Add support for JPEG-XL (file extension: jxl)#1413

Merged

bghira merged 115 commits intobghira:mainfrom

StableLlama:jpeg_xl

Apr 18, 2025

Commits on Apr 17, 2025

Add support for JPEG-XL (file extension: jxl)
StableLlama
committed
add HiDream support

bghira
authored and
StableLlama
committed
better import protection for hidream

bghira
authored and
StableLlama
committed
fix hidream default path

bghira
authored and
StableLlama
committed
gets closer to loading a full pipeline properly

bghira
authored and
StableLlama
committed
extend text encoder quant to include number 4

bghira
authored and
StableLlama
committed
load pipeline using custom tokenisers too

bghira
authored and
StableLlama
committed
loading and embed generation fixes

bghira
authored and
StableLlama
committed
fix typo to 128 tokens

bghira
authored and
StableLlama
committed
allow quant for text encoder 4

bghira
authored and
StableLlama
committed
update torchao, transformers, hf hub

bghira
authored and
StableLlama
committed
reformat

bghira
authored and
StableLlama
committed
add helpers for moving model components around and smarter pipeline return with caching

bghira
authored and
StableLlama
committed
use text2img pipeline for encoding prompts

bghira
authored and
StableLlama
committed
use text2img for encode prompt with hidream

bghira
authored and
StableLlama
committed
format unipc schedule

bghira
authored and
StableLlama
committed
remove flash_attn fallback code for test

bghira
authored and
StableLlama
committed
reformat

bghira
authored and
StableLlama
committed
follow new method for retrieving pipeline

bghira
authored and
StableLlama
committed
update HiDream transformer for newer diffusers gradient checkpointing

bghira
authored and
StableLlama
committed
use helper method to print embed shapes for convenience

bghira
authored and
StableLlama
committed
update base image model class to have a prompt embed collator hook

bghira
authored and
StableLlama
committed
update hidream model class for better t5 vs llama embed handling, clarity

bghira
authored and
StableLlama
committed
update hidream transformer class to properly handle 32 layer bulk inputs

bghira
authored and
StableLlama
committed
add handling for model collator hook in prompt embeds

bghira
authored and
StableLlama
committed
add helper methods for retrieving embed shapes or moving them

bghira
authored and
StableLlama
committed
reorganise hidream transformer class for clarity, move closer to the separated t5/llama embed handling

bghira
authored and
StableLlama
committed
fix unpatchify, disable it for training

bghira
authored and
StableLlama
committed
update pipeline conversion

bghira
authored and
StableLlama
committed
update pipeline signatures for split embeds

bghira
authored and
StableLlama
committed
move embeds to GPU when reformatting inputs

bghira
authored and
StableLlama
committed
reduce gradient tracking in MoEGate

bghira
authored and
StableLlama
committed
MoEGate might benefit from gradient checkpointing instead

bghira
authored and
StableLlama
committed
Rename HiDreamAttention to Attention, for lycoris config convenience

bghira
authored and
StableLlama
committed
FeedForwardSwiGLU -> FeedForward for lycoris config convenience

bghira
authored and
StableLlama
committed
Fix for Attention block rename

bghira
authored and
StableLlama
committed
Revert "FeedForwardSwiGLU -> FeedForward for lycoris config convenience"

bghira
authored and
StableLlama
committed
offload embeds to CPU while formatting

bghira
authored and
StableLlama
committed
Fix a bug where text encoder will not quantise when the base model is not

bghira
authored and
StableLlama
committed
allow HiDream and other validations to fail and not choke on missing eval scores

bghira
authored and
StableLlama
committed
HiDream validations meta tensor error fix

bghira
authored and
StableLlama
committed
use unsloth ungated llama repo in alignment with diffusers example

bghira
authored and
StableLlama
committed
use common fork of llama with correct configs since unsloth has wrong tokenizer class

bghira
authored and
StableLlama
committed
silence transformers behavioural change warning

bghira
authored and
StableLlama
committed
add load balancing loss options for hidream MoEGate

bghira
authored and
StableLlama
committed
add auxiliary loss hook for models and a fix for pipeline not having valid VAE when the time comes for validations later

bghira
authored and
StableLlama
committed
add auxiliary loss handler for hidream

bghira
authored and
StableLlama
committed
pipeline fix for vae latent device location/dtype

bghira
authored and
StableLlama
committed
hidream should pass aux loss init alpha value along through the modules

bghira
authored and
StableLlama
committed
gather and log auxiliary loss values from a model during training

bghira
authored and
StableLlama
committed
style

bghira
authored and
StableLlama
committed
update hidream docs with latest info

bghira
authored and
StableLlama
committed
add hidream to readme

bghira
authored and
StableLlama
committed
resolve error with non-square inputs

bghira
authored and
StableLlama
committed
fix bug with pixart sigma pipeline typo

bghira
authored and
StableLlama
committed
fix differential parent-student loss

bghira
authored and
StableLlama
committed
better error logging when debugging

bghira
authored and
StableLlama
committed
HiDream SDPA support, no more flash_attn required

bghira
authored and
StableLlama
committed
fix publishing of models to the hub

bghira
authored and
StableLlama
committed
default to hidream-full model instead of dev

bghira
authored and
StableLlama
committed
tests go brr

bghira
authored and
StableLlama
committed
fix feedforward training

bghira
authored and
StableLlama
committed
update hidream doc

bghira
authored and
StableLlama
committed
update tail

bghira
authored and
StableLlama
committed
hidream configure script enablement

bghira
authored and
StableLlama
committed
Fix HiDream training for train_batch_size > 1.

mhirki
authored and
StableLlama
committed
Fix parent-student training broken by v2 refactor.

mhirki
authored and
StableLlama
committed
disable fp16 just in case

bghira
authored and
StableLlama
committed
add auraflow to configure script

bghira
authored and
StableLlama
committed
add auraflow to all models list

bghira
authored and
StableLlama
committed
fix single text encoder model loading

bghira
authored and
StableLlama
committed
fix errors and update sd3 model code for consistency

bghira
authored and
StableLlama
committed
add auraflow support

bghira
authored and
StableLlama
committed
update docs

bghira
authored and
StableLlama
committed
fixes for ckpt-only validations when EMA is enabled, and, token length limit application for sd3/auraflow/omnigen fix

bghira
authored and
StableLlama
committed
fix for multigpu training

bghira
authored and
StableLlama
committed
integrate PE torch compile fix from AstraliteHeart

bghira
authored and
StableLlama
committed
enable regularisation loss training on PEFT Standard LoRA

bghira
authored and
StableLlama
committed
model class should always unwrap it for DDP

bghira
authored and
StableLlama
committed
add missing seq len

bghira
authored and
StableLlama
committed
(#1393) enable flux_lora_target for v2 lora loading code in flux

bghira
authored and
StableLlama
committed
when init controlnet datasets, sync their sizes to the base dataset. add some text encoder moving logic that is disabled for now.

bghira
authored and
StableLlama
committed
formatting

bghira
authored and
StableLlama
committed
fixes for init and retrieval of controlnet vs base model

bghira
authored and
StableLlama
committed
formatting

bghira
authored and
StableLlama
committed
controlnet revival for SD1x

bghira
authored and
StableLlama
committed
formatting

bghira
authored and
StableLlama
committed
add state tracker method for retrieving conditioning dataset mapping directly

bghira
authored and
StableLlama
committed
controlnet trainer fixes for loading controlnet model by abstraction layer

bghira
authored and
StableLlama
committed
fix controlnet validation image gathering resizing everything to 64px

bghira
authored and
StableLlama
committed
comment out temp code

bghira
authored and
StableLlama
committed
make sdxl controlnet training work again

bghira
authored and
StableLlama
committed
split_buckets_between_processes: Fix issue with imbalanced steps between processes

clayne
authored and
StableLlama
committed
HiDream: fix pipeline load so that it uses cached value, unlike other models

bghira
authored and
StableLlama
committed
fix sana gradient checkpointing code for intervals

bghira
authored and
StableLlama
committed
remove diffusers utils logspam

bghira
authored and
StableLlama
committed
fix OOM for vae caching on ltxvideo and wan video on 24g cards

bghira
authored and
StableLlama
committed
ltxvideo tokeniser shoudl use AutoTokenizer to avoid class mismatches

bghira
authored and
StableLlama
committed
wan: fix print of model name at startup

bghira
authored and
StableLlama
committed
wandb: for now, only log first video frame. also add more helpful backtrace print for logging errors

bghira
authored and
StableLlama
committed
apple: update diffusers

bghira
authored and
StableLlama
committed
Added check for UniPCMultistepScheduler

ShuyUSTC
authored and
StableLlama
committed
Fix publishing model card for None caption dropout amount

bghira
authored and
StableLlama
committed
fix setting default seq len for sd3

bghira
authored and
StableLlama
committed
update diffusers to latest git main

bghira
authored and
StableLlama
committed
Rename FeedForwardSwiGLU back to FeedForward for convenience.

mhirki
authored and
StableLlama
committed
Add support for JPEG-XL (file extension: jxl)
StableLlama
committed
Fix poetry.lock
StableLlama
committed
Add support for JPEG-XL (file extension: jxl)
StableLlama
committed
Add support for JPEG-XL (file extension: jxl)
StableLlama
committed
Fix poetry.lock
StableLlama
committed
Merge remote-tracking branch 'origin/jpeg_xl' into jpeg_xl
StableLlama
committed
Fix poetry.lock
StableLlama
committed
Handle CV2 code paths as well and fall back to PIL when CV2 doesn't understand an image codec
StableLlama
committed

Commits on Apr 18, 2025

Make `pillow-jxl-plugin` and thus JPEG XL optional.
StableLlama
committed