add device config env for the accelerator (microsoft#5396)

Thank you for [pr](microsoft#5369) and
@delock contribution of ideas.
As mentioned in this
[pr](microsoft#5369), each device has
its own environmental variables.
We create visible_devices_envs() and set_visible_devices_envs() methods
on the accelerator class to enable each accelerator to implement env
settings within the interface , which is more generic to other
accelerators.

this commit has tested on npu, each one has 8 ascend npus

---------

Co-authored-by: yangcheng <yangcheng104@huawei.com>
Co-authored-by: eigen2017 <wobushiliu2@gmail.com>
Co-authored-by: Logan Adams <114770087+loadams@users.noreply.github.com>
Co-authored-by: Olatunji Ruwase <olruwase@microsoft.com>

README.md

@@ -159,11 +159,12 @@ dynamically link them at runtime.
 ## Contributed HW support
 * DeepSpeed now support various HW accelerators.
 
-| Contributor | Hardware | Accelerator Name | Contributor validated | Upstream validated |
-| ----------- | -------- | ---------------- | --------------------- | ------------------ |
-| Intel | Intel(R) Gaudi(R) 2 AI accelerator | hpu | Yes | Yes |
-| Intel | Intel(R) Xeon(R) Processors | cpu | Yes | Yes |
-| Intel | Intel(R) Data Center GPU Max series | xpu | Yes | Yes |
+| Contributor | Hardware                            | Accelerator Name | Contributor validated | Upstream validated |
+|-------------|-------------------------------------|------------------| --------------------- |--------------------|
+| Huawei      | Huawei Ascend NPU                   | npu              | Yes | No                 |
+| Intel       | Intel(R) Gaudi(R) 2 AI accelerator  | hpu              | Yes | Yes                |
+| Intel       | Intel(R) Xeon(R) Processors         | cpu              | Yes | Yes                |
+| Intel       | Intel(R) Data Center GPU Max series | xpu              | Yes | Yes                |
 
 ## PyPI
 We regularly push releases to [PyPI](https://pypi.org/project/deepspeed/) and encourage users to install from there in most cases.

accelerator/abstract_accelerator.py

@@ -287,3 +287,11 @@ def build_extension(self):
     @abc.abstractmethod
     def export_envs(self):
         ...
+
+    @abc.abstractmethod
+    def visible_devices_envs(self):
+        ...
+
+    @abc.abstractmethod
+    def set_visible_devices_envs(self, current_env, local_accelerator_ids):
+        ...

accelerator/cpu_accelerator.py

@@ -322,3 +322,11 @@ def build_extension(self):
 
     def export_envs(self):
         return []
+
+    # TODO: cpu's visible envs is confirmed, keep as CUDA_VISIBLE_DEVICES
+    def visible_devices_envs(self):
+        return ['CUDA_VISIBLE_DEVICES']
+
+    def set_visible_devices_envs(self, current_env, local_accelerator_ids):
+        for env in self.visible_devices_envs():
+            current_env[env] = ",".join(map(str, local_accelerator_ids))

accelerator/cuda_accelerator.py

@@ -360,3 +360,10 @@ def build_extension(self):
 
     def export_envs(self):
         return ['NCCL']
+
+    def visible_devices_envs(self):
+        return ['CUDA_VISIBLE_DEVICES']
+
+    def set_visible_devices_envs(self, current_env, local_accelerator_ids):
+        for env in self.visible_devices_envs():
+            current_env[env] = ",".join(map(str, local_accelerator_ids))

accelerator/hpu_accelerator.py

@@ -294,3 +294,10 @@ def build_extension(self):
 
     def export_envs(self):
         return []
+
+    def visible_devices_envs(self):
+        return ['HABANA_VISIBLE_MODULES']
+
+    def set_visible_devices_envs(self, current_env, local_accelerator_ids):
+        for env in self.visible_devices_envs():
+            current_env[env] = ",".join(map(str, local_accelerator_ids))

accelerator/mps_accelerator.py

@@ -258,3 +258,12 @@ def build_extension(self):
 
     def export_envs(self):
         return []
+
+    # TODO: mpu's visible envs is confirmed, keep as CUDA_VISIBLE_DEVICES
+    def visible_devices_envs(self):
+        # TODO: could not find visible devices env for mps
+        return ['CUDA_VISIBLE_DEVICES']
+
+    def set_visible_devices_envs(self, current_env, local_accelerator_ids):
+        for env in self.visible_devices_envs():
+            current_env[env] = ",".join(map(str, local_accelerator_ids))

accelerator/npu_accelerator.py

@@ -278,3 +278,10 @@ def build_extension(self):
 
     def export_envs(self):
         return ['ASCEND', 'HCCL', 'LD_LIBRARY', 'PATH']
+
+    def visible_devices_envs(self):
+        return ['ASCEND_RT_VISIBLE_DEVICES']
+
+    def set_visible_devices_envs(self, current_env, local_accelerator_ids):
+        for env in self.visible_devices_envs():
+            current_env[env] = ",".join(map(str, local_accelerator_ids))

accelerator/xpu_accelerator.py

@@ -289,3 +289,10 @@ def build_extension(self):
 
     def export_envs(self):
         return []
+
+    def visible_devices_envs(self):
+        return ['ZE_AFFINITY_MASK']
+
+    def set_visible_devices_envs(self, current_env, local_accelerator_ids):
+        for env in self.visible_devices_envs():
+            current_env[env] = ",".join(map(str, local_accelerator_ids))

deepspeed/launcher/launch.py

@@ -22,6 +22,7 @@
 from collections import defaultdict
 from typing import Dict
 from argparse import ArgumentParser, REMAINDER
+from deepspeed.accelerator import get_accelerator
 from ..constants import TORCH_DISTRIBUTED_DEFAULT_PORT
 from ..nebula.constants import DLTS_POD_ENV_PATH
 from ..utils import logger, get_numactl_cmd
@@ -146,8 +147,8 @@ def main():
     node_list = list(world_info.keys())
     args.nnodes = len(node_list)
     local_node = node_list[args.node_rank]
-    local_gpu_ids = world_info[local_node]
-    num_local_procs = len(local_gpu_ids)
+    local_accelerator_ids = world_info[local_node]
+    num_local_procs = len(local_accelerator_ids)
     logger.info(f"nnodes={args.nnodes}, num_local_procs={num_local_procs}, node_rank={args.node_rank}")
 
     global_rank_mapping = defaultdict(list)
@@ -161,8 +162,10 @@ def main():
             curr_global_rank += 1
     logger.info(f"global_rank_mapping={global_rank_mapping}")
     logger.info(f"dist_world_size={dist_world_size}")
-    current_env["CUDA_VISIBLE_DEVICES"] = ",".join(map(str, local_gpu_ids))
-    logger.info(f"Setting CUDA_VISIBLE_DEVICES={current_env['CUDA_VISIBLE_DEVICES']}")
+
+    get_accelerator().set_visible_devices_envs(current_env, local_accelerator_ids)
+    for env in get_accelerator().visible_devices_envs():
+        logger.info(f"Setting {env}={current_env[env]}")
 
     # set PyTorch distributed related environmental variables
     current_env["MASTER_ADDR"] = args.master_addr