Every performance question this week ("can we batch by cores?", "why is dubbing
slow?") was answerable only by measuring, and twice the intuitive answer was wrong:
* Concurrency on Apple Silicon buys NOTHING. Measured, 4 segments:
1 worker 19.3s | 2 workers 20.7s (0.93x) | 3 workers 19.2s (1.00x)
One GPU, already saturated — extra workers interleave. Scaling the GPU pool by
free RAM (the "intelligent batching" that sounds obviously right) would have
added OOM risk on a 16 GB box for zero throughput. _pick_gpu_workers()'s
hardcoded `MPS -> 1` is correct, and now provably so.
* The clone-prompt cache misses on every segment (a dub writes one reference per
segment: 166 distinct keys, cache can never hit). That looked like the dub's
hidden cost. It is 0.40s/segment — ~2% — and it is not even waste: each
reference is genuinely different audio, and encoding it is the *feature*
(per-line prosody). Dropping to per-speaker refs would save ~65s/dub and cost
quality. Not a free win; not taken.
What actually dominates is TTS itself, which scales with text length (3.2s for a
short line, 8.7s for a 2.5x longer one) and is GPU-bound on a GPU that one
inference already fills.
The profiler is deliberately gentle with memory, because a profiler that OOMs the
machine reproduces the very bug class it exists to fix (#1119): stages run one at a
time, models are unloaded between them, a stage is SKIPPED if free RAM is under the
floor rather than starting a load the OS would kill, and each measurement is a fixed
small number of passes — no looping to convergence.
Co-authored-by: mergetest <nizam4103@gmail.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
