Skip to content

Add PEFT and DeepSpeed dependencies; align Qwen2Model usage and TTS prompt with MiMo-Audio #2

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
candlewill wants to merge 4 commits into
base: main
Choose a base branch
from
Open

Add PEFT and DeepSpeed dependencies; align Qwen2Model usage and TTS prompt with MiMo-Audio #2

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
754028e
add peft to requirements
candlewill Oct 17, 2025
cf21ee5
update requirements.txt
candlewill Oct 20, 2025
813cfd8
keep is_causal the same to the MiMo-Audio, see https://github.com/Xia…
candlewill Oct 20, 2025
aa1c009
fix speech_loss_weights missing error
candlewill Oct 20, 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
2 changes: 2 additions & 0 deletions instruct_template.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -70,6 +70,8 @@
]
```

* Note: Change the user prompt from "Please convert this text to speech" to "请将这段文字转换为语音" when synthesizing Chinese text, to better align with the base model's training data.

### Instruct TTS

```
Expand Down
11 changes: 7 additions & 4 deletions mimo_audio_train/models/mimo_audio.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,6 +15,7 @@
)
from transformers.tokenization_utils_fast import PreTrainedTokenizerFast
from peft import get_peft_model
from mimo_audio_train.arguments import CustomArguments

from .src_mimo_audio.process_speechdata import InputSegment, StreamingInputSegment
from .src_mimo_audio.mimo_audio_tokenizer import MiMoAudioTokenizer
Expand Down Expand Up @@ -75,6 +76,7 @@ def __init__(
self.eot_idx = self.tokenizer.convert_tokens_to_ids("<|eot|>")
self.im_start_idx = self.tokenizer.convert_tokens_to_ids("<|im_start|>")
self.im_end_idx = self.tokenizer.convert_tokens_to_ids("<|im_end|>")
self.speech_loss_weights = CustomArguments.speech_loss_weights

model_args = MiMoAudioArguments(
model_name_or_path=self.path,
Expand All @@ -84,6 +86,7 @@ def __init__(
sostm_idx=self.sostm_idx,
eostm_idx=self.eostm_idx,
eot_idx=self.eot_idx,
speech_loss_weights=self.speech_loss_weights,
)

start_loading_time = time.monotonic()
Expand Down Expand Up @@ -510,9 +513,9 @@ def get_tts_sft_prompt(
else:
language = detect_language(input)
if language == "zh":
template = "请将这段文字转换为语音："
template = "请将这段文字转换为语音"
else:
template = "Please convert this text to speech."
template = "Please convert this text to speech"

text = self.preprocess_input(input)
if instruct is None:
Expand Down Expand Up @@ -1397,7 +1400,7 @@ def forward(
add_history=False,
task_name=None,
):

task_sampler = self.get_task_sampler(task_name)

generation_kwargs = self.generate_kwargs.copy()
Expand Down Expand Up @@ -1485,7 +1488,7 @@ def forward(
return detokenized_text
else:
return wav_concat

def asr_sft(self, audio, lang='zh'):
stopping_criteria = [
MiMoStopper(
Expand Down
1 change: 0 additions & 1 deletion mimo_audio_train/models/src_mimo_audio/modeling_mimo_audio.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -486,7 +486,6 @@ def forward(
use_cache=True,
return_dict=True,
cache_position=cache_position,
is_causal=True,
)
hidden_states = outputs.last_hidden_state # [B, new_T_group, hidden_size]

Expand Down
4 changes: 3 additions & 1 deletion requirements.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,4 +27,6 @@ hyperpyyaml==1.2.2
loguru==0.7.3
sox==1.5.0
s3prl==0.4.18
timm
timm
peft
deepspeed