[Core] Ray auto detect nvidia Gpu with pynvml

LICENSE

@@ -448,3 +448,35 @@ distributed under the License is distributed on an "AS IS" BASIS,
 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License.
+
+--------------------------------------------------------------------------------
+
+Code in python/ray/_private/thirdparty/pynvml is adapted from https://pypi.org/project/nvidia-ml-py
+
+Copyright (c) 2011-2022, NVIDIA Corporation.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+* Redistributions of source code must retain the above copyright notice, this
+  list of conditions and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright notice,
+  this list of conditions and the following disclaimer in the documentation
+  and/or other materials provided with the distribution.
+
+* Neither the name of staged-recipes nor the names of its
+  contributors may be used to endorse or promote products derived from
+  this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

python/ray/_private/accelerators/nvidia_gpu.py

@@ -1,17 +1,12 @@
 import re
 import os
-import sys
 import logging
-import subprocess
-import importlib
 from typing import Optional, List, Tuple
-
-try:
-    import GPUtil
-except ImportError:
-    pass
+import ray._private.thirdparty.pynvml as pynvml
+from packaging.version import Version
 
 from ray._private.accelerators.accelerator import AcceleratorManager
+import ray._private.ray_constants as ray_constants
 
 logger = logging.getLogger(__name__)
 
@@ -22,6 +17,9 @@
 # the form "Tesla V100-SXM2-16GB" or "Tesla K80").
 NVIDIA_GPU_NAME_PATTERN = re.compile(r"\w+\s+([A-Z0-9]+)")
 
+# version with mig uuid
+MIG_UUID_DRIVER_VERSION = "470.42.01"
+
 
 class NvidiaGPUAcceleratorManager(AcceleratorManager):
     """Nvidia GPU accelerators."""
@@ -39,7 +37,6 @@ def get_current_process_visible_accelerator_ids() -> Optional[List[str]]:
         cuda_visible_devices = os.environ.get(
             NvidiaGPUAcceleratorManager.get_visible_accelerator_ids_env_var(), None
         )
-
         if cuda_visible_devices is None:
             return None
 
@@ -53,55 +50,82 @@ def get_current_process_visible_accelerator_ids() -> Optional[List[str]]:
 
     @staticmethod
     def get_current_node_num_accelerators() -> int:
-        num_gpus = 0
-        if importlib.util.find_spec("GPUtil"):
-            gpu_list = GPUtil.getGPUs()
-            num_gpus = len(gpu_list)
-        elif sys.platform.startswith("linux"):
-            proc_gpus_path = "/proc/driver/nvidia/gpus"
-            if os.path.isdir(proc_gpus_path):
-                num_gpus = len(os.listdir(proc_gpus_path))
-        elif sys.platform == "win32":
-            props = "AdapterCompatibility"
-            cmdargs = ["WMIC", "PATH", "Win32_VideoController", "GET", props]
-            lines = subprocess.check_output(cmdargs).splitlines()[1:]
-            num_gpus = len([x.rstrip() for x in lines if x.startswith(b"NVIDIA")])
-        return num_gpus
+        try:
+            pynvml.nvmlInit()
+        except pynvml.NVMLError:
+            return 0  # pynvml init failed
+        driver_version = pynvml.nvmlSystemGetDriverVersion()
+        device_count = pynvml.nvmlDeviceGetCount()
+        cuda_devices = []
+        for index in range(device_count):
+            handle = pynvml.nvmlDeviceGetHandleByIndex(index)
+            mig_enabled = os.environ.get(
+                ray_constants.RAY_ENABLE_MIG_DETECTION_ENV_VAR, False
+            )
+            if mig_enabled:
+                try:
+                    max_mig_count = pynvml.nvmlDeviceGetMaxMigDeviceCount(handle)
+                except pynvml.NVMLError_NotSupported:
+                    cuda_devices.append(str(index))
+                    continue
+                for mig_index in range(max_mig_count):
+                    try:
+                        mig_handle = pynvml.nvmlDeviceGetMigDeviceHandleByIndex(
+                            handle, mig_index
+                        )
+                        mig_uuid = ""
+                        if Version(driver_version) >= Version(MIG_UUID_DRIVER_VERSION):
+                            mig_uuid = pynvml.nvmlDeviceGetUUID(mig_handle)
+                        else:
+                            mig_uuid = (
+                                f"MIG-{pynvml.nvmlDeviceGetUUID(handle)}"
+                                f"/{pynvml.nvmlDeviceGetComputeInstanceId(mig_handle)}"
+                                f"/{pynvml.nvmlDeviceGetGpuInstanceId(mig_handle)}"
+                            )
+                        cuda_devices.append(mig_uuid)
+                    except pynvml.NVMLError:
+                        break
+            else:
+                cuda_devices.append(str(index))
+        os.environ[
+            NvidiaGPUAcceleratorManager.get_visible_accelerator_ids_env_var()
+        ] = ",".join(cuda_devices)
+        pynvml.nvmlShutdown()
+        return len(cuda_devices)
 
     @staticmethod
     def get_current_node_accelerator_type() -> Optional[str]:
         try:
-            if importlib.util.find_spec("GPUtil"):
-                gpu_list = GPUtil.getGPUs()
-                if len(gpu_list) > 0:
-                    gpu_list_names = [gpu.name for gpu in gpu_list]
-                    return NvidiaGPUAcceleratorManager._gpu_name_to_accelerator_type(
-                        gpu_list_names.pop()
-                    )
-            elif sys.platform.startswith("linux"):
-                proc_gpus_path = "/proc/driver/nvidia/gpus"
-                if not os.path.isdir(proc_gpus_path):
-                    return None
-                gpu_dirs = os.listdir(proc_gpus_path)
-                if len(gpu_dirs) == 0:
-                    return None
-                gpu_info_path = f"{proc_gpus_path}/{gpu_dirs[0]}/information"
-                info_str = open(gpu_info_path).read()
-                if not info_str:
-                    return None
-                lines = info_str.split("\n")
-                full_model_name = None
-                for line in lines:
-                    split = line.split(":")
-                    if len(split) != 2:
-                        continue
-                    k, v = split
-                    if k.strip() == "Model":
-                        full_model_name = v.strip()
-                        break
-                return NvidiaGPUAcceleratorManager._gpu_name_to_accelerator_type(
-                    full_model_name
+            pynvml.nvmlInit()
+            device_count = pynvml.nvmlDeviceGetCount()
+            cuda_devices_names = []
+            for index in range(device_count):
+                handle = pynvml.nvmlDeviceGetHandleByIndex(index)
+                mig_enabled = os.environ.get(
+                    ray_constants.RAY_ENABLE_MIG_DETECTION_ENV_VAR, False
                 )
+                if mig_enabled:
+                    try:
+                        max_mig_count = pynvml.nvmlDeviceGetMaxMigDeviceCount(handle)
+                    except pynvml.NVMLError_NotSupported:
+                        cuda_devices_names.append(pynvml.nvmlDeviceGetName(handle))
+                        continue
+                    for mig_index in range(max_mig_count):
+                        try:
+                            mig_handle = pynvml.nvmlDeviceGetMigDeviceHandleByIndex(
+                                handle, mig_index
+                            )
+                            cuda_devices_names.append(
+                                pynvml.nvmlDeviceGetName(mig_handle)
+                            )
+                        except pynvml.NVMLError:
+                            break
+                else:
+                    cuda_devices_names.append(pynvml.nvmlDeviceGetName(handle))
+            pynvml.nvmlShutdown()
+            return NvidiaGPUAcceleratorManager._gpu_name_to_accelerator_type(
+                cuda_devices_names.pop()
+            )
         except Exception:
             logger.exception("Could not parse gpu information.")
         return None

python/ray/_private/ray_constants.py

@@ -480,3 +480,5 @@ def gcs_actor_scheduling_enabled():
 
 
 RAY_NODE_IP_FILENAME = "node_ip_address.json"
+
+RAY_ENABLE_MIG_DETECTION_ENV_VAR = "RAY_ENABLE_MIG_DETECTION"

python/ray/_private/resource_spec.py

@@ -199,6 +199,9 @@ def resolve(self, is_head: bool, node_ip_address: Optional[str] = None):
                 num_accelerators = (
                     accelerator_manager.get_current_node_num_accelerators()
                 )
+                visible_accelerator_ids = (
+                    accelerator_manager.get_current_process_visible_accelerator_ids()
+                )
                 # Don't use more accelerators than allowed by visible accelerator ids.
                 if visible_accelerator_ids is not None:
                     num_accelerators = min(

python/ray/_private/thirdparty/pynvml/__init__.py

@@ -0,0 +1 @@
+from ray._private.thirdparty.pynvml.pynvml import *