add model downloads

inspiremusic/cli/frontend.py

@@ -12,8 +12,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from functools import partial
+import onnxruntime
 import torch
+import numpy as np
+import whisper
 from typing import Callable
+import torchaudio.compliance.kaldi as kaldi
+import torchaudio
+import os
 import re
 import inflect
 from inspiremusic.cli.model import InspireMusicModel

inspiremusic/cli/inference.py

@@ -49,15 +49,14 @@ class InspireMusicUnified:
                  fp16: bool = True,
                  gpu: int = 0,
                  result_dir: str = None,
-                 hub: str = "modelscope"):
+                 hub="modelscope"):
         os.environ['CUDA_VISIBLE_DEVICES'] = str(gpu)
 
         # Set model_dir or default to downloading if it doesn't exist
         if model_dir is None:
              model_dir = f"../../pretrained_models/{model_name}"
-        print(model_dir)
+
         download_model_dir = os.path.dirname(model_dir)
-        print(download_model_dir)
         if not os.path.isfile(f"{model_dir}/llm.pt"):
             if hub == "modelscope":
                 from modelscope import snapshot_download
@@ -68,9 +67,7 @@ class InspireMusicUnified:
             elif hub == "huggingface":
                 from huggingface_hub import snapshot_download
                 model_dir_tmp = snapshot_download(repo_id=f"FunAudioLLM/{model_name}", cache_dir=download_model_dir)
-            print(download_model_dir, model_dir_tmp, model_dir)
             shutil.move(model_dir_tmp, model_dir)
-            # shutil.rmtree(model_dir_tmp)
 
         self.model_dir = model_dir
 
@@ -88,6 +85,7 @@ class InspireMusicUnified:
         use_cuda = gpu >= 0 and torch.cuda.is_available()
         self.device = torch.device('cuda' if use_cuda else 'cpu')
         self.model = InspireMusic(self.model_dir, load_jit=load_jit, load_onnx=load_onnx, fast=fast, fp16=fp16)
+        self.model.model.llm = self.model.model.llm.to(torch.float16)
 
         logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
 
@@ -202,10 +200,6 @@ class InspireMusicUnified:
 
             else:
                 logging.error(f"Generated audio length is shorter than minimum required audio length.")
-        if music_fn:
-            return music_fn
-        else:
-            return None
 
 def get_args():
     parser = argparse.ArgumentParser(description='Run inference with your model')

inspiremusic/cli/inspiremusic.py

@@ -39,14 +39,9 @@ class InspireMusic:
             elif hub == "huggingface":
                 from huggingface_hub import snapshot_download
                 model_dir_tmp = snapshot_download(repo_id=f"FunAudioLLM/{model_name}", cache_dir=download_model_dir)
-            print(model_dir_tmp, model_dir)
             shutil.move(model_dir_tmp, model_dir)
-            # shutil.rmtree(model_dir_tmp)
-
-        self.model_dir = model_dir
 
         assert os.path.exists(f'{model_dir}/inspiremusic.yaml')
-
         with open('{}/inspiremusic.yaml'.format(model_dir), 'r') as f:
             configs = load_hyperpyyaml(f)
 

inspiremusic/llm/llm.py

@@ -35,7 +35,7 @@ class SinusoidalEmbedding(nn.Module):
         emb = torch.tensor(log(10000) / (half_dim - 1), device=device)
         emb = torch.exp(torch.arange(half_dim, device=device) * -emb)
         emb = rearrange(x, "i -> i 1") * rearrange(emb, "j -> 1 j")
-        return torch.cat((emb.sin(), emb.cos()), dim=-1).to(torch.float32)
+        return torch.cat((emb.sin(), emb.cos()), dim=-1).to(torch.float16)
 
 class LLM(torch.nn.Module):
     def __init__(