E2e device cuda (#575)

* use torch.cuda.current_device() instead of local_rank

* ignore NVML errors for gpu stats

* llama lora packing e2e tests

src/axolotl/utils/bench.py

@@ -2,6 +2,7 @@
 
 import pynvml
 import torch
+from pynvml.nvml import NVMLError
 
 
 def gpu_memory_usage(device=0):
@@ -20,11 +21,13 @@ def gpu_memory_usage_smi(device=0):
         device = device.index
     if isinstance(device, str) and device.startswith("cuda:"):
         device = int(device[5:])
-
-    pynvml.nvmlInit()
-    handle = pynvml.nvmlDeviceGetHandleByIndex(device)
-    info = pynvml.nvmlDeviceGetMemoryInfo(handle)
-    return info.used / 1024.0**3
+    try:
+        pynvml.nvmlInit()
+        handle = pynvml.nvmlDeviceGetHandleByIndex(device)
+        info = pynvml.nvmlDeviceGetMemoryInfo(handle)
+        return info.used / 1024.0**3
+    except NVMLError:
+        return 0.0
 
 
 def log_gpu_memory_usage(log, msg, device):

src/axolotl/utils/config.py

@@ -29,7 +29,7 @@ def choose_device(cfg):
         cfg.device_map = "auto"
     else:
         if cfg.device.startswith("cuda"):
-            cfg.device_map = {"": cfg.local_rank}
+            cfg.device_map = {"": torch.cuda.current_device()}
         else:
             cfg.device_map = {"": cfg.device}