Merge branch 'main' of https://git.azuze.fr/kawa/KawAI

backends/memory.py

@@ -30,9 +30,20 @@ def apply_memory_strategy(pipe) -> None:
         except Exception:
             pass
 
+<<<<<<< HEAD
     if backend in ("cuda", "rocm"):
         # ROCm builds expose the cuda API, so accelerate offload hooks work the same way.
         # Offload only if VRAM tight.
+=======
+<<<<<<< HEAD
+    if backend in ("cuda", "rocm"):
+        # ROCm builds expose the cuda API, so accelerate offload hooks work the same way.
+        # Offload only if VRAM tight.
+=======
+    if backend == "cuda":
+        # Offload only if VRAM tight. cpu_offload is CUDA-only via accelerate hooks.
+>>>>>>> refs/remotes/azuze/main
+>>>>>>> 965a3d97c6dae38fa25174559b1ea0f3050788f9
         if vram < 10:
             try:
                 pipe.enable_sequential_cpu_offload()
@@ -57,6 +68,10 @@ def apply_memory_strategy(pipe) -> None:
             pipe.to("cpu")
         return
 
+<<<<<<< HEAD
+=======
+<<<<<<< HEAD
+>>>>>>> 965a3d97c6dae38fa25174559b1ea0f3050788f9
     if backend == "mps":
         # Apple Silicon shares unified memory with CPU. accelerate's sequential offload
         # has spotty MPS support; rely on slicing/tiling already enabled above.
@@ -79,5 +94,10 @@ def apply_memory_strategy(pipe) -> None:
             pipe.to("cpu")
         return
 
+<<<<<<< HEAD
+=======
+=======
+>>>>>>> refs/remotes/azuze/main
+>>>>>>> 965a3d97c6dae38fa25174559b1ea0f3050788f9
     # CPU
     pipe.to("cpu")