diff --git a/lollms/server/endpoints/lollms_hardware_infos.py b/lollms/server/endpoints/lollms_hardware_infos.py
index 3c88924..55e6803 100644
--- a/lollms/server/endpoints/lollms_hardware_infos.py
+++ b/lollms/server/endpoints/lollms_hardware_infos.py
@@ -17,6 +17,7 @@ from pathlib import Path
 from typing import List
 import psutil
 import subprocess
+from ascii_colors import trace_exception
 
 # ----------------------- Defining router and main class ------------------------------
 router = APIRouter()
@@ -64,31 +65,85 @@ def ram_usage():
         "percent_usage":ram.percent,
         "ram_usage": ram.used
         }
-
-@router.get("/vram_usage")
+@router.get("/vram_usage") # Or @app.route("/vram_usage", methods=['GET']) for Flask
 def vram_usage():
+    """
+    Fetches GPU VRAM usage information using nvidia-smi.
+
+    Returns:
+        dict: A dictionary containing a 'gpus' key. The value is a list of
+              dictionaries, each representing a GPU with its VRAM details.
+              Returns {"gpus": []} if nvidia-smi fails or no GPUs are detected.
+    """
+    gpus_list = []
     try:
-        output = subprocess.check_output(['nvidia-smi', '--query-gpu=memory.total,memory.used,gpu_name', '--format=csv,nounits,noheader'])
-        lines = output.decode().strip().split('\n')
-        vram_info = [line.split(',') for line in lines]
-    except (subprocess.CalledProcessError, FileNotFoundError):
-        return {
-        "nb_gpus": 0
-        }
-    
-    ram_usage = {
-        "nb_gpus": len(vram_info)
-    }
-    
-    if vram_info is not None:
-        for i, gpu in enumerate(vram_info):
-            ram_usage[f"gpu_{i}_total_vram"] = int(gpu[0])*1024*1024
-            ram_usage[f"gpu_{i}_used_vram"] = int(gpu[1])*1024*1024
-            ram_usage[f"gpu_{i}_model"] = gpu[2].strip()
-    else:
-        # Set all VRAM-related entries to None
-        ram_usage["gpu_0_total_vram"] = None
-        ram_usage["gpu_0_used_vram"] = None
-        ram_usage["gpu_0_model"] = None
-    
-    return ram_usage
\ No newline at end of file
+        # Execute nvidia-smi to get total memory, used memory, and GPU name
+        # The output is in MiB (Mebibytes)
+        output = subprocess.check_output(
+            ['nvidia-smi', '--query-gpu=memory.total,memory.used,gpu_name', '--format=csv,nounits,noheader'],
+            encoding='utf-8' # Decode the output directly
+        )
+
+        # Split the output into lines, one line per GPU
+        lines = output.strip().split('\n')
+
+        # Process each line
+        for line in lines:
+            if not line.strip(): # Skip potential empty lines
+                continue
+
+            parts = line.split(',')
+            if len(parts) != 3:
+                print(f"Warning: Unexpected nvidia-smi output line format: {line}")
+                continue # Skip this malformed line
+
+            try:
+                # Parse values (nvidia-smi returns MiB)
+                total_vram_mib = int(parts[0].strip())
+                used_vram_mib = int(parts[1].strip())
+                gpu_model = parts[2].strip()
+
+                # Convert MiB to Bytes (as expected by filesize library)
+                total_vram_bytes = total_vram_mib * 1024 * 1024
+                used_vram_bytes = used_vram_mib * 1024 * 1024
+
+                # Calculate available space and percentage
+                available_vram_bytes = total_vram_bytes - used_vram_bytes
+                percentage_used = 0.0
+                if total_vram_bytes > 0:
+                    # Calculate percentage and round to 2 decimal places
+                    percentage_used = round((used_vram_bytes / total_vram_bytes) * 100, 2)
+
+                # Create the dictionary structure expected by the frontend
+                gpu_info = {
+                    "total_vram": total_vram_bytes,
+                    "used_vram": used_vram_bytes,
+                    "available_space": available_vram_bytes, # Frontend uses this in details
+                    "gpu_model": gpu_model,
+                    "percentage": percentage_used # Frontend uses this directly
+                }
+                gpus_list.append(gpu_info)
+
+            except ValueError as ve:
+                print(f"Warning: Could not parse VRAM numbers from line: {line}. Error: {ve}")
+                continue # Skip this line if parsing fails
+
+    except FileNotFoundError:
+        # nvidia-smi command not found (likely no NVIDIA drivers or not in PATH)
+        print("nvidia-smi command not found. Assuming no compatible GPU.")
+        # Return empty list as expected by frontend when no GPUs detected
+        gpus_list = []
+    except subprocess.CalledProcessError as e:
+        # nvidia-smi command failed to execute properly
+        print(f"nvidia-smi failed with error code {e.returncode}: {e.output}")
+        # Return empty list
+        gpus_list = []
+    except Exception as e:
+        # Catch any other unexpected errors during processing
+        trace_exception(e)
+        # Return empty list
+        gpus_list = []
+
+    # Always return the structure the frontend expects, even if the list is empty
+    # If using Flask, use jsonify: return jsonify({"gpus": gpus_list})
+    return {"gpus": gpus_list}
\ No newline at end of file