New gpu response for gpu packets with multiple gpus

svc/router.py

@@ -53,6 +53,7 @@ async def deep_eval_status():
 @router.get("/deepeval/hardware")
 def hardware_status():
     info = get_gpu_tier()
+    print("Hardware Response:", info)
     return info
 
 @router.post("/chat",  response_model=TaskResponse)
@@ -155,17 +156,20 @@ def get_gpu_tier():
     if not torch.cuda.is_available():
         return {"gpu": "CPU", "tier": "cpu"}
 
-    gpu_name = torch.cuda.get_device_name(0).lower()
+    device_count = torch.cuda.device_count()
+    gpu_names = [torch.cuda.get_device_name(i).lower() for i in range(device_count)]
 
-    # Normalize GPU model to your custom tier system
-    if "t4" in gpu_name:
-        # You can improve this by checking memory or other context
+    # Count how many of each GPU type we care about
+    l4_count = sum("l4" in name and "l40s" not in name for name in gpu_names)
+    l40s_count = sum("l40s" in name for name in gpu_names)
+
+    if l4_count == device_count:
+        return {"gpu": "NVIDIA L4", "tier": f"l4x{l4_count}"}
+    elif l40s_count == device_count:
+        return {"gpu": "NVIDIA L40S", "tier": f"l40sx{l40s_count}"}
+    elif "t4" in gpu_names[0]:
         return {"gpu": "Tesla T4", "tier": "t4-medium"}
-    elif "l4" in gpu_name:
-        return {"gpu": "NVIDIA L4", "tier": "l4x1"}
-    elif "l40s" in gpu_name:
-        return {"gpu": "NVIDIA L40S", "tier": "l40sx1"}
-    elif "a10g" in gpu_name:
+    elif "a10g" in gpu_names[0]:
         return {"gpu": "NVIDIA A10G", "tier": "a10g"}
     else:
-        return {"gpu": gpu_name, "tier": "unknown"}
+        return {"gpu": gpu_names[0], "tier": "unknown"}