Update app.py

app.py

@@ -9,12 +9,34 @@ import torch.nn.functional as F
 from Bio.Blast import NCBIWWW, NCBIXML
 from transformers import AutoTokenizer, AutoModelForSequenceClassification, AutoModel
 from huggingface_hub import hf_hub_download
-import spaces  # REQUIRED FOR ZEROGPU
 
 # ===================================
-# 1. GLOBAL VARIABLES
+# 1. LOAD AI MODELS (GLOBALLY CACHED)
 # ===================================
-print("Waking up the Genomic Oracle... Awaiting sequence for Immediate GPU Access.\n")
+# Because "import spaces" is not called yet, PyTorch is safe from the Meta Device!
+print("Waking up the Genomic Oracle... Loading models safely into CPU RAM.\n")
+
+# A. Kadir's Gatekeeper 
+clf_coding = joblib.load("coding_classifier_universal.joblib")
+
+# B. Base DNABERT 
+tokenizer_base = AutoTokenizer.from_pretrained("DNABERT_Local", trust_remote_code=True)
+model_base = AutoModel.from_pretrained("DNABERT_Local", trust_remote_code=True, low_cpu_mem_usage=False)
+model_base.eval()
+
+# C. DNABERT-2 Promoter Model 
+tokenizer_promoter = AutoTokenizer.from_pretrained("llm_promoter_classifier_v2", trust_remote_code=True)
+model_promoter = AutoModelForSequenceClassification.from_pretrained("llm_promoter_classifier_v2", trust_remote_code=True, low_cpu_mem_usage=False)
+model_promoter.eval()
+
+# D. Multi-Feature LightGBM 
+lgbm_path = hf_hub_download(repo_id="Geonomic/Genomic-Oracle-Weights", filename="dnabert_lightgbm_model_feature_type.pkl")
+lightgbm_model = joblib.load(lgbm_path)
+
+# E. Custom ALiBi Lean/Obese BERT 
+tokenizer_pheno = AutoTokenizer.from_pretrained("Geonomic/Genomic-Oracle-Weights", trust_remote_code=True)
+model_pheno = AutoModelForSequenceClassification.from_pretrained("Geonomic/Genomic-Oracle-Weights", trust_remote_code=True, low_cpu_mem_usage=False)
+model_pheno.eval()
 
 # Structural Feature Dictionary
 FEATURE_DICT = {
@@ -25,31 +47,13 @@ FEATURE_DICT = {
 # ==============================================
 # 2. CORE INFERENCE ENGINE (ZeroGPU Accelerated)
 # ==============================================
-@spaces.GPU  # This tells Hugging Face to teleport this specific math to the A100 GPU
+import spaces  # <--- THE MAGIC DELAY: We import ZeroGPU ONLY AFTER models are safely loaded!
+
+@spaces.GPU  # This tells Hugging Face to teleport the CPU models to the A100 GPU dynamically
 def run_deep_learning_cascade(dna_sequence):
     device = torch.device("cuda")  # Wakes up the A100 connection
     clean_seq = "".join(dna_sequence.split()).upper()
     
-    print("🚀 Allocated GPU is Awake! Loading ALiBi models directly into VRAM...")
-    
-    # --- JUST-IN-TIME MODEL LOADING ---
-    clf_coding = joblib.load("coding_classifier_universal.joblib")
-    
-    tokenizer_base = AutoTokenizer.from_pretrained("DNABERT_Local", trust_remote_code=True)
-    model_base = AutoModel.from_pretrained("DNABERT_Local", trust_remote_code=True, low_cpu_mem_usage=False).to(device)
-    model_base.eval()
-    
-    tokenizer_promoter = AutoTokenizer.from_pretrained("llm_promoter_classifier_v2", trust_remote_code=True)
-    model_promoter = AutoModelForSequenceClassification.from_pretrained("llm_promoter_classifier_v2", trust_remote_code=True, low_cpu_mem_usage=False).to(device)
-    model_promoter.eval()
-    
-    lgbm_path = hf_hub_download(repo_id="Geonomic/Genomic-Oracle-Weights", filename="dnabert_lightgbm_model_feature_type.pkl")
-    lightgbm_model = joblib.load(lgbm_path)
-    
-    tokenizer_pheno = AutoTokenizer.from_pretrained("Geonomic/Genomic-Oracle-Weights", trust_remote_code=True)
-    model_pheno = AutoModelForSequenceClassification.from_pretrained("Geonomic/Genomic-Oracle-Weights", trust_remote_code=True, low_cpu_mem_usage=False).to(device)
-    model_pheno.eval()
-    
     # --- LEVEL 1: Base Embedding & Kadir's Gatekeeper ---
     inputs = tokenizer_base([clean_seq], return_tensors="pt", max_length=300, truncation=True, padding=True)
     inputs = {k: v.to(device) for k, v in inputs.items()}
@@ -203,7 +207,7 @@ def gradio_inference(dna_sequence, run_mapping):
     return summary, "\n".join(stats_lines), context_output, ""
 
 # --- THE UI LAYOUT ---
-with gr.Blocks(theme=gr.themes.Soft(), title="🧬 The Genomic Oracle 🧬") as demo:
+with gr.Blocks(theme=gr.themes.Soft(), title="🧬 The Genomic Oracle") as demo:
     # The Integrated Landing Page
     gr.Markdown(
         """