fix: Silero VAD must not leave the global backend at one thread

[localai-org-maint-bot]

Problem

Silero VAD computes each 32 ms chunk with run_graph(0, 1, ...). When --threads is not given, run_graph wrote that count of 1 to the process-global backend and kept it. Every later graph then ran on one thread, including the ASR decode after a VAD pass. This affects parakeet-cli transcribe --vad --vad-model silero.gguf without --threads, and any host that does not set a thread count.

Fix

A per-call n_threads in run_graph now applies to that call only: the previous backend count is restored after the compute. Silero still runs its tiny graphs on one thread. A positive global override (--threads) still wins, and leased pool backends are unaffected. pk::backend_thread_count() is added so tests can read the count.

The only other caller that passes a positive count on the global path and is not a test is the CTC head (4 threads). It had the same side effect and is now scoped too. On a 67 s clip with the 110M hybrid model and --decoder ctc, wall time was 0.91 to 0.95 s before and 0.94 to 0.98 s after, which is within the noise of that machine.

Measurement

Build: CPU only, Release. Model: 110M hybrid, Q8_0. Input: a 66.9 s speech clip (a 7.4 s LibriSpeech utterance repeated). Silero model: F16. Wall time of the whole transcribe command, 3 runs interleaved. The machine was not quiet (load average about 9 to 11), so these are not measurements from a quiet machine; the gap between before and after is much larger than the noise.

Backend thread count after the Silero pass, --threads unset: 1 before the fix, 8 after.

Run	Before (s, RTF)	After (s, RTF)
no VAD	1.05, 1.10, 1.13 (0.016 to 0.017)	0.97, 1.00, 1.16 (0.014 to 0.017)
--vad, no --threads	2.78, 2.78, 2.83 (0.042)	0.79, 0.85, 0.91 (0.012 to 0.014)
--vad --threads 8	1.00, 0.98, 0.94 (0.015)	0.82, 0.94, 0.95 (0.012 to 0.014)

Before the fix, --vad without --threads took about 2.7 times as long as the same run with --threads 8, and the process stayed at about 100 percent CPU. After the fix the two match.

Tests

• New test_run_graph_threads (no model needed): checks the backend count across per-call counts and, when PARAKEET_TEST_SILERO_GGUF is set, across a Silero pass. It fails on the old code (3 failures) and passes now.
• ctest -LE model: 35 of 35 pass.
• test_silero_vad, test_silero_load_negative, test_silero_framer, test_capi_vad_silero, test_transcribe_vad_silero, test_vad_options: pass.

Limits

• The Ultra and Redux head VAD tests were not run: those models were not available. The head path computes through the fused encoder with a count of 0 and does not use a per-call count, so it does not trigger this bug.
• Only CPU was measured.

🤖 Generated with Claude Code