[Bug] shipped dynamic-sampling filters crash on fan-out groups (list[list[Sample]]) from multi-turn agent rollouts

[kugua2025aigc-code]

Summary

The shipped dynamic-sampling filters assume groups are list[Sample], but multi-turn agent rollouts fan out into list[list[Sample]] — a shape that generate_and_rm_group itself documents and that every other consumer of the group handles explicitly. Enabling --dynamic-sampling-filter-path with any custom-generate function that returns a list of samples crashes the rollout within the first training step:

AttributeError: 'list' object has no attribute 'get_reward_value'

Evidence

1. The shipped filter calls sample.get_reward_value(args) for each element of samples, assuming samples: list[Sample]:
   • slime/rollout/filter_hub/dynamic_sampling_filters.py:9-11
2. generate_and_rm_group documents both shapes (emphasis in the original docstring): "the group is list[Sample] for plain rollouts and list[list[Sample]] for the fan-out case":
   • slime/rollout/sglang_rollout.py:329-338
3. The rollout loop itself defends against both shapes everywhere else — first-sample logging and final sorting both branch on isinstance(group[0], list):
   • slime/rollout/sglang_rollout.py:466-470, 494-497
4. But call_dynamic_filter(dynamic_filter, args, group) at sglang_rollout.py:469 passes the raw group into the filter, which has no such branch.

Reproduction

Any rollout whose custom-generate function emits a list per trajectory (multi-turn agent fan-out, e.g. the pattern in examples/coding_agent_rl), with:

--dynamic-sampling-filter-path \
  slime.rollout.filter_hub.dynamic_sampling_filters.check_reward_nonzero_std_with_fallback

Minimal Python repro (no GPU):

import types
from slime.rollout.filter_hub.dynamic_sampling_filters import check_reward_nonzero_std
from slime.utils.types import Sample

args = types.SimpleNamespace(reward_key=None)
group = [[Sample(prompt="p", label="l", reward=0.2)],   # fan-out: list[list[Sample]]
         [Sample(prompt="p", label="l", reward=1.0)]]
check_reward_nonzero_std(args, group)
# AttributeError: 'list' object has no attribute 'get_reward_value'

Observed on slime main @ a3f50097, single-GPU H100, agentic GRPO run — reproduced at both N_SAMPLES=4 and N_SAMPLES=8.

Suggested fix

Handle the fan-out shape in the shipped filters, mirroring the pattern already used by the rollout loop:

def _entry_reward(args, entry):
    # fan-out fragments of one trajectory share the trajectory reward
    return entry[0].get_reward_value(args) if isinstance(entry, list) else entry.get_reward_value(args)

def check_reward_nonzero_std(args, samples: list, **kwargs):
    rewards = [_entry_reward(args, s) for s in samples]
    ...

We shipped a fan-out-safe variant locally as a workaround and are happy to send a PR if the maintainers agree with the approach (fragments of one trajectory sharing entry[0]'s reward).

Related

• [Bug] save_model fires unconditionally at the final step regardless of --save-interval; final optimizer-state write can kill the job #2370 (same "agentic/multi-turn path is under-tested" family: final-step unconditional save)