Skip to content

add support for nsys profile #667

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
Kipok merged 5 commits into from
Aug 14, 2025
Merged

add support for nsys profile #667

Show file tree
Hide file tree
Changes from 4 commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
a9ab333
add support for nsys profile
wedu-nvidia Aug 13, 2025
0ccc215
refractor logic
wedu-nvidia Aug 13, 2025
c78f1f7
Merge branch 'main' of github.com:NVIDIA/NeMo-Skills into wedu/nsys_p…
wedu-nvidia Aug 13, 2025
a70e02a
update nsys profiling for both sft and grpo
wedu-nvidia Aug 13, 2025
7a1a8d0
update
wedu-nvidia Aug 14, 2025
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
62 changes: 38 additions & 24 deletions nemo_skills/pipeline/nemo_rl/grpo.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,6 +31,8 @@
get_timeout,
resolve_mount_paths,
run_exp,
temporary_env_update,
get_nsight_cmd,
)
from nemo_skills.utils import get_logger_name, setup_logging

Expand All @@ -57,6 +59,7 @@ class NemoRLTask:
log_dir: str
env_variables: dict
backend: str
profile_step_range: str
extra_arguments: str = ""

def format_train_args(self):
Expand Down Expand Up @@ -94,10 +97,11 @@ def format_wandb_args(self):

def get_cmd(self):
self.logging_params = self.format_wandb_args()

nsight_cmd = get_nsight_cmd(self.profile_step_range)
cmd = (
f"export PYTHONPATH=$PYTHONPATH:/nemo_run/code:/opt/NeMo-RL && "
f"export UV_PROJECT=/opt/NeMo-RL && "
f"{nsight_cmd}"
f"echo 'Starting training' && "
f"uv run --active python /nemo_run/code/nemo_skills/training/nemo_rl/start_grpo.py "
f" {self.format_train_args()} "
Expand Down Expand Up @@ -125,6 +129,7 @@ def get_training_cmd(
log_dir,
env_variables,
backend,
profile_step_range,
):
timeout = get_timeout(cluster_config, partition)

Expand All @@ -144,6 +149,7 @@ def get_training_cmd(
log_dir=log_dir,
env_variables=env_variables,
backend=backend,
profile_step_range=profile_step_range,
)

return task.get_cmd()
Expand Down Expand Up @@ -197,6 +203,12 @@ def grpo_nemo_rl(
wandb_project: str = typer.Option("nemo-skills", help="Weights & Biases project name"),
wandb_group: str = typer.Option(None, help="Weights & Biases group name."),
disable_wandb: bool = typer.Option(False, help="Disable wandb logging"),
profile_step_range: str = typer.Option(
None,
help="Environment variable to control which training steps the profiler captures. "

@Kipok Kipok Aug 13, 2025

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
help="Environment variable to control which training steps the profiler captures. "
help="Controls which training steps the nsys profiler captures. "

"Format: START:STOP (1-indexed, STOP exclusive, same as slice syntax arr[start:stop]). "
"Example: '3:5' profiles steps 3 and 4 only. NOTE: START must be ≥ 1, so '0:10' is invalid."
),
partition: str = typer.Option(
None, help="Can specify if need interactive jobs or a specific non-default partition"
),
Expand Down Expand Up @@ -302,34 +314,36 @@ def grpo_nemo_rl(
log_dir=f"{log_dir}/training-logs",
env_variables=env_variables,
backend=backend,
profile_step_range=profile_step_range,
)

server_config = None
with get_exp(expname, cluster_config, _reuse_exp) as exp:
prev_task = _task_dependencies
for job_id in range(num_training_jobs):
prev_task = add_task(
exp,
cmd=train_cmd,
task_name=f'{expname}-grpo-{job_id}',
log_dir=f"{log_dir}/training-logs",
container=cluster_config["containers"]["nemo-rl"],
num_gpus=num_gpus,
num_nodes=num_nodes,
cluster_config=cluster_config,
server_config=server_config,
partition=partition,
time_min=time_min,
run_after=run_after,
reuse_code=reuse_code,
reuse_code_exp=reuse_code_exp,
task_dependencies=[prev_task] if prev_task is not None else None,
slurm_kwargs={"exclusive": exclusive} if exclusive else None,
heterogeneous=True if server_config is not None else False,
with_sandbox=with_sandbox,
with_ray=True,
installation_command=installation_command,
)
with temporary_env_update(cluster_config, {"RAY_LOG_SYNC_FREQUENCY": 20}):

@Kipok Kipok Aug 13, 2025

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

we should only do this if profile is specified. You can use {} as a second argument in other cases

@wedu-nvidia wedu-nvidia Aug 14, 2025

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done

for job_id in range(num_training_jobs):
prev_task = add_task(
exp,
cmd=train_cmd,
task_name=f'{expname}-grpo-{job_id}',
log_dir=f"{log_dir}/training-logs",
container=cluster_config["containers"]["nemo-rl"],
num_gpus=num_gpus,
num_nodes=num_nodes,
cluster_config=cluster_config,
server_config=server_config,
partition=partition,
time_min=time_min,
run_after=run_after,
reuse_code=reuse_code,
reuse_code_exp=reuse_code_exp,
task_dependencies=[prev_task] if prev_task is not None else None,
slurm_kwargs={"exclusive": exclusive} if exclusive else None,
heterogeneous=True if server_config is not None else False,
with_sandbox=with_sandbox,
with_ray=True,
installation_command=installation_command,
)

prev_task = add_task(
exp,
Expand Down
79 changes: 47 additions & 32 deletions nemo_skills/pipeline/nemo_rl/sft.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,6 +31,8 @@
get_timeout,
resolve_mount_paths,
run_exp,
temporary_env_update,
get_nsight_cmd,
)
from nemo_skills.utils import get_logger_name, setup_logging

Expand Down Expand Up @@ -59,6 +61,7 @@ class NemoRLTask:
log_dir: str
env_variables: dict
backend: str
profile_step_range: str
extra_arguments: str = ""

def format_train_args(self):
Expand Down Expand Up @@ -94,20 +97,22 @@ def format_wandb_args(self):

def get_cmd(self):
self.logging_params = self.format_wandb_args()

nsight_cmd = get_nsight_cmd(self.profile_step_range)
cmd = (
f"export PYTHONPATH=$PYTHONPATH:/nemo_run/code:/opt/NeMo-RL && "
f"export UV_PROJECT=/opt/NeMo-RL && "
f"echo 'Starting training' && "
f"NRL_FORCE_REBUILD_VENVS=true uv run --active python /nemo_run/code/nemo_skills/training/nemo_rl/start_sft.py "
f" {self.format_train_args()} "
f" {self.format_data_args()} "
f" {self.logging_params} "
f" {self.extra_arguments} "
"export PYTHONPATH=$PYTHONPATH:/nemo_run/code:/opt/NeMo-RL && "
"export UV_PROJECT=/opt/NeMo-RL && "
f"{nsight_cmd}"
"echo 'Starting training' && "
"NRL_FORCE_REBUILD_VENVS=true uv run --active "
"python /nemo_run/code/nemo_skills/training/nemo_rl/start_sft.py "
f"{self.format_train_args()} {self.format_data_args()} "
f"{self.logging_params} {self.extra_arguments}"
)

return cmd



def get_training_cmd(
cluster_config,
partition,
Expand All @@ -125,6 +130,7 @@ def get_training_cmd(
log_dir,
env_variables,
backend,
profile_step_range,
):
timeout = get_timeout(cluster_config, partition)

Expand All @@ -144,6 +150,7 @@ def get_training_cmd(
log_dir=log_dir,
env_variables=env_variables,
backend=backend,
profile_step_range=profile_step_range,
)

return task.get_cmd()
Expand Down Expand Up @@ -197,6 +204,12 @@ def sft_nemo_rl(
wandb_project: str = typer.Option("nemo-skills", help="Weights & Biases project name"),
wandb_group: str = typer.Option(None, help="Weights & Biases group name."),
disable_wandb: bool = typer.Option(False, help="Disable wandb logging"),
profile_step_range: str = typer.Option(

@Kipok Kipok Aug 13, 2025

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

same comments here as in the grpo file

@wedu-nvidia wedu-nvidia Aug 14, 2025

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done

None,
help="Environment variable to control which training steps the profiler captures. "
"Format: START:STOP (1-indexed, STOP exclusive, same as slice syntax arr[start:stop]). "
"Example: '3:5' profiles steps 3 and 4 only. NOTE: START must be ≥ 1, so '0:10' is invalid."
),
partition: str = typer.Option(
None, help="Can specify if need interactive jobs or a specific non-default partition"
),
Expand Down Expand Up @@ -301,34 +314,36 @@ def sft_nemo_rl(
log_dir=f"{log_dir}/training-logs",
env_variables=env_variables,
backend=backend,
profile_step_range=profile_step_range,
)

server_config = None
with get_exp(expname, cluster_config, _reuse_exp) as exp:
prev_task = _task_dependencies
for job_id in range(num_training_jobs):
prev_task = add_task(
exp,
cmd=train_cmd,
task_name=f'{expname}-sft-{job_id}',
log_dir=f"{log_dir}/training-logs",
container=cluster_config["containers"]["nemo-rl"],
num_gpus=num_gpus,
num_nodes=num_nodes,
cluster_config=cluster_config,
server_config=server_config,
partition=partition,
time_min=time_min,
run_after=run_after,
reuse_code=reuse_code,
reuse_code_exp=reuse_code_exp,
task_dependencies=[prev_task] if prev_task is not None else None,
slurm_kwargs={"exclusive": exclusive} if exclusive else None,
heterogeneous=True if server_config is not None else False,
with_sandbox=False,
with_ray=True,
installation_command=installation_command,
)
with temporary_env_update(cluster_config, {"RAY_LOG_SYNC_FREQUENCY": 20}):
for job_id in range(num_training_jobs):
prev_task = add_task(
exp,
cmd=train_cmd,
task_name=f'{expname}-sft-{job_id}',
log_dir=f"{log_dir}/training-logs",
container=cluster_config["containers"]["nemo-rl"],
num_gpus=num_gpus,
num_nodes=num_nodes,
cluster_config=cluster_config,
server_config=server_config,
partition=partition,
time_min=time_min,
run_after=run_after,
reuse_code=reuse_code,
reuse_code_exp=reuse_code_exp,
task_dependencies=[prev_task] if prev_task is not None else None,
slurm_kwargs={"exclusive": exclusive} if exclusive else None,
heterogeneous=True if server_config is not None else False,
with_sandbox=False,
with_ray=True,
installation_command=installation_command,
)

prev_task = add_task(
exp,
Expand Down
1 change: 1 addition & 0 deletions nemo_skills/pipeline/utils/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -38,6 +38,7 @@
get_exp_handles,
get_sandbox_command,
run_exp,
get_nsight_cmd,
)
from nemo_skills.pipeline.utils.generation import (
configure_client,
Expand Down
12 changes: 12 additions & 0 deletions nemo_skills/pipeline/utils/exp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -633,3 +633,15 @@ def get_exp(expname, cluster_config, _reuse_exp=None):
if cluster_config['executor'] == 'local':
return run.Experiment(expname, clean_mode=True)
return run.Experiment(expname, clean_mode=True, log_level="WARN")


def get_nsight_cmd(profile_step_range):
cmd = ''
if profile_step_range is not None:
cmd = (
f'export LD_LIBRARY_PATH="/usr/local/cuda/lib64:/usr/local/cuda/lib:/usr/local/nvidia/lib64:/usr/local/nvidia/lib:/usr/lib/x86_64-linux-gnu" && '
f"export NRL_NSYS_PROFILE_STEP_RANGE={profile_step_range} && "
'export NRL_NSYS_WORKER_PATTERNS="*policy*,*vllm*" && '

)
return cmd