Skip to content

fix facebook/hf-seamless-m4t-medium crash #1433

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

Merged
regisss merged 3 commits into from
Oct 30, 2024
Merged

fix facebook/hf-seamless-m4t-medium crash #1433

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
e1b0fd8
fix facebook/hf-seamless-m4t-medium crash
sywangyi Oct 17, 2024
3c76192
Merge branch 'main' into seamless_fix
sywangyi Oct 25, 2024
4d84946
update README
sywangyi Oct 25, 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
1 change: 1 addition & 0 deletions examples/text-to-speech/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,4 +36,5 @@ python3 run_pipeline.py \
```
Models that have been validated:
- [microsoft/speecht5_tts](https://huggingface.co/microsoft/speecht5_tts)
- [facebook/hf-seamless-m4t-medium](https://huggingface.co/facebook/hf-seamless-m4t-medium)
- [facebook/mms-tts-eng](https://huggingface.co/facebook/mms-tts-eng)
17 changes: 13 additions & 4 deletions optimum/habana/transformers/generation/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -237,14 +237,20 @@ def _prepare_decoder_input_ids_for_generation(
if token_idx is None:
decoder_input_ids = torch.cat([decoder_start_token_id, decoder_input_ids], dim=-1)
else:
max_length = max_new_tokens + 2 if max_new_tokens is not None else self.generation_config.max_length
decoder_input_ids_len = decoder_input_ids.shape[-1]
max_length = (
max_new_tokens + decoder_input_ids_len + 1
if max_new_tokens is not None
else self.generation_config.max_length
)
if max_length != decoder_start_token_id.shape[-1]:
decoder_start_token_id = torch.nn.functional.pad(
decoder_start_token_id,
(0, max_length - decoder_start_token_id.shape[-1]),
value=pad_token_id,
)
decoder_input_ids = decoder_start_token_id.index_copy(1, token_idx, decoder_input_ids)
decoder_start_token_id[:, 1 : 1 + decoder_input_ids_len, ...] = decoder_input_ids
decoder_input_ids = decoder_start_token_id
token_idx.add_(1)
if "decoder_attention_mask" in model_kwargs:
decoder_attention_mask = model_kwargs["decoder_attention_mask"]
Expand Down Expand Up @@ -1103,11 +1109,14 @@ def generate(
)
else:
assert generation_config.bucket_size <= 0, "Untested path for bucket>0"
token_idx = 1
if model_kwargs.get("decoder_input_ids", None) is None:
token_idx = 1
else:
token_idx = model_kwargs["decoder_input_ids"].shape[-1]
model_kwargs["token_idx"] = torch.tensor(token_idx, device=inputs_tensor.device)
if model_kwargs.get("decoder_attention_mask", None) is None and generation_config.use_cache:
max_length = (
generation_config.max_new_tokens + 1
generation_config.max_new_tokens + token_idx
if generation_config.max_new_tokens is not None
else generation_config.max_length
)
Expand Down