disable huggingface cache completely for Spaces

Dockerfile

@@ -9,7 +9,6 @@ RUN pip install --no-cache-dir -r requirements.txt
 
 EXPOSE 7860
 
-ENV HF_HOME=/home/user/.cache/huggingface
-RUN mkdir -p /home/user/.cache/huggingface
+ENV HF_HUB_DISABLE_CACHE=1
 
 CMD ["python", "main.py"]

main.py

@@ -4,9 +4,9 @@ from langdetect import detect
 import torch
 import torch.nn.functional as F
 from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline
-HF_CACHE_DIR = "/home/user/.cache/huggingface"
-os.makedirs(HF_CACHE_DIR, exist_ok=True)
-os.environ["HF_HOME"] = HF_CACHE_DIR
+os.environ["HF_HUB_DISABLE_SYMLINKS_WARNING"] = "1"
+os.environ["TRANSFORMERS_OFFLINE"] = "0"
+os.environ["HF_HUB_DISABLE_CACHE"] = "1"
 
 app = Flask(__name__)
 
@@ -20,17 +20,27 @@ device = "cuda" if torch.cuda.is_available() else "cpu"
 # vi_tokenizer = AutoTokenizer.from_pretrained(VI_MODEL_NAME, use_fast=False)
 # vi_model = AutoModelForSequenceClassification.from_pretrained(VI_MODEL_NAME).to(device)
 # vi_model.eval()
-vi_tokenizer = AutoTokenizer.from_pretrained(VI_MODEL_NAME, use_fast=False)
-vi_model = AutoModelForSequenceClassification.from_pretrained(VI_MODEL_NAME)
+# vi_tokenizer = AutoTokenizer.from_pretrained(VI_MODEL_NAME, use_fast=False)
+# vi_model = AutoModelForSequenceClassification.from_pretrained(VI_MODEL_NAME)
+# vi_model.eval()
+# sentiment_pipeline = pipeline("sentiment-analysis", model=vi_model, tokenizer=vi_tokenizer)
+
+
+# # English model
+# en_tokenizer = AutoTokenizer.from_pretrained(EN_MODEL_NAME)
+# en_model = AutoModelForSequenceClassification.from_pretrained(EN_MODEL_NAME).to(device)
+# en_model.eval()
+
+print("🔄 Loading Vietnamese model from Hugging Face Hub (no cache)...")
+vi_tokenizer = AutoTokenizer.from_pretrained(VI_MODEL_NAME, use_fast=False, local_files_only=False)
+vi_model = AutoModelForSequenceClassification.from_pretrained(VI_MODEL_NAME, local_files_only=False)
 vi_model.eval()
 sentiment_pipeline = pipeline("sentiment-analysis", model=vi_model, tokenizer=vi_tokenizer)
 
-
-# English model
-en_tokenizer = AutoTokenizer.from_pretrained(EN_MODEL_NAME)
-en_model = AutoModelForSequenceClassification.from_pretrained(EN_MODEL_NAME).to(device)
+print("🔄 Loading English model from Hugging Face Hub (no cache)...")
+en_tokenizer = AutoTokenizer.from_pretrained(EN_MODEL_NAME, local_files_only=False)
+en_model = AutoModelForSequenceClassification.from_pretrained(EN_MODEL_NAME, local_files_only=False)
 en_model.eval()
-
 # Label mapping cho PhoBERT
 vi_label_map = {
     0: ("NEGATIVE", "Tiêu cực"),