Fix all: writable /tmp cache, no login(), proper permissions

Dockerfile

@@ -2,9 +2,8 @@ FROM python:3.10-slim
 
 WORKDIR /app
 
-ENV HF_HOME=/app/.cache/huggingface
-ENV TRANSFORMERS_CACHE=/app/.cache/huggingface
-RUN mkdir -p /app/.cache/huggingface
+ENV HF_HOME=/tmp/hf_cache
+ENV TRANSFORMERS_CACHE=/tmp/hf_cache
 
 RUN pip install --no-cache-dir \
     torch \
@@ -17,4 +16,6 @@ RUN pip install --no-cache-dir \
 
 COPY train_on_hf.py .
 
+RUN chmod -R 777 /app
+
 CMD ["sh", "-c", "python train_on_hf.py --hf-token $HF_TOKEN"]

train_on_hf.py

@@ -13,15 +13,15 @@ Usage (on HF with GPU):
 
 import argparse
 import json
-import os
-import sys
 from pathlib import Path
 
 
 def setup(hf_token: str):
     """Login and download data from HF."""
-    from huggingface_hub import login, hf_hub_download, snapshot_download
-    login(token=hf_token)
+    import os
+    os.environ["HF_TOKEN"] = hf_token
+    os.environ["HUGGING_FACE_HUB_TOKEN"] = hf_token
+    from huggingface_hub import hf_hub_download, snapshot_download
 
     # Download training data
     data_dir = Path("data")
@@ -326,8 +326,6 @@ def merge_and_push(hf_token: str):
     """Merge LoRA, push merged model to HF Hub."""
     from transformers import AutoModelForCausalLM, AutoTokenizer
     from peft import PeftModel
-    from huggingface_hub import login
-    login(token=hf_token)
 
     config = json.load(open("data/preprocessing_config.json"))
     MODEL_NAME = config["model"]["name"]