Skip to content

fix(kokoro): export Prosody with fp32 compute (FluidAudio #947) #108

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
Alex-Wengg wants to merge 1 commit into
base: main
Choose a base branch
from
Open

fix(kokoro): export Prosody with fp32 compute (FluidAudio #947) #108

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
9 changes: 6 additions & 3 deletions models/tts/kokoro-v1.1-zh/coreml/scripts/convert-coreml.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -778,7 +778,10 @@ def bench(path, feed, n_runs=10):
else:
print(f'\n[3/7] Alignment — SKIP (reusing {align_path.name})')

# ═══ 4/7: Prosody (fp16+int8pal) ═══
# ═══ 4/7: Prosody (fp16 I/O, fp32 compute, int8pal) ═══
# fp32 compute: the Core ML CPU/ANE fp16 path corrupts F0/N at the start
# of the sequence for many T_a >= 400 (FluidAudio #947); GPU fp16 and
# fp32 are exact. KokoroProsody_v2 on HF.
pros_path = OUTDIR / 'KokoroProsody.mlpackage'
pros_feed = {"en": en.numpy().astype(np.float16),
"style_s": s.numpy().astype(np.float16)}
Expand All @@ -789,8 +792,8 @@ def bench(path, feed, n_runs=10):
with torch.no_grad():
traced = torch.jit.trace(prosody, (en, s), strict=False)
ml = ct.convert(traced,
inputs=[ct.TensorType(name="en", shape=(1, 640, T_a_dim), dtype=np.float32),
ct.TensorType(name="style_s", shape=(1, 128), dtype=np.float32)],
inputs=[ct.TensorType(name="en", shape=(1, 640, T_a_dim), dtype=np.float16),
ct.TensorType(name="style_s", shape=(1, 128), dtype=np.float16)],
outputs=[ct.TensorType(name="F0"), ct.TensorType(name="N")],
convert_to="mlprogram", minimum_deployment_target=ct.target.iOS17,
compute_precision=ct.precision.FLOAT32, compute_units=ct.ComputeUnit.ALL)
Expand Down
9 changes: 6 additions & 3 deletions models/tts/kokoro/laishere-coreml/convert-coreml.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -715,12 +715,15 @@ def bench(path, feed, n_runs=10):
else:
print(f'\n[3/7] Alignment — SKIP (reusing {align_path.name})')

# ═══ 4/7: Prosody (fp16+int8pal) ═══
# ═══ 4/7: Prosody (fp16 I/O, fp32 compute, int8pal) ═══
# fp32 compute: the Core ML CPU/ANE fp16 path corrupts F0/N at the start
# of the sequence for many T_a >= 400 (FluidAudio #947); GPU fp16 and
# fp32 are exact. KokoroProsody_v2 on HF.
pros_path = OUTDIR / 'KokoroProsody.mlpackage'
pros_feed = {"en": en.numpy().astype(np.float16),
"style_s": s.numpy().astype(np.float16)}
if 'prosody' in selected:
print('\n[4/7] Prosody (fp16+int8pal)...')
print('\n[4/7] Prosody (fp32 compute, int8pal)...')
prosody = CoreMLProsodyF0N(model.predictor)
prosody.eval()
with torch.no_grad():
Expand All @@ -730,7 +733,7 @@ def bench(path, feed, n_runs=10):
ct.TensorType(name="style_s", shape=(1, 128), dtype=np.float16)],
outputs=[ct.TensorType(name="F0"), ct.TensorType(name="N")],
convert_to="mlprogram", minimum_deployment_target=ct.target.iOS17,
compute_precision=ct.precision.FLOAT16, compute_units=ct.ComputeUnit.ALL)
compute_precision=ct.precision.FLOAT32, compute_units=ct.ComputeUnit.ALL)
ml = cto.palettize_weights(ml, pal_config)
ml.save(str(pros_path))
bench(pros_path, pros_feed)
Expand Down