fix (hf): use gpu from hf space

README.md

@@ -9,7 +9,7 @@ python_version: "3.13"
 app_file: app.py
 pinned: false
 license: mit
-hardware: t4-small
+hardware: zero-gpu
 short_description: AI wardrobe. catalog, combine and ask about your clothes
 ---
 

app.py

@@ -9,8 +9,24 @@ Built for the Build Small Hackathon (HuggingFace x Gradio, June 2026).
 import io
 import logging
 import os
+import sys
 from pathlib import Path
 
+# Python 3.13 raises spurious ValueError during event-loop GC when httpx
+# (used internally by Gradio) creates and discards temporary loops.
+# Suppress the noise — it has no functional impact.
+if sys.version_info >= (3, 13):
+    _original_unraisablehook = sys.unraisablehook
+
+    def _quiet_unraisablehook(unraisable):
+        if isinstance(unraisable.exc_value, ValueError) and "file descriptor" in str(
+            unraisable.exc_value
+        ):
+            return
+        _original_unraisablehook(unraisable)
+
+    sys.unraisablehook = _quiet_unraisablehook
+
 from dotenv import load_dotenv
 
 load_dotenv(Path(__file__).resolve().parent / ".env")
@@ -38,6 +54,19 @@ from src.detector import detect_garments as detect_boxes, crop_garments, list_av
 from src import settings
 from gradio_image_annotation import image_annotator
 
+try:
+    import spaces
+except ImportError:
+    class _NoOpGPU:
+        """Fallback when not running on HF Spaces (local dev)."""
+        def __call__(self, fn=None, *, duration=None):
+            if fn is None:
+                return lambda f: f
+            return fn
+
+    class spaces:  # type: ignore[no-redef]
+        GPU = _NoOpGPU()
+
 os.environ.setdefault("CUDA_VISIBLE_DEVICES", "0")
 logging.basicConfig(level=logging.INFO, format="%(name)s | %(message)s")
 
@@ -79,6 +108,7 @@ def _format_results(added: list[dict]) -> tuple[str, list]:
     return status, table_data
 
 
+@spaces.GPU(duration=120)
 def auto_detect(annotation, mode, progress=gr.Progress(track_tqdm=False)):
     """Run automatic detection on the uploaded image.
 
@@ -116,6 +146,7 @@ def auto_detect(annotation, mode, progress=gr.Progress(track_tqdm=False)):
     return _format_results(added)
 
 
+@spaces.GPU(duration=120)
 def process_manual_selection(annotation, progress=gr.Progress(track_tqdm=False)):
     """Process manually drawn bounding boxes from the annotator.
 
@@ -185,6 +216,7 @@ SAMPLE_DATASETS = [
 TARGET_GARMENTS = 50
 
 
+@spaces.GPU(duration=300)
 def obtener_dataset(dataset_key: str):
     """Download a HF dataset and process garments in-process with real-time UI updates.
 
@@ -421,6 +453,7 @@ def clear_all():
 # Chat handlers
 # ---------------------------------------------------------------------------
 
+@spaces.GPU(duration=90)
 def chat_respond(message, history):
     """Handle chat messages with streaming responses."""
     if not message:
@@ -463,6 +496,7 @@ def _get_combo_display(combo: dict | None) -> tuple[str | None, str | None, str,
     return top_img, bottom_img, top_text, bottom_text
 
 
+@spaces.GPU(duration=60)
 def init_combinations(state, context):
     """Initialize or refresh the combination queue, optionally ranked by context."""
     combos = generate_combinations()
@@ -818,6 +852,7 @@ def _build_custom_server():
         return {"garments": catalog, "count": len(catalog)}
 
     @server.api(name="add_photo")
+    @spaces.GPU(duration=120)
     def api_add_photo(image_path: str) -> dict:
         results = extract_garments(image_path)
         if not results:
@@ -828,6 +863,7 @@ def _build_custom_server():
         return {"garments": added, "count": len(added)}
 
     @server.api(name="get_combinations")
+    @spaces.GPU(duration=60)
     def api_get_combinations(context: str = "") -> dict:
         combos = generate_combinations()
         if not combos:
@@ -852,12 +888,14 @@ def _build_custom_server():
         return {"status": "ok", "liked": liked}
 
     @server.api(name="ask_question")
+    @spaces.GPU(duration=90)
     def api_ask_question(question: str) -> str:
         if not question or not question.strip():
             return "Please ask a question about your wardrobe."
         return ask(question.strip())
 
     @server.api(name="load_dataset")
+    @spaces.GPU(duration=300)
     def api_load_dataset(dataset_key: str) -> dict:
         from datasets import load_dataset as hf_load
 

requirements.txt

@@ -1,4 +1,5 @@
 --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu124
+spaces
 gradio==6.17.3
 llama-cpp-python>=0.3.28
 huggingface-hub>=1.18.0

src/model_loader.py

@@ -1,11 +1,15 @@
 """Singleton model loader with VRAM management.
 
 Handles loading/unloading of GGUF models. Only one model is kept in
-memory at a time to fit within 8 GB VRAM.
+memory at a time to fit within available VRAM.
+
+Supports ZeroGPU (dynamic GPU allocation on HF Spaces) by detecting
+CUDA availability at load time rather than import time.
 """
 
 import gc
 import logging
+import os
 from pathlib import Path
 from dataclasses import dataclass
 
@@ -90,6 +94,19 @@ class _ModelManager:
 
         return model_path, mmproj_path
 
+    @staticmethod
+    def _detect_gpu_layers() -> int:
+        """Return -1 (all layers on GPU) if CUDA is available, else 0 (CPU)."""
+        if os.environ.get("CUDA_VISIBLE_DEVICES") == "":
+            return 0
+        try:
+            import ctypes
+            ctypes.CDLL("libcudart.so.12")
+            return -1
+        except OSError:
+            logger.warning("CUDA runtime not found — running on CPU")
+            return 0
+
     def _is_same_model(self, config: ModelConfig) -> bool:
         if self._current_config is None:
             return False
@@ -119,12 +136,16 @@ class _ModelManager:
             from llama_cpp.llama_chat_format import Qwen25VLChatHandler
             chat_handler = Qwen25VLChatHandler(clip_model_path=str(mmproj_path))
 
-        logger.info("Loading model: %s (handler: %s)", config.model_file, config.handler_type)
+        n_gpu_layers = self._detect_gpu_layers()
+        logger.info(
+            "Loading model: %s (handler: %s, gpu_layers: %s)",
+            config.model_file, config.handler_type, n_gpu_layers,
+        )
 
         self._llm = Llama(
             model_path=str(model_path),
             chat_handler=chat_handler,
-            n_gpu_layers=-1,
+            n_gpu_layers=n_gpu_layers,
             n_ctx=config.n_ctx,
             verbose=False,
         )