Upload folder using huggingface_hub

.gitattributes

@@ -43,3 +43,4 @@ app/frontend/public/ir/Narrow[[:space:]]Bumpy[[:space:]]Space.wav filter=lfs dif
 app/frontend/public/ir/Nice[[:space:]]Drum[[:space:]]Room.wav filter=lfs diff=lfs merge=lfs -text
 app/frontend/public/ir/Scala[[:space:]]Milan[[:space:]]Opera[[:space:]]Hall.wav filter=lfs diff=lfs merge=lfs -text
 app/frontend/public/interface.png filter=lfs diff=lfs merge=lfs -text
+utils/vendor/wheels/antlr4_python3_runtime-4.9.3-py3-none-any.whl filter=lfs diff=lfs merge=lfs -text

app/core/config.py

@@ -43,11 +43,14 @@ class ProjectConfig:
 
         # Writable paths live under user_data_dir (diverges from project_root in frozen mode);
         # read-only code/assets stay under project_root.
+        fine_tuned_override = os.environ.get("FRAGMENTA_FINE_TUNED_DIR")
+        fine_tuned_dir = Path(fine_tuned_override) if fine_tuned_override else self.user_data_dir / "models" / "fine_tuned"
+
         self.paths: Dict[str, Path] = {
             "models": self.user_data_dir / "models",
             "models_config": self.user_data_dir / "models" / "config",
             "models_pretrained": self.user_data_dir / "models" / "pretrained",
-            "models_fine_tuned": self.user_data_dir / "models" / "fine_tuned",
+            "models_fine_tuned": fine_tuned_dir,
             "data": self.user_data_dir / "data",
             "logs": self.user_data_dir / "logs",
             "output": self.user_data_dir / "output",

requirements.txt

@@ -1,34 +1,56 @@
 --extra-index-url https://download.pytorch.org/whl/cu128
 
-Flask>=2.3.0
-Flask-CORS>=4.0.0
-requests>=2.28.0
+# --- Web stack ---
+# Flask 3.x is fine; cap below the next major to avoid surprise breakage.
+Flask>=2.3.0,<4.0
+# Flask-CORS jumped to 6.x with API changes; cap below the next major.
+Flask-CORS>=4.0.0,<7.0
+requests>=2.28.0,<3.0
 
+# --- PyTorch (CUDA 12.8 wheels) ---
 torch>=2.5,<=2.8
 torchvision<0.24
 torchaudio>=2.5,<=2.8
-transformers>=4.30.0
-diffusers>=0.20.0
-accelerate>=0.20.0
-peft>=0.4.0
-datasets>=2.14.0
-huggingface-hub>=0.16.0
 
-librosa>=0.10.0
-soundfile>=0.12.0
-scipy>=1.10.0
+# --- HuggingFace stack ---
+# transformers 5.x and huggingface-hub 1.x are major rewrites that break
+# stable-audio-tools' import paths (e.g. T5EncoderModel, AutoTokenizer).
+# Stay on the 4.x / 0.x lines that SAO was written against.
+transformers>=4.41.0,<5.0
+huggingface-hub>=0.23.0,<1.0
+# safetensors 0.8.0rc0 is a pre-release; pip occasionally selected it under
+# the old loose pin. Cap below 1.0 and avoid rc tags.
+safetensors>=0.4.3,<1.0
+diffusers>=0.27.0,<1.0
+accelerate>=0.30.0,<2.0
+peft>=0.10.0,<1.0
+# datasets 4.x dropped some legacy loaders; stay on 3.x/2.x.
+datasets>=2.18.0,<4.0
+
+# --- Audio / DSP ---
+# numba and scipy interact tightly. numba 0.61+ requires Python>=3.10 and
+# narrows the scipy window; pinning both ranges stops pip from backtracking
+# through 6+ scipy versions on every install.
+librosa>=0.10.0,<0.12
+soundfile>=0.12.0,<0.14
+scipy>=1.10.0,<1.14
+numba>=0.59,<0.61
 numpy==1.23.5
 pandas==2.0.2
 
+# --- Build / runtime utilities ---
+# setuptools<70 is required because stable-audio-tools' build chain pulls in
+# a few packages that still use deprecated setuptools APIs.
 setuptools<70
-tqdm>=4.65.0
-psutil>=5.9.0
-omegaconf>=2.3.0
-click>=8.1.0
-Pillow>=9.0.0
-python-dotenv>=1.0.0
+tqdm>=4.65.0,<5.0
+psutil>=5.9.0,<8.0
+omegaconf>=2.3.0,<3.0
+click>=8.1.0,<9.0
+Pillow>=9.0.0,<13.0
+python-dotenv>=1.0.0,<2.0
 
-pywebview>=4.4.1
+# --- Desktop / native ---
+pywebview>=4.4.1,<7.0
 Pycairo ; sys_platform == 'linux'
 PyGObject<3.49 ; sys_platform == 'linux'
-flash-attn>=2.8.3 ; sys_platform == 'linux'
+flash-attn>=2.8.3 ; sys_platform == 'linux'

stable-audio-tools/setup.py

@@ -6,35 +6,43 @@ setup(
     url='https://github.com/Stability-AI/stable-audio-tools.git',
     author='Stability AI',
     description='Training and inference tools for generative audio models from Stability AI',
-    packages=find_packages(),  
+    packages=find_packages(),
     install_requires=[
+        # Inference path (always loaded by Fragmenta)
         'alias-free-torch==0.0.6',
-        'auraloss==0.4.0',
-        'descript-audio-codec==1.0.0',
         'einops',
         'einops-exts',
-        'ema-pytorch==0.2.3',
-        'encodec==0.1.1',
-        'gradio>=5.20.0',
         'huggingface_hub',
-        'importlib-resources==5.12.0',
         'k-diffusion==0.1.1',
-        'laion-clap==1.1.4',
-        'local-attention==1.8.6',
-        'pandas==2.0.2',
-        'prefigure==0.0.9',
-        'pytorch_lightning==2.1.0',
         'PyWavelets==1.4.1',
         'safetensors',
         'sentencepiece==0.1.99',
         'torch>=2.5,<=2.8',
         'torchaudio>=2.5,<=2.8',
-        'torchmetrics==0.11.4',
         'tqdm',
         'transformers',
-        'v-diffusion-pytorch==0.0.2',
-        'vector-quantize-pytorch==1.14.41',
+
+        # Training subprocess (train.py); top-level imports must resolve
+        'auraloss==0.4.0',
+        'descript-audio-codec==1.0.0',
+        'ema-pytorch==0.2.3',
+        'pandas==2.0.2',
+        'prefigure==0.0.9',
+        'pytorch_lightning==2.1.0',
         'wandb==0.15.4',
-        'webdataset==0.2.100'
+        'webdataset==0.2.100',
+
+        # Lazy-loaded by certain model configs (kept for compatibility with
+        # configs other than SAO 1.0 / Small):
+        'encodec==0.1.1',
+        'laion-clap==1.1.4',
+        'local-attention==1.8.6',
+        'vector-quantize-pytorch==1.14.41',
+
+        # Dropped (Fragmenta does not exercise these paths):
+        # 'gradio>=5.20.0'             - run_gradio.py only; Fragmenta uses React
+        # 'v-diffusion-pytorch==0.0.2' - zero imports in this codebase
+        # 'torchmetrics==0.11.4'       - zero imports in this codebase
+        # 'importlib-resources==5.12.0' - stdlib in Python 3.9+
     ],
 )

utils/vendor/wheels/antlr4_python3_runtime-4.9.3-py3-none-any.whl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0944bf551a20f0dd5fa4fcd5d893b8622a672fd5108f72b26e2b7f689719c30f
+size 144590