Create vince_server.py

services/vince_server.py

@@ -0,0 +1,43 @@
+# services/vince_server.py
+from common.config import load_config, create_object
+from pathlib import Path
+import gc, torch
+
+class VinceSingleton:
+    def __init__(self, config_path: str, overrides: list[str]):
+        self.config = load_config(config_path, overrides)
+        self.gen = create_object(self.config)
+        self.gen.configure_persistence()
+        self.gen.configure_models()
+        self.gen.configure_diffusion()
+
+    def _set_steps(self, steps: int | None):
+        if steps and hasattr(self.gen, "sampler") and hasattr(self.gen.sampler, "timesteps"):
+            ts = self.gen.sampler.timesteps
+            if hasattr(ts, "__len__") and len(ts) > 0:
+                steps = min(int(steps), len(ts))
+                if steps < len(ts):
+                    idx = torch.linspace(0, len(ts) - 1, steps).round().long().tolist()
+                    self.gen.sampler.timesteps = [ts[i] for i in idx]
+
+    def generate_multi_turn(self, image_path, turns, out_dir, *, steps=None, cfg_scale=None, aspect_ratio=None, resolution=None):
+        g = self.gen.config.generation
+        g.output.dir = str(out_dir)
+        g.positive_prompt = {"image_path": [str(image_path)], "prompts": list(turns)}
+        if cfg_scale is not None: g.cfg_scale = float(cfg_scale)
+        if aspect_ratio is not None: g.aspect_ratio = str(aspect_ratio)
+        if resolution is not None: g.resolution = int(resolution)
+        self._set_steps(steps)
+        self.gen.inference_loop()
+        try:
+            torch.cuda.synchronize()
+        except Exception:
+            pass
+        gc.collect()
+        try:
+            torch.cuda.empty_cache()
+            torch.cuda.memory.reset_peak_memory_stats()
+        except Exception:
+            pass
+        return str(out_dir)
+