'm trying to train with 1x 4090 and I'm stilling maxing out my VRAM with 1020x1024, TRAIN_BATCH_SIZE=1, and adam_bf16 optimizer. Updated training config.env. Does this sound right? The FLUX quickstart guide mentions training with 20 GB of VRAM, but I have yet to see my usage fall below at least 23.5 GB of VRAM. Thank you for your time and assistance.
gradient accumulation steps have some memory overhead on their own, though i'm not sure it is that much
you also might not want to use gradient accumulation steps with bf16 because of accuracy issues in the accumulation stage
Oddly enough when it crashes with SIGKILL 9, neither my RAM or VRAM are maxed out. My VRAM utilization was at about 2 GB when it crashed. Picture captured shortly after it crashed: https://files.catbox.moe/5ddmlq.jpg
it might just not have shown it. can you add torch_dtype=torch.bfloat16 to line 776 of train.py?
from diffusers.models import FluxTransformer2DModel
transformer = FluxTransformer2DModel.from_pretrained(
args.pretrained_model_name_or_path,
subfolder="transformer", torch_dype=torch.bfloat16,
**pretrained_load_args,
)
alternatively you can try the main branch latest commits. i have put that in, so we load model directly into the correct precision
with recent changes in main branch i get constant SIGKILL. 731ccdf5 works though (with quanto, getting SIGKILL on 731ccdf5 too...)
Have a look at the output dmesg command. If it's saying something about OOM killer, you're running out of RAM.
Commit 3310c672d4c7b688373026a8faf0eba2ea267709 may have increased RAM usage if the text encoders remain in RAM during training. One idea would be to partially revert this commit so that text encoders are moved to the "meta" device.
Have a look at the output
dmesgcommand. If it's saying something about OOM killer, you're running out of RAM.
unfortunately can't run dmesg as I'm running in a container. dmesg: read kernel buffer failed: Operation not permitted
gradient accumulation steps have some memory overhead on their own, though i'm not sure it is that much
it actually is that much, I'm looking at my wandb and it takes this much vram (can't look at the actual usage not via wandb, don't ask why), and the training lasts 5h, however, as soon as I shift from batch 1 gr_a_steps 1 to gr_a_steps 2, it all goes bonkers and I get 9 hours of training
the line which gets SIGKILLed is the freezing of the model in _qunato_model . Specifically
freeze(model)
try grep oom /proc/vmstat which will show you how many oomkiller events have occurred without dmesg access
it needs a lot of memory to quantise
So this is a error which can be solved by installing more RAM in the computer? or it's VRAM related?
System RAM usage should now be several gigabytes lower now that pull request #694 has been merged.