Skip to content

Demo: Fix hardcoded CUDA references to support Apple Silicon MPS and fix input_sample_rate #30

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
kainoj wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Demo: Fix hardcoded CUDA references to support Apple Silicon MPS and fix input_sample_rate #30

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
60c8192
Fix hardcoded CUDA references to support Apple Silicon MPS
kainoj Jun 3, 2026
317901b
Fix input_sample_rate to match WebRTC native 48kHz
kainoj Jun 3, 2026
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
6 changes: 3 additions & 3 deletions src/liquid_audio/demo/chat.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,7 +8,7 @@

from liquid_audio import ChatState, LFMModality

from .model import lfm2_audio, mimi, proc
from .model import device, lfm2_audio, mimi, proc


def chat_producer(
Expand Down Expand Up @@ -91,7 +91,7 @@ def chat_response(audio: tuple[int, np.ndarray], _id: str, chat: ChatState, temp
chat.append(
text=torch.stack(out_text, 1),
audio_out=torch.stack(out_audio, 1),
modality_flag=torch.tensor(out_modality, device="cuda"),
modality_flag=torch.tensor(out_modality, device=device),
)

chat.end_turn()
Expand Down Expand Up @@ -122,7 +122,7 @@ def clear():
webrtc.stream(
ReplyOnPause(
chat_response, # type: ignore[arg-type]
input_sample_rate=24_000,
input_sample_rate=48_000,
output_sample_rate=24_000,
can_interrupt=False,
),
Expand Down
15 changes: 11 additions & 4 deletions src/liquid_audio/demo/model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,19 +8,26 @@

logger = logging.getLogger(__name__)

__all__ = ["lfm2_audio", "mimi", "proc"]
__all__ = ["lfm2_audio", "mimi", "proc", "device"]

HF_DIR = "LiquidAI/LFM2.5-Audio-1.5B"

if torch.cuda.is_available():
device = "cuda"
elif torch.backends.mps.is_available():
device = "mps"
else:
device = "cpu"

logging.info("Loading processor")
proc = LFM2AudioProcessor.from_pretrained(HF_DIR).eval()
proc = LFM2AudioProcessor.from_pretrained(HF_DIR, device=device).eval()
logging.info("Loading model")
lfm2_audio = LFM2AudioModel.from_pretrained(HF_DIR).eval()
lfm2_audio = LFM2AudioModel.from_pretrained(HF_DIR, device=device).eval()
logging.info("Loading tokenizer")
mimi = proc.mimi.eval()

logging.info("Warmup tokenizer")
with mimi.streaming(1), torch.no_grad():
for _ in range(5):
x = torch.randint(2048, (1, 8, 1), device="cuda")
x = torch.randint(2048, (1, 8, 1), device=device)
mimi.decode(x)
6 changes: 4 additions & 2 deletions src/liquid_audio/processor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -111,7 +111,9 @@ def mimi(self) -> MimiModel:
from safetensors.torch import load_file

mimi_model = moshi.models.loaders.get_mimi(None, device=self.device)
mimi_weights = load_file(self.mimi_weights_path, device=str(self.device))
# safetensors only supports cpu/cuda as load targets
load_device = str(self.device) if self.device.type in ("cpu", "cuda") else "cpu"
mimi_weights = load_file(self.mimi_weights_path, device=load_device)
mimi_model.load_state_dict(mimi_weights, strict=True)

self._mimi = mimi_model
Expand Down Expand Up @@ -148,7 +150,7 @@ def rename_layer(
assert isinstance(detok_config.layer_types, list)
detok_config.layer_types = [rename_layer(layer) for layer in detok_config.layer_types] # type: ignore[arg-type]

detok = LFM2AudioDetokenizer(detok_config).eval().cuda()
detok = LFM2AudioDetokenizer(detok_config).eval().to(self.device)

detok_weights_path = Path(self.detokenizer_path) / "model.safetensors"
from safetensors.torch import load_file
Expand Down