Skip to content

feature(tj): add MoE expert selection stats and gradient conflict monitoring #470

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

Open
tAnGjIa520 wants to merge 2 commits into
base: main
Choose a base branch
from
Open

feature(tj): add MoE expert selection stats and gradient conflict monitoring #470

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
244808e
feat(tj): add MoE expert selection stats and gradient conflict monito…
tAnGjIa520 Feb 3, 2026
8604002
feature(tj): add MoE gradient conflict experiment reproduction README…
tAnGjIa520 Feb 12, 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
17 changes: 16 additions & 1 deletion lzero/entry/train_unizero_multitask_segment_ddp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,6 +26,11 @@
symlog,
inv_symlog,
)
from lzero.entry.utils import (
collect_and_log_moe_statistics,
TemperatureScheduler,
log_buffer_memory_usage
)
# NOTE: The following imports are for type hinting purposes.
# The actual GameBuffer is selected dynamically based on the policy type.
from lzero.mcts import UniZeroGameBuffer
Expand All @@ -52,7 +57,8 @@ def train_unizero_multitask_segment_ddp(
model_path: Optional[str] = None,
max_train_iter: Optional[int] = int(1e10),
max_env_step: Optional[int] = int(1e10),
benchmark_name: str = "atari"
benchmark_name: str = "atari",
cal_moe_profile: bool = True
) -> 'Policy':
"""
Overview:
Expand All @@ -68,6 +74,7 @@ def train_unizero_multitask_segment_ddp(
- max_train_iter (:obj:`Optional[int]`): The maximum number of policy update iterations during training.
- max_env_step (:obj:`Optional[int]`): The maximum number of environment interaction steps to collect.
- benchmark_name (:obj:`str`): The name of the benchmark, e.g., "atari" or "dmc".
- cal_moe_profile (:obj:`bool`, optional): Whether to enable MoE expert selection statistics and heatmap logging. Default is True.

Returns:
- policy (:obj:`Policy`): The converged policy.
Expand Down Expand Up @@ -201,6 +208,8 @@ def train_unizero_multitask_segment_ddp(
# Create a TensorBoard logger.
log_dir = os.path.join('./{}/log'.format(cfg.exp_name), f'serial_rank_{rank}')
tb_logger = SummaryWriter(log_dir)
# Inject TensorBoard logger into policy for gradient conflict and MoE statistics logging
policy.logger = tb_logger

# Create the shared learner.
learner = BaseLearner(cfg.policy.learn.learner, policy.learn_mode, tb_logger, exp_name=cfg.exp_name)
Expand Down Expand Up @@ -498,6 +507,12 @@ def train_unizero_multitask_segment_ddp(
# DDP automatically synchronizes gradients and parameters during training.
log_vars = learner.train(train_data_multi_task, envstep_multi_task, policy_kwargs=learn_kwargs)

if cal_moe_profile and cfg.policy.model.world_model_cfg.multiplication_moe_in_transformer and cfg.policy.model.world_model_cfg.num_experts_of_moe_in_transformer:
# Collect and log MoE expert selection statistics (heatmaps, distribution divergence)
moe_log_interval = getattr(cfg.policy, 'moe_log_interval', 1)
if learner.train_iter % moe_log_interval == 0:
collect_and_log_moe_statistics(policy, tb_logger, learner.train_iter, world_size, rank)

# Check if task_exploitation_weight needs to be calculated.
if i == 0:
# Calculate task weights.
Expand Down
Loading