revert to non-persistence of weights, describe persistence issue in issues_doc

app.py

@@ -16,8 +16,6 @@ from rfd3.engine import RFD3InferenceConfig, RFD3InferenceEngine
 
 cmd =  f"foundry install rfd3 ligandmpnn rf3 --checkpoint-dir /data/checkpoints"
 print("Global PATH:", os.environ.get("PATH"))
-os.environ["HF_HOME"] = "/data"  # HF Spaces persistent dir
-os.environ["FOUNDRY_CHECKPOINT_DIRS"] = "/data/checkpoints"
 
 result = subprocess.run(cmd, shell=True, capture_output=True, text=True)
 if result.returncode == 0:

issues_doc.md

@@ -0,0 +1,6 @@
+- I'm trying to make it so that pushing new code (editing app.py) and building again does not need to redownload the weights
+- In a space like BoltzGen, all the installation happens in requirements.txt. downloading Boltzgen downloads the weights. This means that when I push app.py code, HF detects that requirements.txt did not change, it pulls the built image from cache and there is no need for redownload
+- In my space, this is not the case, pip install rc-foundry is cached (from requirements.txt) but I need to run the command "foundry install rfd3 ligandmpnn rf3" to download the weights. I do the installation in the header of app.py. This means that the installation reruns every code push to app.py
+- I considered fixing this using Docker instead of Gradio as I could create a Docker image with the weights downloaded but Docker is not compatible with ZeroGPU, needed for the hackathon
+- I tried to use persistent storage using ./data to store the weights but persistent storage is only available at runtime not during build so I get a "Permission Denied" error.
+- This problem is annoying for development but not really once the space is done and is being used as there will be no more code pushes.