train.py

# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0

import ray
from sglang.srt.constants import GPU_MEMORY_TYPE_KV_CACHE, GPU_MEMORY_TYPE_WEIGHTS

try:
    from sglang.srt.constants import GPU_MEMORY_TYPE_CUDA_GRAPH
except ImportError:
    GPU_MEMORY_TYPE_CUDA_GRAPH = None

from slime.ray.placement_group import create_placement_groups, create_rollout_manager, create_training_models
from slime.utils.arguments import parse_args
from slime.utils.logging_utils import configure_logger
from slime.utils.tracking_utils import init_tracking


def train(args):
    configure_logger()
    # allocate the GPUs
    pgs = create_placement_groups(args)
    init_tracking(args)

    # create the rollout manager, with sglang engines inside.
    # need to initialize rollout manager first to calculate num_rollout
    rollout_manager, num_rollout_per_epoch = create_rollout_manager(args, pgs["rollout"])

    # create the actor and critic models
    actor_model, critic_model = create_training_models(args, pgs, rollout_manager)

    if args.offload_rollout:
        ray.get(rollout_manager.onload.remote(tags=[GPU_MEMORY_TYPE_WEIGHTS]))

    if args.offload_train and not args.enable_weights_backuper:
        actor_model.onload()
    # always update weight first so that sglang has the loaded weights from training.
    actor_model.update_weights()
    if args.offload_train and not args.enable_weights_backuper:
        actor_model.offload()

    if args.offload_rollout:
        if GPU_MEMORY_TYPE_CUDA_GRAPH is not None:
            ray.get(rollout_manager.onload.remote(tags=[GPU_MEMORY_TYPE_CUDA_GRAPH]))
        ray.get(rollout_manager.onload.remote(tags=[GPU_MEMORY_TYPE_KV_CACHE]))

    # special case for eval-only
    if args.num_rollout == 0 and args.eval_interval is not None:
        ray.get(rollout_manager.eval.remote(rollout_id=0))

    def offload_train():
        if args.offload_train:
            if args.use_critic:
                critic_model.offload()
                if rollout_id >= args.num_critic_only_steps:
                    actor_model.offload()
            else:
                actor_model.offload()
        else:
            actor_model.clear_memory()

    def onload_rollout():
        if args.offload_rollout:
            ray.get(rollout_manager.onload.remote(tags=[GPU_MEMORY_TYPE_WEIGHTS]))

    # train loop.
    # note that for async training, one can change the position of the sync operation(ray.get).
    for rollout_id in range(args.start_rollout_id, args.num_rollout):
        # TODO extract the duplicated eval logic
        if args.eval_interval is not None and rollout_id == 0:
            ray.get(rollout_manager.eval.remote(rollout_id))

        rollout_data_ref = ray.get(rollout_manager.generate.remote(rollout_id))

        if args.offload_rollout:
            ray.get(rollout_manager.offload.remote())

        if args.use_critic:
            critic_train_handle = critic_model.async_train(rollout_id, rollout_data_ref)
            if rollout_id >= args.num_critic_only_steps:
                ray.get(actor_model.async_train(rollout_id, rollout_data_ref))
            ray.get(critic_train_handle)
        else:
            ray.get(actor_model.async_train(rollout_id, rollout_data_ref))

        if args.save_interval is not None and (
            (rollout_id + 1) % args.save_interval == 0
            or (num_rollout_per_epoch is not None and (rollout_id + 1) % num_rollout_per_epoch == 0)
        ):
            if (not args.use_critic) or (rollout_id >= args.num_critic_only_steps):
                actor_model.save_model(rollout_id)
            if args.use_critic:
                critic_model.save_model(rollout_id)
            if args.rollout_global_dataset:
                ray.get(rollout_manager.save.remote(rollout_id))

        # Policy lag: only sync rollout engine weights every update_weights_interval steps.
        # When lag > 1, the rollout engine runs with stale weights; IS correction
        # (--use-rollout-logprobs + --use-tis) handles the resulting distribution shift.
        should_sync = (rollout_id - args.start_rollout_id + 1) % args.update_weights_interval == 0

        if args.enable_weights_backuper:
            offload_train()
            onload_rollout()
            if should_sync:
                actor_model.update_weights()
        else:
            actor_model.clear_memory()
            onload_rollout()
            if should_sync:
                actor_model.update_weights()
            offload_train()

        if args.offload_rollout:
            if GPU_MEMORY_TYPE_CUDA_GRAPH is not None:
                ray.get(rollout_manager.onload.remote(tags=[GPU_MEMORY_TYPE_CUDA_GRAPH]))
            ray.get(rollout_manager.onload.remote(tags=[GPU_MEMORY_TYPE_KV_CACHE]))

        if args.eval_interval is not None and (
            (rollout_id + 1) % args.eval_interval == 0
            or (num_rollout_per_epoch is not None and (rollout_id + 1) % num_rollout_per_epoch == 0)
        ):
            ray.get(rollout_manager.eval.remote(rollout_id))

    ray.get(rollout_manager.dispose.remote())


if __name__ == "__main__":
    args = parse_args()
    train(args)