Initial commit

verl/single_controller/base/megatron/__init__.py

@@ -0,0 +1,13 @@
+# Copyright 2024 Bytedance Ltd. and/or its affiliates
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# 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.

verl/single_controller/base/megatron/worker.py

@@ -0,0 +1,39 @@
+# Copyright 2024 Bytedance Ltd. and/or its affiliates
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# 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.
+
+import os
+from dataclasses import dataclass
+from verl.single_controller.base.worker import Worker, DistRankInfo, DistGlobalInfo
+
+
+class MegatronWorker(Worker):
+
+    def __init__(self, cuda_visible_devices=None) -> None:
+        super().__init__(cuda_visible_devices)
+
+    def get_megatron_global_info(self):
+        from megatron.core import parallel_state as mpu
+        tp_size = mpu.get_tensor_model_parallel_world_size()
+        dp_size = mpu.get_data_parallel_world_size()
+        pp_size = mpu.get_pipeline_model_parallel_world_size()
+        info = DistGlobalInfo(tp_size=tp_size, dp_size=dp_size, pp_size=pp_size)
+        return info
+
+    def get_megatron_rank_info(self):
+        from megatron.core import parallel_state as mpu
+        tp_rank = mpu.get_tensor_model_parallel_rank()
+        dp_rank = mpu.get_data_parallel_rank()
+        pp_rank = mpu.get_pipeline_model_parallel_rank()
+        info = DistRankInfo(tp_rank=tp_rank, dp_rank=dp_rank, pp_rank=pp_rank)
+        return info

verl/single_controller/base/megatron/worker_group.py

@@ -0,0 +1,51 @@
+# Copyright 2024 Bytedance Ltd. and/or its affiliates
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# 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.
+
+from typing import Dict
+
+from .worker import DistRankInfo, DistGlobalInfo
+from verl.single_controller.base import ResourcePool, WorkerGroup
+
+
+class MegatronWorkerGroup(WorkerGroup):
+
+    def __init__(self, resource_pool: ResourcePool, **kwargs):
+        super().__init__(resource_pool=resource_pool, **kwargs)
+        self._megatron_rank_info = None
+        self._megatron_global_info: DistGlobalInfo = None
+
+    def init_megatron(self, default_megatron_kwargs: Dict = None):
+        raise NotImplementedError(f"MegatronWorkerGroup.init_megatron should be overwritten")
+
+    def get_megatron_rank_info(self, rank: int) -> DistRankInfo:
+        assert 0 <= rank < self.world_size, f'rank must be from [0, world_size), Got {rank}'
+        return self._megatron_rank_info[rank]
+
+    @property
+    def tp_size(self):
+        assert self._megatron_global_info is not None, "MegatronWorkerGroup._megatron_global_info must be initialized"
+        return self._megatron_global_info.tp_size
+
+    @property
+    def dp_size(self):
+        assert self._megatron_global_info is not None, "MegatronWorkerGroup._megatron_global_info must be initialized"
+        return self._megatron_global_info.dp_size
+
+    @property
+    def pp_size(self):
+        assert self._megatron_global_info is not None, "MegatronWorkerGroup._megatron_global_info must be initialized"
+        return self._megatron_global_info.pp_size
+
+    def get_megatron_global_info(self):
+        return self._megatron_global_info