[CI] guard initial actor/ref KL checks under R3 (routing replay)

[aoshen02]

What

Backports slime #1987 + #1990 (one logical fix, split across two upstream commits because #1990 amends #1987's lines and adds the matching model.py guard).

Under --use-routing-replay (R3), the actor forward replays rollout routing while the reference forward intentionally falls through to normal routing. Their log-probs / KL are therefore not expected to match bit-for-bit at the initial step. But our CI asserts assume exact zero-KL:

• data.py: at rollout_id == 0, assert abs(log_probs - ref_log_probs) < 1e-8
• model.py: at step 0, assert train/kl_loss < 1e-8

So any routing-replay + --ci-test run raises a false failure. vime exercises this path today (e.g. the Qwen3-30B-A3B test passes --use-routing-replay).

Fix

• data.py: gate the rollout_id == 0 log_probs-vs-ref_log_probs assert on not use_rollout_routing_replay and not ci_disable_kl_checker (both via getattr default-False for safety).
• model.py: gate the step-0 train/kl_loss assert on not use_rollout_routing_replay.

Non-R3 runs are unaffected (the asserts still fire exactly as before).

Decoupling

Independent of #92 (GLM bridge) and the upcoming logprob_abs_diff assert (slime #1968) / critic-temperature (slime #1928) PRs — no shared lines.

Test

ast.parse clean on both files. Logic change is guard-only; non-R3 CI behavior identical.

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request adds guards to bypass initial actor/reference zero-KL and KL loss checks in CI when routing replay is enabled, as the actor and reference paths route differently under this mode. The feedback suggests checking the base flag use_routing_replay instead of use_rollout_routing_replay in both data.py and model.py to ensure that tests running with only --use-routing-replay do not trigger false assertion failures.

[gemini-code-assist]

The PR description mentions that the false failures occur under --use-routing-replay (R3), which is exercised by tests like Qwen3-30B-A3B. However, the guard here checks use_rollout_routing_replay instead of use_routing_replay.

If a test is run with --use-routing-replay but without --use-rollout-routing-replay, use_rollout_routing_replay will be False, causing this guard to evaluate to True and the assertion to still run and fail.

Since use_routing_replay is the base flag and is automatically set to True when use_rollout_routing_replay is enabled (via args.use_routing_replay = True in slime/utils/arguments.py), checking use_routing_replay instead will correctly cover both routing replay modes.

Suggested change

	if (
	rollout_id == 0
	and not getattr(args, "ci_disable_kl_checker", False)
	and not getattr(args, "use_rollout_routing_replay", False)
	and "rollout/log_probs" in reduced_log_dict
	and "rollout/ref_log_probs" in reduced_log_dict
	):
	if (
	rollout_id == 0
	and not getattr(args, "ci_disable_kl_checker", False)
	and not getattr(args, "use_routing_replay", False)
	and "rollout/log_probs" in reduced_log_dict
	and "rollout/ref_log_probs" in reduced_log_dict
	):

[gemini-code-assist]

Similar to the change in data.py, this guard should check use_routing_replay instead of use_rollout_routing_replay to ensure that tests running with --use-routing-replay alone do not trigger false failures.

Suggested change

	if (
	accumulated_step_id == 0
	and not getattr(args, "use_rollout_routing_replay", False)
	and "train/kl_loss" in log_dict
	):
	if (
	accumulated_step_id == 0
	and not getattr(args, "use_routing_replay", False)
	and "train/kl_loss" in log_dict
	):