-
diffusers - Modified Stable Diffusion XL pipeline in
diffusers/pipelines/stable_diffusion_xl. The code was cloned from Hugging Face Diffusers and adapted to incorporate approximate caching. Lines modified/added for the caching process. -
noise - Directory to store noises generated for image generation using approximate caching.
-
pipeline - Contains cache scripts and utilities, classifier scripts and utilities, cache_logs to display queries and their retrieved K values for a sample, along with a small set of examples with images, and a notebook to see the pipeline end-to-end. The
pipeline-system.ipynbnotebook is left empty for anonymous submission. -
generate_approx_cache.ipynb - A notebook to generate noise and use it to generate images for query prompts using approximate caching.
-
csv files - Workload traces for Normal and Bursty cases.
-
solver_milp_oda - Sample pseudo-code for Distribution Aligner (ODA) and Allocation (MILP) solvers.
-
Results_Analysis - Plots for bursty distribution and load time
For installing required dependecies, please use following:
pip install torch==2.0.0,
pip install diffusers==0.21.2,
pip install transformers==4.26.1
Use the generate_approx_caching.ipynb notebook to:
- Generate noise for any cache prompt at all K values.
- Generate images for query prompts using the noise generated in step 1 at specified K values.
- Render to see cached image (baseline GPTCache), image generated using approximate caching (proposed NIRVANA), and standard diffusion generation (baseline Vanilla).
This script uses the stable diffusion XL pipeline, which takes 15 seconds for Vanilla generation on an A10g GPU. We have modified the script in the diffusers library (diffusers/pipelines/stable_diffusion_xl) and provided a custom diffusers directory.
- Note: The full end-to-end code implementation is currently under construction.