Skip to content

fix: don't cast Silero VAD to fp16 (MPSGraph abort with --fp16 --vad) #32

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
anubissbe wants to merge 2 commits into
base: main
Choose a base branch
from
Open

fix: don't cast Silero VAD to fp16 (MPSGraph abort with --fp16 --vad) #32

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
f66ed42
fix: don't cast Silero VAD to fp16
anubissbe Jul 21, 2026
bb82249
fix: correct stale class reference in VAD comment
anubissbe Jul 21, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
16 changes: 12 additions & 4 deletions include/parakeet/api/transcribe.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -492,8 +492,12 @@ class Transcriber {
/// Enable VAD preprocessing. Call after to_half()/to_gpu().
void enable_vad(const std::string &vad_weights_path) {
vad_ = std::make_unique<audio::SileroVAD>(vad_weights_path);
if (use_fp16_)
vad_->to_half();
// The VAD deliberately stays fp32 even when the ASR model is fp16.
// SileroVAD::to_half() casts the weights but leaves the context
// buffer, LSTM state and audio input fp32, which aborts the process
// under MPSGraph ('mps.add' requires the same element type). The VAD
// is a ~1.2 MB preprocessor that only emits segment boundaries, so
// running it in fp32 costs almost nothing.
if (use_gpu_)
vad_->to_gpu();
}
Expand Down Expand Up @@ -841,8 +845,12 @@ class TDTTranscriber {
/// Enable VAD preprocessing. Call after to_half()/to_gpu().
void enable_vad(const std::string &vad_weights_path) {
vad_ = std::make_unique<audio::SileroVAD>(vad_weights_path);
if (use_fp16_)
vad_->to_half();
// The VAD deliberately stays fp32 even when the ASR model is fp16.
// SileroVAD::to_half() casts the weights but leaves the context
// buffer, LSTM state and audio input fp32, which aborts the process
// under MPSGraph ('mps.add' requires the same element type). The VAD
// is a ~1.2 MB preprocessor that only emits segment boundaries, so
// running it in fp32 costs almost nothing.
if (use_gpu_)
vad_->to_gpu();
}
Expand Down
4 changes: 2 additions & 2 deletions src/api/diarize.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -75,8 +75,8 @@ void DiarizedTranscriber::to_half() {

void DiarizedTranscriber::enable_vad(const std::string &vad_weights_path) {
vad_ = std::make_unique<audio::SileroVAD>(vad_weights_path);
if (use_fp16_)
vad_->to_half();
// See Transcriber::enable_vad — casting the VAD to fp16 mixes an fp32
// context buffer with fp16 weights and aborts under MPSGraph.
if (use_gpu_)
vad_->to_gpu();
}
Expand Down