Fix MIG GPU support

clearml/utilities/gpu/gpustat.py

@@ -56,6 +56,21 @@ def uuid(self):
         """
         return self.entry['uuid']
 
+    @property
+    def mig_index(self):
+        """
+        Returns the index of the MIG partition (as in nvidia-smi).
+        """
+        return self.entry.get("mig_index")
+
+    @property
+    def mig_uuid(self):
+        """
+        Returns the uuid of the MIG partition returned by nvidia-smi when running in MIG mode,
+        e.g. MIG-12345678-abcd-abcd-uuid-123456abcdef
+        """
+        return self.entry.get("mig_uuid")
+
     @property
     def name(self):
         """
@@ -160,6 +175,7 @@ class GPUStatCollection(object):
     _initialized = False
     _device_count = None
     _gpu_device_info = {}
+    _mig_device_info = {}
 
     def __init__(self, gpu_list, driver_version=None, driver_cuda_version=None):
         self.gpus = gpu_list
@@ -190,7 +206,7 @@ def _decode(b):
                 return b.decode()  # for python3, to unicode
             return b
 
-        def get_gpu_info(index, handle):
+        def get_gpu_info(index, handle, is_mig=False):
             """Get one GPU information specified by nvml handle"""
 
             def get_process_info(nv_process):
@@ -226,12 +242,13 @@ def get_process_info(nv_process):
                     pass
                 return process
 
-            if not GPUStatCollection._gpu_device_info.get(index):
+            device_info = GPUStatCollection._mig_device_info if is_mig else GPUStatCollection._gpu_device_info
+            if not device_info.get(index):
                 name = _decode(N.nvmlDeviceGetName(handle))
                 uuid = _decode(N.nvmlDeviceGetUUID(handle))
-                GPUStatCollection._gpu_device_info[index] = (name, uuid)
+                device_info[index] = (name, uuid)
 
-            name, uuid = GPUStatCollection._gpu_device_info[index]
+            name, uuid = device_info[index]
 
             try:
                 temperature = N.nvmlDeviceGetTemperature(
@@ -327,8 +344,36 @@ def get_process_info(nv_process):
         for index in range(GPUStatCollection._device_count):
             handle = N.nvmlDeviceGetHandleByIndex(index)
             gpu_info = get_gpu_info(index, handle)
-            gpu_stat = GPUStat(gpu_info)
-            gpu_list.append(gpu_stat)
+            mig_cnt = 0
+            # noinspection PyBroadException
+            try:
+                mig_cnt = N.nvmlDeviceGetMaxMigDeviceCount(handle)
+            except Exception:
+                pass
+
+            if mig_cnt <= 0:
+                gpu_list.append(GPUStat(gpu_info))
+                continue
+
+            got_mig_info = False
+            for mig_index in range(mig_cnt):
+                try:
+                    mig_handle = N.nvmlDeviceGetMigDeviceHandleByIndex(handle, mig_index)
+                    mig_info = get_gpu_info(mig_index, mig_handle, is_mig=True)
+                    mig_info["mig_name"] = mig_info["name"]
+                    mig_info["name"] = gpu_info["name"]
+                    mig_info["mig_index"] = mig_info["index"]
+                    mig_info["mig_uuid"] = mig_info["uuid"]
+                    mig_info["index"] = gpu_info["index"]
+                    mig_info["uuid"] = gpu_info["uuid"]
+                    mig_info["temperature.gpu"] = gpu_info["temperature.gpu"]
+                    mig_info["fan.speed"] = gpu_info["fan.speed"]
+                    gpu_list.append(GPUStat(mig_info))
+                    got_mig_info = True
+                except Exception as e:
+                    pass
+            if not got_mig_info:
+                gpu_list.append(GPUStat(gpu_info))
 
         # 2. additional info (driver version, etc).
         if get_driver_info:

clearml/utilities/resource_monitor.py

@@ -12,6 +12,7 @@
 from .process.mp import BackgroundMonitor
 from ..backend_api import Session
 from ..binding.frameworks.tensorflow_bind import IsTensorboardInit
+from ..config import config
 
 try:
     from .gpu import gpustat
@@ -46,6 +47,11 @@ def __init__(self, task, sample_frequency_per_sec=2., report_frequency_sec=30.,
         self._last_process_pool = {}
         self._last_process_id_list = []
         self._gpu_memory_per_process = True
+        self._default_gpu_utilization = config.get("resource_monitoring.default_gpu_utilization", 100)
+        # allow default_gpu_utilization as null in the config, in which case we don't log anything
+        if self._default_gpu_utilization is not None:
+            self._default_gpu_utilization = int(self._default_gpu_utilization)
+        self._gpu_utilization_warning_sent = False
 
         # noinspection PyBroadException
         try:
@@ -314,13 +320,18 @@ def mem_usage_children(a_mem_size, pr, parent_mem=None):
 
         return mem_size
 
-    def _skip_nonactive_gpu(self, idx, gpu):
+    def _skip_nonactive_gpu(self, gpu):
         if not self._active_gpus:
             return False
         # noinspection PyBroadException
         try:
             uuid = getattr(gpu, "uuid", None)
-            return str(idx) not in self._active_gpus and (not uuid or uuid not in self._active_gpus)
+            mig_uuid = getattr(gpu, "mig_uuid", None)
+            return (
+                str(gpu.index) not in self._active_gpus
+                and (not uuid or uuid not in self._active_gpus)
+                and (not mig_uuid or mig_uuid not in self._active_gpus)
+            )
         except Exception:
             pass
         return False
@@ -349,7 +360,7 @@ def _get_gpu_stats(self):
                     self._gpu_memory_per_process = False
                     break
                 # only monitor the active gpu's, if none were selected, monitor everything
-                if self._skip_nonactive_gpu(i, g):
+                if self._skip_nonactive_gpu(g):
                     continue
 
                 gpu_mem[i] = 0
@@ -369,10 +380,27 @@ def _get_gpu_stats(self):
 
         for i, g in enumerate(gpu_stat.gpus):
             # only monitor the active gpu's, if none were selected, monitor everything
-            if self._skip_nonactive_gpu(i, g):
+            if self._skip_nonactive_gpu(g):
                 continue
             stats["gpu_%d_temperature" % i] = float(g["temperature.gpu"])
-            stats["gpu_%d_utilization" % i] = float(g["utilization.gpu"])
+            if g["utilization.gpu"] is not None:
+                stats["gpu_%d_utilization" % i] = float(g["utilization.gpu"])
+            else:
+                stats["gpu_%d_utilization" % i] = self._default_gpu_utilization
+                if not self._gpu_utilization_warning_sent:
+                    if g.mig_index is not None:
+                        self._task.get_logger().report_text(
+                            "Running inside MIG, Nvidia driver cannot export utilization, pushing fixed value {}".format(  # noqa
+                                self._default_gpu_utilization
+                            )
+                        )
+                    else:
+                        self._task.get_logger().report_text(
+                            "Nvidia driver cannot export utilization, pushing fixed value {}".format(
+                                self._default_gpu_utilization
+                            )
+                        )
+                    self._gpu_utilization_warning_sent = True
             stats["gpu_%d_mem_usage" % i] = 100. * float(g["memory.used"]) / float(g["memory.total"])
             # already in MBs
             stats["gpu_%d_mem_free_gb" % i] = float(g["memory.total"] - g["memory.used"]) / 1024
@@ -400,7 +428,7 @@ def _get_machine_specs(self):
             if self._gpustat:
                 gpu_stat = self._gpustat.new_query(shutdown=True, get_driver_info=True)
                 if gpu_stat.gpus:
-                    gpus = [g for i, g in enumerate(gpu_stat.gpus) if not self._skip_nonactive_gpu(i, g)]
+                    gpus = [g for i, g in enumerate(gpu_stat.gpus) if not self._skip_nonactive_gpu(g)]
                     specs.update(
                         gpu_count=int(len(gpus)),
                         gpu_type=', '.join(g.name for g in gpus),