Update app.py

app.py

@@ -4,7 +4,7 @@ import json
 import os
 from transformers import AutoModelForCausalLM, AutoTokenizer
 from peft import PeftModel
-from huggingface_hub import InferenceClient
+from huggingface_hub import InferenceClient, snapshot_download
 
 # ==============================================================================
 # 1. CONFIGURATION
@@ -13,19 +13,19 @@ HF_TOKEN = os.getenv("HF_TOKEN")
 PROJECT_TITLE = "The Janus Interface: Semantic Decoupling Architecture"
 
 # Models
-BASE_MODEL_ID = "microsoft/Phi-3.5-mini-instruct" # Official repo for better CPU compatibility
+BASE_MODEL_ID = "microsoft/Phi-3.5-mini-instruct" 
 ADAPTER_ID = "st192011/janus-gold-lora" 
 CLOUD_MODEL_ID = "meta-llama/Meta-Llama-3-8B-Instruct" 
 
 # ==============================================================================
-# 2. ENGINE INITIALIZATION (CPU Optimized)
+# 2. ENGINE INITIALIZATION (CPU Optimized + Config Sanitizer)
 # ==============================================================================
 print("⏳ Initializing Neural Backbone...")
 
 try:
     tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL_ID)
     
-    # CRITICAL FIX: attn_implementation="eager" prevents the DynamicCache error on CPU
+    # Load Base Model
     base_model = AutoModelForCausalLM.from_pretrained(
         BASE_MODEL_ID,
         torch_dtype=torch.bfloat16, 
@@ -34,8 +34,27 @@ try:
         attn_implementation="eager" 
     )
     
-    print(f"⏳ Mounting Janus Adapter ({ADAPTER_ID})...")
-    model = PeftModel.from_pretrained(base_model, ADAPTER_ID)
+    print(f"⏳ Downloading and sanitizing adapter ({ADAPTER_ID})...")
+    
+    # 1. Download the adapter files locally
+    local_adapter_path = snapshot_download(repo_id=ADAPTER_ID, token=HF_TOKEN)
+    
+    # 2. Load the config JSON
+    config_path = os.path.join(local_adapter_path, "adapter_config.json")
+    with open(config_path, "r") as f:
+        config_data = json.load(f)
+    
+    # 3. Remove the key that causes the crash
+    if "alora_invocation_tokens" in config_data:
+        print("🧹 Cleaning incompatible Unsloth config keys...")
+        del config_data["alora_invocation_tokens"]
+        
+        # Save the clean config back to disk
+        with open(config_path, "w") as f:
+            json.dump(config_data, f, indent=2)
+            
+    # 4. Load the adapter from the local sanitized folder
+    model = PeftModel.from_pretrained(base_model, local_adapter_path)
     model.eval()
     print("✅ System Online.")
 
@@ -54,7 +73,6 @@ def clean_output(text):
     clean = text.replace("<|end|>", "").replace("<|endoftext|>", "")
     if "Output:" in clean: clean = clean.split("Output:")[-1]
     
-    # Remove conversational filler lines
     lines = clean.split('\n')
     valid_lines = [line for line in lines if ":" in line and "Note" not in line]
     return " ".join(valid_lines).strip()
@@ -82,7 +100,7 @@ RAW NOTE:
                 max_new_tokens=256, 
                 temperature=0.1,
                 do_sample=True,
-                use_cache=True # Enabled for speed
+                use_cache=True
             )
         
         text = tokenizer.batch_decode(outputs)[0]
@@ -136,11 +154,11 @@ PRIVATE_DB:
         with torch.no_grad():
             outputs = model.generate(
                 **inputs, 
-                max_new_tokens=600, # Reduced slightly to ensure completion on CPU
+                max_new_tokens=600, 
                 temperature=0.1, 
                 repetition_penalty=1.05,
                 do_sample=True,
-                use_cache=True # Enabled for speed
+                use_cache=True
             )
         
         text = tokenizer.batch_decode(outputs)[0]
@@ -228,6 +246,7 @@ To achieve high fidelity without using private patient data, we developed a **Te
 *   **Base Model:** Microsoft Phi-3.5-mini-instruct (3.8B Parameters).
 *   **Framework:** **Unsloth** (Optimized QLoRA).
 *   **Technique:** **DoRA (Weight-Decomposed Low-Rank Adaptation)**.
+    *   *Why DoRA?* Standard LoRA struggles with strict syntax/coding tasks. DoRA updates both magnitude and direction vectors, allowing the model to learn the strict `JanusScript` grammar effectively.
 *   **Loss Masking:** We used `train_on_responses_only`. The model was **never** trained on the input text, only on the output. This prevents the model from memorizing patient PII from the training set.
 *   **Hyperparameters:** Rank 16, Alpha 16, Learning Rate 2e-4, 2 Epochs (306 samples).