tests : load backends before init when built with GGML_BACKEND_DL

[apollo-2006]

Fixes #4030.

ggml_backend_load_all() at the top of main() in the four tests that
initialise a context. Every example already does this:
examples/cli/cli.cpp, bench, quantize, lsp,
vad-speech-segments. The tests were the only callers that did not.

With GGML_BACKEND_DL=ON no backend is registered until something calls it,
so whisper_init_* and parakeet_init_* ran with devices = 0 and aborted
on GGML_ASSERT(device) in ggml_backend_dev_backend_reg.

No new include: ggml_backend_load_all() is already reachable through
whisper.h and parakeet.h, both of which pull in ggml-cpu.h, which
includes ggml-backend.h.

Before / after

Full suite on a -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON build,
macOS 26.6.2 / arm64, Apple clang 21.0.0:

test	before	after
test-whisper-zero-samples	aborted	passes
test-vad	aborted	passes
test-parakeet	aborted	passes
test-vad-full	aborted	reaches init, see below

ctest -L gh goes from 2 of 4 to 4 of 4. The full suite goes from three
aborts to none.

test-vad-full is labelled base;en rather than gh because it needs a real
models/ggml-base.en.bin, which I do not have locally. It had the same fault
and the fix is verified the same way: before, it printed devices = 0 and hit
GGML_ASSERT(device); after, it prints devices = 3 and gets through
initialisation. Substituting the stub model it then fails its own
assert(n_segments == 1), identically on a GGML_BACKEND_DL and a normal
build, which is the stub having no segments rather than anything to do with
this change.

Regression check

The same suite on a build without GGML_BACKEND_DL is unchanged; 14 of 15
before and after, the one failure being test-vad-full for the missing model
described above. Both configurations now produce identical results.

Why CI did not catch it

None of the workflows pairs the two. build-clang.yml, build-gcc.yml and
build-sanitize.yml run ctest -L gh but never set GGML_BACKEND_DL;
release.yml sets it but runs no tests. Adding a ctest step to a
GGML_BACKEND_DL job would cover the combination, but that felt like a
separate change from fixing the tests, so I have left it out here.

[Copilot]

🟢 Approval recommended

The changes are minimal and directly address the documented GGML_BACKEND_DL test abort by ensuring backends are registered before context initialization.

Pull request overview

Ensures tests that initialize a whisper_context / whisper_vad_context / parakeet_context explicitly load GGML backends first, matching the behavior of existing examples and preventing devices = 0 aborts when built with GGML_BACKEND_DL=ON.

Changes:

• Add ggml_backend_load_all() at the start of main() in four affected test binaries.
• Align test startup behavior with example binaries that already perform backend loading.
• Fixes aborts caused by no backends/devices being registered under dynamic-backend loading builds.

File summaries

File	Description
tests/test-whisper-zero-samples.cpp	Load GGML backends before initializing a Whisper context.
tests/test-vad.cpp	Load GGML backends before initializing VAD context/model usage.
tests/test-vad-full.cpp	Load GGML backends before Whisper/VAD initialization path.
tests/test-parakeet.cpp	Load GGML backends before initializing Parakeet contexts.

Review details

• Files reviewed: 4/4 changed files
• Comments generated: 0
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[apollo-2006]

@danbev one thing I could not settle from the source. test-vad is labelled unit and test-vad-full is base;en, so neither runs under ctest -L gh, while test-whisper-buffer-loader beside it is unit;gh. That puts two of the four tests this PR fixes outside the label CI actually runs, so a regression here would go unnoticed the same way this did.

Is that split deliberate, keeping anything that needs a model out of the gh job, or should test-vad pick up gh the way the buffer-loader test does?

[danbev]

Is that split deliberate, keeping anything that needs a model out of the gh job, or should test-vad pick up gh the way the buffer-loader test does?

There is a separate workflow, build-vad.yml which runs test-vad so it does not require the gh label.
And test-vad-full is left as a manual test as it requires a larger whisper model.