fix(cpu): force DDIM in get_bald; add /health/gpu diagnostics

Hair_stable_new_fresh/infer_full.py

@@ -170,6 +170,12 @@ class StableHair:
     def get_bald(self, id_image, scale):
         H, W = id_image.size
         scale = float(scale)
+        # Enforce a CPU/GPU-safe scheduler every call in case upstream resets it
+        try:
+            from diffusers import DDIMScheduler as _DDIM
+            self.remove_hair_pipeline.scheduler = _DDIM.from_config(self.remove_hair_pipeline.scheduler.config)
+        except Exception:
+            pass
         image = self.remove_hair_pipeline(
             prompt="",
             negative_prompt="",

Hair_stable_new_fresh/server.py

@@ -56,6 +56,21 @@ def root():
     return {"status": "ok"}
 
 
+@app.get("/health/gpu")
+def health_gpu():
+    try:
+        import torch as _t
+        return {
+            "cuda_available": bool(_t.cuda.is_available()),
+            "cuda_device_count": int(_t.cuda.device_count()) if _t.cuda.is_available() else 0,
+            "cuda_device_name": _t.cuda.get_device_name(0) if _t.cuda.is_available() else None,
+            "torch_version": getattr(_t, "__version__", None),
+            "torch_cuda_version": getattr(_t.version, "cuda", None),
+        }
+    except Exception as e:
+        return {"error": str(e)}
+
+
 class HairSwapRequest(BaseModel):
     source_id: str
     reference_id: str
@@ -169,11 +184,11 @@ def get_hairswap(req: HairSwapRequest, _=Depends(verify_bearer)):
                 source_image=source_path,
                 reference_image=reference_path,
                 random_seed=-1,
-                step=30,
+                step=20,
                 guidance_scale=req.guidance_scale,
                 scale=req.scale,
                 controlnet_conditioning_scale=req.controlnet_conditioning_scale,
-                size=512,
+                size=448,
             )
             LOGGER.info("Hair transfer completed successfully")
         except Exception as e: