0 0 110 C++

perf(gigaam): speed up inference (#132)

* perf(gigaam): batch host decoder projections

* perf(gigaam): parallelize mel extraction and simplify flash attention

* honor thread budget in mel frontend

* modify tolerances

---------

Co-authored-by: CJ Pais <cj@cjpais.com>

Ilya Gulya committed 1mo ago

8b289ff729b6b4617effcae558930f5d641756af

Parent: 923d4a0

Committed by GitHub <noreply@github.com> on 8/19/2026, 5:56:45 AM