Skip to content

[Refactor] Seperate colocate and disaggregated produce flow #1884

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
jayhenry merged 2 commits into from
Jun 11, 2026
Merged

[Refactor] Seperate colocate and disaggregated produce flow #1884

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
7012f5a
refactor rl production code separation
jayhenry Jun 11, 2026
52c0d8c
refine code
jayhenry Jun 11, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
492 changes: 492 additions & 0 deletions docs/design/sep_code.md
View file
Open in desktop

Large diffs are not rendered by default.

1,472 changes: 1,472 additions & 0 deletions docs/design/sep_code_demo.py
View file
Open in desktop

Large diffs are not rendered by default.

33 changes: 16 additions & 17 deletions examples/v1/config/rl_disagg_multi.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,15 +35,16 @@
from xtuner.v1.rl.agent_loop import SingleTurnAgentLoopConfig
from xtuner.v1.rl.agent_loop_manager import (
AgentLoopManagerConfig,
AsyncProduceStrategyConfig,
DisaggAsyncProduceStrategyConfig,
DisaggAgentLoopManagerConfig,
DisaggTaskSpecConfig,
SamplerConfig,
SyncProduceStrategyConfig,
TaskSpecConfig,
)
from xtuner.v1.rl.evaluator import EvaluatorConfig
from xtuner.v1.rl.judger import DapoMathJudgerConfig
from xtuner.v1.rl.loss import GRPOLossConfig
from xtuner.v1.rl.replay_buffer import SyncReplayBufferConfig
from xtuner.v1.rl.replay_buffer import AsyncReplayBufferConfig
from xtuner.v1.rl.rollout.worker import RolloutConfig
from xtuner.v1.rl.trainer import WorkerConfig
from xtuner.v1.rl.utils import AcceleratorResourcesConfig, get_eos_token
Expand Down Expand Up @@ -221,27 +222,25 @@
),
)

if over_sample_threshold > 0 or partial_rollout:
produce_strategy_config = AsyncProduceStrategyConfig(
over_sample_threshold=over_sample_threshold,
enable_partial_rollout=partial_rollout,
tail_batch_trigger_size=tail_batch_trigger_size,
max_staleness=max_staleness,
)
else:
produce_strategy_config = SyncProduceStrategyConfig()

agent_loop_manager_cfg = AgentLoopManagerConfig(
# 非共卡后台 producer 使用独立的 Disagg* config，不复用共卡 AsyncProduceStrategyConfig。
produce_strategy_config = DisaggAsyncProduceStrategyConfig(
over_sample_threshold=over_sample_threshold,
enable_partial_rollout=partial_rollout,
tail_batch_trigger_size=tail_batch_trigger_size,
max_staleness=max_staleness,
)

agent_loop_manager_cfg = DisaggAgentLoopManagerConfig(
tasks=[
TaskSpecConfig(
DisaggTaskSpecConfig(
task_name="train_task:dapo_math",
weight=dapo_task_weight,
agent_loop_config=dapo_train_agent_loop_config,
judger_config=judger_config,
produce_strategy_config=produce_strategy_config,
sampler_config=dapo_train_sampler_config,
),
TaskSpecConfig(
DisaggTaskSpecConfig(
task_name="train_task:gsm8k",
weight=gsm8k_task_weight,
agent_loop_config=gsm8k_train_agent_loop_config,
Expand Down Expand Up @@ -335,7 +334,7 @@ def compute_metric(samples):
train_worker_cfg=train_worker_cfg,
rollout_config=rollout_config,
tokenizer_path=model_path,
replay_buffer_config=SyncReplayBufferConfig(),
replay_buffer_config=AsyncReplayBufferConfig(),
agent_loop_manager_cfg=agent_loop_manager_cfg,
eval_agent_loop_manager_cfg=eval_agent_loop_manager_cfg,
evaluator_config=evaluator_config,
Expand Down
29 changes: 14 additions & 15 deletions examples/v1/config/rl_disagg_single.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -45,15 +45,16 @@
from xtuner.v1.rl.agent_loop import SingleTurnAgentLoopConfig
from xtuner.v1.rl.agent_loop_manager import (
AgentLoopManagerConfig,
AsyncProduceStrategyConfig,
DisaggAsyncProduceStrategyConfig,
DisaggAgentLoopManagerConfig,
DisaggTaskSpecConfig,
SamplerConfig,
SyncProduceStrategyConfig,
TaskSpecConfig,
)
from xtuner.v1.rl.evaluator import EvaluatorConfig
from xtuner.v1.rl.judger import GSM8KJudgerConfig
from xtuner.v1.rl.loss import GRPOLossConfig
from xtuner.v1.rl.replay_buffer import SyncReplayBufferConfig
from xtuner.v1.rl.replay_buffer import AsyncReplayBufferConfig
from xtuner.v1.rl.rollout.worker import RolloutConfig
from xtuner.v1.rl.trainer import WorkerConfig
from xtuner.v1.rl.utils import AcceleratorResourcesConfig
Expand Down Expand Up @@ -193,17 +194,15 @@
hf_checkpoint=model_path,
sample_params=training_sample_params,
)
if over_sample_threshold > 0 or partial_rollout:
produce_strategy_config = AsyncProduceStrategyConfig(
over_sample_threshold=over_sample_threshold,
enable_partial_rollout=partial_rollout,
tail_batch_trigger_size=tail_batch_trigger_size,
max_staleness=max_staleness,
)
else:
produce_strategy_config = SyncProduceStrategyConfig()
agent_loop_manager_cfg = AgentLoopManagerConfig(
tasks=TaskSpecConfig(
# 非共卡后台 producer 使用独立的 Disagg* config，不复用共卡 AsyncProduceStrategyConfig。
produce_strategy_config = DisaggAsyncProduceStrategyConfig(
over_sample_threshold=over_sample_threshold,
enable_partial_rollout=partial_rollout,
tail_batch_trigger_size=tail_batch_trigger_size,
max_staleness=max_staleness,
)
agent_loop_manager_cfg = DisaggAgentLoopManagerConfig(
tasks=DisaggTaskSpecConfig(
task_name="train_task",
agent_loop_config=agent_loop_config,
judger_config=judger_config,
Expand Down Expand Up @@ -258,7 +257,7 @@
train_worker_cfg=train_worker_cfg,
rollout_config=rollout_config,
tokenizer_path=model_path,
replay_buffer_config=SyncReplayBufferConfig(),
replay_buffer_config=AsyncReplayBufferConfig(),
agent_loop_manager_cfg=agent_loop_manager_cfg,
eval_agent_loop_manager_cfg=eval_agent_loop_manager_cfg,
evaluator_config=evaluator_config,
Expand Down
Loading
Loading