[Sync][C] CI/docs/mechanical sweep (slime #1939/#1940/#1942/#1943/#1980/#1967/#1938/#1988)

[aoshen02]

Mega-PR C — CI / docs / mechanical sweep

Part of the #107 slime→vime sync (mega-PR C, A-independent slice). All cpu-validatable; the GitHub e2e-test-plugin-contracts job (ubuntu-latest) is the cpu gate.

Ported

slime PR	what	files
#1939 + #1940	reward-fn cpu tests + cpu-on-main push trigger	pr-test.yml(.j2), tests/test_rm_*.py (5), vime/utils/misc.py (lazy ray)
#1942	sync customization/trace/pd-disagg docs	docs/{en,zh}/... (6)
#1943	remove redundant replay tool + harden unwrap_model import-compat	tools/replay_openai_jsonl.py (−535), model.py
#1980	FLOPs accounting fix for non-MLA attention (gate on multi_latent_attention)	vime/utils/flops_utils.py
#1967	PYTHONBUFFERED=16 → PYTHONUNBUFFERED=1 (name and value)	12 scripts + command_utils.py
#1938	guard vllm_speculative_config read under --debug-train-only (getattr)	vime/ray/rollout.py
#1988	rewrite fault-tolerance doc (rebranded sglang→vllm, /health_generate→/health, link vllm-config.md)	docs/en/advanced/fault-tolerance.md

Dropped after source review (evidence for #107 disposition)

• #1987 / #1990 (R3 ci ref/logprob + kl guards) — already in main via vime [CI] guard initial actor/ref KL checks under R3 (routing replay) #93 (b1009dd "Backports slime #1987 + #1990"); identical two-file, two-guard change.
• #1974 / #1989 — touch examples/coding_agent_rl + its README bullet; that feature isn't in vime yet → move to the agent mega-PR.
• #1975 — sglang conda-ci image resolution + a test vime already has (test_qwen3_0.6B_parallel_check.py) + a build_conda.sh micromamba fix; no real version-file port for vime → N/A.

Deferred to after the rollout-data-model PR

• #1986 / #1985 — they edit that PR's test files (dp_schedule / cp_utils / fanout / rm); will land when rebased on the merged data-model PR.
• #1939 tests/test_sample.py — the other 7 files of #1939 (5 reward-fn cpu tests + lazy-ray + cpu-on-main trigger) landed here; test_sample.py constructs Sample(rollout_id=...), and Sample.rollout_id only exists after the data-model refactor (slime #1933/#1984), so this one file defers with it. (Drop was previously undocumented; recorded here.)
• customization.md "Returning multiple training samples" section — removed in c8a0b99. It came from slime #1961 (not #1942, despite the earlier commit message), documents list[Sample] return + group_id + Sample.rollout_id, depends on the same data-model refactor (vime code has 0 occurrences of group_id/rollout_id), and is out of mega-C scope. Re-add with the data-model PR.

Validation

• py_compile clean (flops_utils / rollout / command_utils); pr-test.yml re-rendered from .j2 — no drift.
• codex review VALID (no bugs; MLA q_lora_rank=None fallback confirmed against slime's moonlight.sh).
• rm cpu tests are unchanged from the previously-green [CI] add reward-fn cpu tests + cpu-on-main trigger (port slime #1939+#1940) #129 (107/107 cpu); local venv lacks pytest/torch so the authoritative cpu run is the GitHub job on this PR.

🤖 Generated with Claude Code

Corrections (c8a0b99)

• #1938 rollout.py: dropped the extra 4-line comment — slime #1938 is a pure one-line getattr change.
• #1943 model.py: moved the unwrap_model try/except to slime's exact position (after tqdm) — import block now byte-identical to slime.

---

Fidelity ledger (audited 2026-06-04)

op	detail
Ported	#1938/1939/1940/1942/1943/1967/1980/1988 — CI/lint/docs cluster + test_rm_×5 (deepscaler/f1/gpqa/math/math_dapo) wired into cpu job
Translated	none (no sglang seam)
Dropped/N-A	docs/version-only bits
Verified	test_rm_×5 PASS on rebuilt stack (gb200 gate_full 8914)

[gemini-code-assist]

Code Review

This pull request rebrands the project from 'slime' to 'vime' across documentation, adds comprehensive CPU unit tests for various reward models (deepscaler, f1, gpqa, math, and dapo), and deletes the OpenAI replay tool. It also introduces several code improvements, such as lazy-importing Ray to support CPU-only execution, refactoring FLOPs calculation for Multi-Latent Attention (MLA), correcting the environment variable to PYTHONUNBUFFERED=1, and adding a guard for speculative decoding metrics. A review comment correctly points out a potential ZeroDivisionError in _compute_spec_metrics if all_samples is empty, which should be addressed by extending the guard condition.

[gemini-code-assist]

If all_samples is empty, num_samples will be 0, which leads to a ZeroDivisionError when calculating spec_accept_rate and spec_accept_length. Adding a check for not all_samples in the guard condition prevents this potential crash.

Suggested change

	# Guard against --debug-train-only / --load-debug-rollout-data, where the
	# vllm arg-parse phase is skipped (skip_vllm) and vllm_speculative_config is
	# never set on args. Mirrors slime #1938. getattr(..., None) keeps this safe
	# whether or not the vllm parse ran.
	if getattr(args, "vllm_speculative_config", None) is None:
	# Guard against --debug-train-only / --load-debug-rollout-data, where the
	# vllm arg-parse phase is skipped (skip_vllm) and vllm_speculative_config is
	# never set on args. Mirrors slime #1938. getattr(..., None) keeps this safe
	# whether or not the vllm parse ran.
	if getattr(args, "vllm_speculative_config", None) is None or not all_samples: