Skip to content
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

Fix vits low-precision dtype #35418

Open
wants to merge 3 commits into
base: main
Choose a base branch
from
Open

Fix vits low-precision dtype #35418

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
e73927c
fix vits dtype
jiqing-feng Dec 26, 2024
6e8f286
Merge branch 'main' into vits
jiqing-feng Jan 3, 2025
671fee8
add tests
jiqing-feng Jan 3, 2025
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
4 changes: 2 additions & 2 deletions src/transformers/models/vits/modeling_vits.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1407,9 +1407,9 @@ def forward(
raise NotImplementedError("Training of VITS is not supported yet.")

if attention_mask is not None:
input_padding_mask = attention_mask.unsqueeze(-1).float()
input_padding_mask = attention_mask.unsqueeze(-1).to(self.dtype)
else:
input_padding_mask = torch.ones_like(input_ids).unsqueeze(-1).float()
input_padding_mask = torch.ones_like(input_ids).unsqueeze(-1).to(self.dtype)

if self.config.num_speakers > 1 and speaker_id is not None:
if not 0 <= speaker_id < self.config.num_speakers:
Expand Down
32 changes: 32 additions & 0 deletions tests/models/vits/test_modeling_vits.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,6 +27,7 @@
is_flaky,
is_torch_available,
require_torch,
require_torch_fp16,
require_torch_multi_gpu,
slow,
torch_device,
Expand Down Expand Up @@ -434,3 +435,34 @@ def test_forward(self):
)
# fmt: on
self.assertTrue(torch.allclose(outputs.waveform[0, 10000:10030].cpu(), EXPECTED_LOGITS, atol=1e-4))

@require_torch_fp16
def test_forward_fp16(self):
# GPU gives different results than CPU
torch_device = "cpu"

model = VitsModel.from_pretrained("facebook/mms-tts-eng", torch_dtype=torch.float16)
model.to(torch_device)

tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-eng")

set_seed(555) # make deterministic

input_text = "Mister quilter is the apostle of the middle classes and we are glad to welcome his gospel!"
input_ids = tokenizer(input_text, return_tensors="pt").input_ids.to(torch_device)

with torch.no_grad():
outputs = model(input_ids)

self.assertEqual(outputs.waveform.shape, (1, 87040))
# fmt: off
EXPECTED_LOGITS = torch.tensor(
[
0.0101, 0.0318, 0.0489, 0.0627, 0.0728, 0.0865, 0.1053, 0.1279,
0.1514, 0.1703, 0.1827, 0.1829, 0.1694, 0.1509, 0.1332, 0.1188,
0.1066, 0.0978, 0.0936, 0.0867, 0.0724, 0.0493, 0.0197, -0.0141,
-0.0501, -0.0817, -0.1065, -0.1223, -0.1311, -0.1339
]
).to(torch.float16)
# fmt: on
self.assertTrue(torch.allclose(outputs.waveform[0, 10000:10030].cpu(), EXPECTED_LOGITS, atol=1e-4))