Summary

Two recipes under src/megatron/bridge/recipes/ have no unit-test coverage. Add them.

• qwen2_audio.qwen2_audio_7b_finetune_config — 18 tests
• kimi_vl.kimi_k25_vl_sft_config (and _get_kimi_k25_vl_pipeline_layout) — 16 tests

Both modules monkeypatch AutoBridge to bypass HuggingFace Hub I/O and assert the recipe's documented contract end-to-end. Pattern mirrors the existing test_kimi_k2.py and test_glm_45v_recipes.py.

Refs #3177.

Why this matters

The original issue notes that Kimi K2 had a recipe but no recipe test, which let an MCore-side API breaking change land undetected. Kimi K2 itself has since gained tests/unit_tests/recipes/kimi/test_kimi_k2.py, but a sweep of src/megatron/bridge/recipes/ vs tests/unit_tests/recipes/ still surfaces several recipe modules with no recipe-test counterpart. This PR closes the easiest two, both of which are entirely missing today.

What's covered

test_qwen2_audio_recipes.py

• ConfigContainer shape, default training cadence (train_iters=2000, GBS=32, MBS=1), validation settings.
• Default parallelism (TP=1, PP=1), seq_length propagation into both model_cfg and the HFDatasetConversationProvider.
• Dataset wiring: make_default_audio_dataset maker, default train / dev split selection, processor path.
• Default freeze flags (language model, audio model, audio projection) and override propagation.
• Full-SFT vs PEFT lr selection (the entry point's branch): peft=None and peft=\"none\" → lr=5e-6; peft=\"lora\" → lr=1e-4; peft=\"dora\" attaches a DoRA instance; unknown PEFT scheme raises a clear ValueError.
• User-supplied finetune_lr overrides the default.
• Parallelism / seq_length overrides apply.
• DDP defaults (no overlap, fp32 grad reduce, dist optimizer on, optim+grads+params sharding), checkpoint format (torch_dist, parallel save), RNG seed, val/test maker overrides.

test_kimi_k25_vl.py

• _get_kimi_k25_vl_pipeline_layout helper:
  • PP=1, VP=1 returns None.
  • PP=16, VP=1 (the SFT default) matches the documented stage breakdown.
  • All 7 known (PP, VP) combinations return non-empty list-of-lists.
  • vp_size=None is normalized to 1.
  • Unknown combinations raise ValueError.
  • Each call returns a fresh copy (mutation-safe).
• kimi_k25_vl_sft_config:
  • ConfigContainer shape, default training cadence, validation cadence.
  • Default parallelism (TP=2, PP=16, EP=32, SP on, expert TP=1).
  • Full activation recompute (uniform, 1 layer), pipeline split flags, layout matches helper output.
  • Muon-compatible DDP wiring (use_distributed_optimizer=False, overlap_param_gather=False, no_shard strategy).
  • HF processor path, NullTokenizer with vocab_size pulled from the provider.
  • bf16 mixed precision, fp32 optimizer precision, dist_muon selected.
  • MoE wiring (alltoall + deepep flex backend, fused permute, grouped GEMM, shared-expert overlap on).
  • TE backend with CUDA graphs disabled by default, kernel selections (cross_entropy_fusion_impl=\"te\").
  • Comm overlap defaults (tp_comm_overlap=False, delay_wgrad_compute=False).
  • Checkpoint cadence and async-save default.
  • Rope-fusion gating contract: when model.apply_rope_fusion=False the experimental dist flag stays False; when True the recipe flips cfg.dist.enable_megatron_core_experimental=True.

Test plan

• python3 -m ast parse of new files
• ruff check clean
• ruff format --check clean
• CI: L0 unit tests pick up the new modules automatically (no workflow file changes needed).

Risk

Zero — tests only. No production code changes.

Notes for reviewers

• Both stub providers expose only the attribute surface the recipe touches, so adding new attribute reads/writes in either recipe will surface as AttributeError here, matching the existing test_kimi_k2.py style.
• qwen2_audio reuses the existing HFDatasetConversationProvider dataclass, which is import-safe (no HF I/O at construction).
• The kimi_k25_vl test's rope-fusion-on case overrides the bridge inside the test to flip apply_rope_fusion=True, exercising the recipe's if cfg.model.apply_rope_fusion: branch.