Default ASR to Granite Speech TurboCTC, and raise transformers to 5.16

[aviv1ron1]

Makes Granite Speech TurboCTC the default ASR backend for the audio cascade, and raises transformers to the release that ships its architecture.

Four commits, no change to the decoder, switch or LoRA paths.

1. TurboCTC becomes the ASR default (ef43e69)

DEFAULT_ASR_MODEL_ID moves from distil-whisper/distil-small.en to ibm-granite/granite-speech-5.0-470m-turboctc — a 470M English conformer CTC encoder.

Being CTC rather than generative changes three things that callers can see, all documented in docs/AUDIO.md:

• No generate() to steer. Decode kwargs (language, task) are dropped rather than forwarded, so the per-request language override is inert on the default model. A multilingual generative backend is still one --asr-model away.
• Transcripts are lowercase and unpunctuated (what is the capital of israel). They are spliced into the prompt as ordinary text, so the model reads them that way.
• Defaults retuned for a fixed-window encoder: asr_self_chunks=False and asr_chunk_length_s=120.0, so clips up to two minutes reach the backend whole and only longer ones are split by our own chunker. The HF pipeline's own CTC chunking is deliberately never used — it rescales stride by inputs_to_logits_ratio, which this checkpoint does not publish, so it would silently fall back to 1 and trim every seam at the wrong offset.

asr_device now defaults to cuda and asr_dtype resolves to bfloat16 there — the checkpoint's own dtype, and it keeps float32's exponent range, which suits an encoder carrying BatchNorm in every conv block.

Dependency change: transformers>=5.5.1,<5.17.0 in the core package, and transformers>=5.16.0 on the audio extra — 5.16 is where granite_speech5_ctc landed. The requirement sits on the extra so the rest of the package still works on an older release. uv.lock regenerated: transformers 5.8.1 → 5.16.1.

Two consequences of that bump are handled in the same commit:

• ATTENTION_LAYER_TYPES now accepts both attention and full_attention. transformers 5.16 renamed the layer-type value and rewrites it inside PreTrainedConfig.__init__, so comparing against the bare string silently dropped the attention LoRA target groups and left adapters with MLP targets only.
• IsHybrid is no longer declared. It requires mamba-state hooks this model neither implements nor needs — the composer normalizes every layer to attention, so a composed checkpoint has zero mamba layers. vLLM's escape hatch for exactly that shape compares the literal "attention", so the rename flipped is_hybrid true and engine init began failing for a mamba dtype hook. Dropping the marker is the fix rather than implementing the hooks: several other vLLM behaviours gate on is_hybrid and none should apply to a pure-attention model.

2. A test helper stops leaking a mock (7b7faa6) — closes #121

_patched_pipeline in tests/unit/test_asr.py patched transformers.pipelines.pipeline before transformers.pipeline. mock.patch.__enter__ records the current value to restore later, and transformers is a lazy module: resolving transformers.pipeline when it is not yet cached goes through transformers.pipelines. With the submodule patched first, the second patch read back the mock, recorded it as "the original", and faithfully restored it on exit — leaving the mock installed for the rest of the process.

Swapping the two lines is the whole fix. TestPatchedPipelineRestores guards both attributes and forces the lazy-resolve precondition rather than relying on collection order; it was verified to fail on the old ordering, so it is a real guard rather than a tautology.

3. Supply vLLM's missing layer-type key (8a07b33) — closes #122

vLLM ≤0.25 keys its Granite-hybrid layer table on the pre-5.16 spelling:

ALL_DECODER_LAYER_TYPES = {"attention": ..., "mamba": ...}
layer_class = ALL_DECODER_LAYER_TYPES[config.layer_types[layer_idx]]   # KeyError: 'full_attention'

Since transformers ≥5.16 can only produce the new spelling, register() now aliases it onto the same layer class. Two things this repairs:

• The equivalence tier's reference model. Those tests build their upstream comparison with transformers, save_pretrained it, and hand the directory to vllm.LLM; the saved config carries the new spelling and its architectures field routes to vLLM's own class, so the reference could not load. GraniteSwitchForCausalLM was never affected — it builds layers through a closure and does not consult that table.
• Serving an un-composed Granite 4.0 hybrid checkpoint. In a venv built from this branch the pairing is transformers>=5.16 (from our audio extra) with vLLM pinned ≤0.25, so vllm serve ibm-granite/granite-4.0-micro raised. Composed Granite Switch checkpoints did not.

Why it lives in register(): the lookup runs in vLLM's spawned engine-core process, so nothing patched in the parent survives. vLLM loads its plugins in that process during init, reading the vllm.general_plugins entry point we already declare — it is the only hook we own that executes there.

Aliasing adds a name, not behaviour: both keys resolve to the identical class object. setdefault makes the block a no-op on vLLM 0.26+, where upstream added the key themselves, so it can be deleted on a version bump without touching anything else.

4. Audio support for the Granite MoE architecture

No source code changes, only documentration and tests.

• docs/AUDIO.md — the audio cascade over a pure sparse MoE (granitemoe) base. Hand-merged rather than overwritten, since this file had already diverged here.
• tests/composer/test_audio_marker_output_row.py — the marker/reserved-row copy is now parametrized over MLP topology as well as embedding tying, 2 cases → 4 (untied/tied × dense/sparse-MoE). The fixup only ever touches embedding rows, so both topologies must behave identically; the sparse arm is what would notice a shared-MLP-shaped assumption creeping into the model construction it runs against.
• tests/composer/test_granitemoe_audio_compose.py — new, 9 tests: audio does not resurrect the shared MLP, does not widen the adapter surface and leaves no zero-width parameter; the control LUT goes stale on the marker and refresh fixes it idempotently and passes the validator; the marker output row; and survival across save/load.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!