Update app.py

app.py

@@ -18,11 +18,9 @@ from concurrent.futures import ThreadPoolExecutor
 app = Flask(__name__)
 CORS(app)
 
-# dynamic hardware detection
 has_gpu = torch.cuda.is_available()
 device_type = "cuda" if has_gpu else "cpu"
-# use lower sample rate on cpu for speed; higher on gpu for quality
-sr = 44100 if has_gpu else 22050
+sr = 44100
 target_loudness = -9.0
 
 def convert_to_wav(input_path):
@@ -35,9 +33,7 @@ def convert_to_wav(input_path):
 def load_mono(file_path):
     if not os.path.exists(file_path):
         return np.zeros(sr * 5)
-    # limit to 60s on cpu to prevent hangs
-    duration = None if has_gpu else 60
-    y, _ = librosa.load(file_path, sr=sr, mono=True, duration=duration)
+    y, _ = librosa.load(file_path, sr=sr, mono=True)
     return y
 
 def normalize_audio(y):
@@ -83,7 +79,7 @@ def separate_stems(input_file, job_id):
         input_file
     ]
     if not has_gpu:
-        cmd.extend(["-j", "2"]) # limit threads on cpu to avoid oom
+        cmd.extend(["-j", "1"])
     
     subprocess.run(cmd, check=True)
     base = os.path.splitext(os.path.basename(input_file))[0]
@@ -113,16 +109,12 @@ def fuse_api():
         m_req.save(m_path)
         temp_files.extend([t_path, m_path])
 
-        # parallel on gpu, sequential on cpu to manage resources
-        if has_gpu:
-            with ThreadPoolExecutor(max_workers=2) as executor:
-                f_t = executor.submit(separate_stems, t_path, f"t_{job_id}")
-                f_m = executor.submit(separate_stems, m_path, f"m_{job_id}")
-                t_stems, t_dir = f_t.result()
-                m_stems, m_dir = f_m.result()
-        else:
-            t_stems, t_dir = separate_stems(t_path, f"t_{job_id}")
-            m_stems, m_dir = separate_stems(m_path, f"m_{job_id}")
+        max_workers = 2 if has_gpu else 1
+        with ThreadPoolExecutor(max_workers=max_workers) as executor:
+            f_t = executor.submit(separate_stems, t_path, f"t_{job_id}")
+            f_m = executor.submit(separate_stems, m_path, f"m_{job_id}")
+            t_stems, t_dir = f_t.result()
+            m_stems, m_dir = f_m.result()
         
         cleanup_dirs.extend([t_dir, m_dir])