Skip to content

log all trajectory repetitions per training step (not just first) #1388

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

log all trajectory repetitions per training step (not just first) #1388

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
1d86728
log all trajectory repetitions per training step (not just first)
arteemg Mar 25, 2026
dd76564
update for step-wise training mode
arteemg Mar 25, 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
26 changes: 18 additions & 8 deletions skyrl/train/trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -248,14 +248,24 @@ async def train(self):
with Timer("postprocess_generator_output", self.all_timings):
generator_output = self.postprocess_generator_output(generator_output, uids)

# 2. print example just for debugging
vis = self.tokenizer.decode(generator_output["response_ids"][0])
log_example(
logger,
prompt=generator_input["prompts"][0],
response=vis,
reward=generator_output["rewards"][0],
)
# 2. print all trajectories for debugging
# Prefer trajectory_ids from generator_output (step-wise mode expands them per step),
# fall back to generator_input for standard mode.
trajectory_ids = generator_output.get("trajectory_ids") or generator_input.get("trajectory_ids", [])
input_prompts = generator_input.get("prompts", [])
for i, (response_ids, reward) in enumerate(
zip(generator_output["response_ids"], generator_output["rewards"])
):
vis = self.tokenizer.decode(response_ids)
tid = trajectory_ids[i] if i < len(trajectory_ids) else None
tid_str = f"instance={tid.instance_id}, rep={tid.repetition_id}" if tid else f"idx={i}"
prompt = input_prompts[i] if i < len(input_prompts) else "<prompt unavailable>"
log_example(
logger,
prompt=prompt,
response=f"[Trajectory {tid_str}]\n{vis}",
reward=reward,
)

# 3. Convert GeneratorOutput to TrainingInputBatch
with Timer("convert_to_training_input", self.all_timings):
Expand Down
Loading