Remove hardcoded cuda in triton.testing.get_dram_gbps (#5455) (#8717)

Signed-off-by: Anatoly Myachev <[email protected]>

Author: anmyachev

python/triton/testing.py

@@ -420,11 +420,10 @@ def perf_report(benchmarks):
 
 def get_dram_gbps(device=None):
     ''' return DRAM bandwidth in GB/s '''
-    import torch
 
     from .runtime import driver
-    if not device:
-        device = torch.cuda.current_device()
+    if device is None:
+        device = driver.active.get_device_interface().current_device()
     mem_clock_khz = driver.active.utils.get_device_properties(device)["mem_clock_rate"]  # in kHz
     bus_width = driver.active.utils.get_device_properties(device)["mem_bus_width"]
     bw_gbps = mem_clock_khz * bus_width * 2 / 1e6 / 8  # In GB/s