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

[Whisper] patch float type on mps #35295

Merged
merged 2 commits into from
Dec 16, 2024
Merged

[Whisper] patch float type on mps #35295

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift click to select a range
e7340a3
fix float type on mps
eustlb Dec 16, 2024
b618dc9
make
eustlb Dec 16, 2024
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
17 changes: 13 additions & 4 deletions src/transformers/models/whisper/generation_whisper.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -632,7 632,9 @@ def generate(
cur_bsz=cur_bsz,
batch_idx_map=batch_idx_map,
)
time_offset = seek.to(torch.float64) * time_precision / input_stride
time_offset = (
seek.to(torch.float32 if device.type == "mps" else torch.float64) * time_precision / input_stride
)
seek_num_frames = (max_frames - seek).clamp(max=num_segment_frames)

# 6.2 cut out next 30s segment from input features
Expand Down Expand Up @@ -1805,6 1807,7 @@ def _retrieve_segment(
timestamp_segment_indices = torch.where(timestamp_tokens[:-1] & timestamp_tokens[1:])[0]
timestamp_segment_indices.add_(1)
token_timestamps = seek_outputs[idx]["token_timestamps"] if return_token_timestamps else []
device = seek_sequence.device

# If whisper predicted a "end of segment" via a timestep token, let's go ever each
# "end of segment" prediction and slice the decoding into segments accordingly
Expand All @@ -1828,8 1831,12 @@ def _retrieve_segment(
end_timestamp_pos = sliced_tokens[idx_sliced_tokens] - timestamp_begin
segments.append(
{
"start": time_offset[prev_idx] start_timestamp_pos.to(torch.float64) * time_precision,
"end": time_offset[prev_idx] end_timestamp_pos.to(torch.float64) * time_precision,
"start": time_offset[prev_idx]
start_timestamp_pos.to(torch.float32 if device.type == "mps" else torch.float64)
* time_precision,
"end": time_offset[prev_idx]
end_timestamp_pos.to(torch.float32 if device.type == "mps" else torch.float64)
* time_precision,
"tokens": sliced_tokens,
"result": seek_outputs[idx],
}
Expand All @@ -1856,7 1863,9 @@ def _retrieve_segment(
last_timestamp_pos = int(seek_num_frames[prev_idx] * time_precision_features / time_precision)
if timestamps.numel() > 0 and timestamps[-1] != timestamp_begin:
# no consecutive timestamps but it has a timestamp; use the last one.
last_timestamp_pos = (timestamps[-1] - timestamp_begin).to(torch.float64)
last_timestamp_pos = (timestamps[-1] - timestamp_begin).to(
torch.float32 if device.type == "mps" else torch.float64
)
segments = [
{
"start": time_offset[prev_idx],
Expand Down