Improve lora implementation

lora unloading initial

Merge pull request #14 from flowpoint/rtx4000ada_bench

add benchmarks numbers for rtx4000ada (non-sff)

Update README.md, Fix license link

Improved precision / reduced frequency of nan outputs, allow bf16 t5, f32 rmsnorm, larger clamp

Create LICENSE

Merge pull request #11 from ClashLuke/patch-1

remove wrongly added lines

add h100

remove torchao dependency, quantize entirely via linear

Fix issues with loading F8Linear from state dict when init_scale not initialized & loaded from meta device

Fix issue where lora alpha is not correct if lora from transformers checkpoint

Small fix for issue where f16 CublasLinear layers weren't being used even when available.

Ensure repo only accesses CublasLinear lazily

Fix issue loading loras

Add lora loading

Merge pull request #3 from aredden/improved_precision

Update readme

Add quantize embedders/modulation to argparse options

Remove f8 flux, instead configure at load, improved quality & corrected configs

Fix issue where torch.dtype throws error when converting to dtype

Move compile out of FluxPipeline init

Dynamic swap with cublas linear / optional improved precision with vram drawback

Allow overriding config values from load_pipeline_from_config_path

Merge pull request #2 from XmYx/main

python 3.10 compatibility

Merge pull request #1 from dsingal0/main

update README

Remove unnecessary tokenization (still needs work)

Add case where seed is string & try/catch if invalid

Remove unnecessary synchronize, add more universal seeding & limit if run on windows

Remove more unnecessary code, fix small typing hickup

Remove unnecessary code, hide prints behind debug flag, hide warnings

Update README.md - Fix config path names

Small README update for clarification

Add all relevent args to argparser & update readme

Merge branch 'main' of https://github.com/aredden/flux-fp16-acc-api into main

Adding device specific configs & more input image type options + small model spec from args change

Update README.md

make table more clear- again

make table more clear

Add speed comparison

Add uvicorn requirement

Add fields to configs, fix issue with offload from bnb, remove extra random text code

Make README.md more comprehensive

Fix for nightly

Fix non-prequantized inference

cuda version checks

Fix non-offload inference & add option to load from prequantized flux