Update app.py

app.py

@@ -1,6 +1,6 @@
 import gradio as gr
-from peft import AutoPeftModelForCausalLM
-from transformers import AutoTokenizer
+from peft import AutoPeftModelForCausalLM, PeftConfig
+from transformers import AutoTokenizer, AutoConfig
 from huggingface_hub import login, snapshot_download
 import torch
 import os
@@ -9,25 +9,34 @@ import json
 # Login using secret (secure, no hardcode)
 login(os.environ["HF_TOKEN"])
 
-# Model setup (loads once on Space startup; switched to Llama 3 base)
-model_id = "agarkovv/CryptoTrader-LM"  # Keep PEFT adapter if compatible; otherwise fine-tune on Llama
-base_model_id = "meta-llama/Meta-Llama-3-8B-Instruct"
+# Model setup (loads once on Space startup)
+model_id = "agarkovv/CryptoTrader-LM"
+base_model_id = "mistralai/Ministral-8B-Instruct-2410"
 MAX_LENGTH = 32768
 DEVICE = "cuda" if torch.cuda.is_available() else "cpu"  # Use GPU if available (ZeroGPU on HF)
 
-# Workaround: Download model files, edit adapter_config.json to remove 'model_type' if present
-local_dir = snapshot_download(repo_id=model_id)
-config_path = os.path.join(local_dir, "adapter_config.json")
+# Download adapter files
+adapter_local_dir = snapshot_download(repo_id=model_id)
+config_path = os.path.join(adapter_local_dir, "adapter_config.json")
 with open(config_path, 'r') as f:
-    config = json.load(f)
-if 'model_type' in config:
-    del config['model_type']
+    adapter_config = json.load(f)
+if 'model_type' in adapter_config:
+    del adapter_config['model_type']
 with open(config_path, 'w') as f:
-    json.dump(config, f)
+    json.dump(adapter_config, f)
 
-# Now load the model from modified local dir, passing token for gated base model
+# Download base model config locally to avoid gated access issues
+base_local_dir = snapshot_download(repo_id=base_model_id, allow_patterns="config.json")
+base_config_path = os.path.join(base_local_dir, "config.json")
+base_config = AutoConfig.from_pretrained(base_config_path)
+
+# Load model with explicit base config
 token = os.environ["HF_TOKEN"]
-model = AutoPeftModelForCausalLM.from_pretrained(local_dir, token=token)
+model = AutoPeftModelForCausalLM.from_pretrained(
+    adapter_local_dir,
+    config=base_config,
+    token=token
+)
 tokenizer = AutoTokenizer.from_pretrained(base_model_id, token=token)
 model = model.to(DEVICE)
 model.eval()
@@ -62,8 +71,8 @@ demo = gr.Interface(
     fn=predict_trading_decision,
     inputs=gr.Textbox(label="Input Prompt (News + Prices)"),
     outputs=gr.Textbox(label="Trading Decision"),
-    title="CryptoTrader-LM with Llama MCP Tool",
-    description="Predict buy/sell/hold for BTC/ETH using Llama 3 base."
+    title="CryptoTrader-LM MCP Tool",
+    description="Predict buy/sell/hold for BTC/ETH."
 )
 
 # Launch with MCP support