Auto-detect best GPU with sm_120 skip: prefer RTX 5060 Ti, fall back to 4060 Ti when nvcc/CuPy does not support sm_120 yet

lidar_pipeline/gpu.py

@@ -50,7 +50,7 @@ def _pick_gpu() -> int | None:
             mem_mi = int(parts[3])
             major, minor = (int(x) for x in cap_str.split('.'))
             score = major * 1000 + minor * 100 + mem_mi
-            gpus.append((idx, name, cap_str, mem_mi, score))
+            gpus.append((idx, name, cap_str, mem_mi, score, major))
 
         if not gpus:
             return None
@@ -58,13 +58,21 @@ def _pick_gpu() -> int | None:
         _NUM_GPUS = len(gpus)
         gpus.sort(key=lambda g: g[4], reverse=True)
 
-        # Use the best GPU (highest compute capability)
-        best = gpus[0]
-        _best_gpu_id = best[0]
-        _gpu_name = best[1]
-        _gpu_mem_gb = best[3] // 1024
-        HAS_GPU = True
-        return _best_gpu_id
+        # Try GPUs in order of capability.
+        # CuPy 13.4 + CUDA 11.8 JIT works for sm_89 (RTX 40xx).
+        # sm_120 (RTX 50xx) is NOT supported by any nvcc yet (CUDA ≤ 12.9).
+        for gpu in gpus:
+            idx, name, cap_str, mem_mi, score, major = gpu
+            if major >= 12:
+                logger.debug(f"  GPU {idx}: {name} (sm_{cap_str}) — non supporté par nvcc/CuPy")
+                continue
+            _best_gpu_id = idx
+            _gpu_name = name
+            _gpu_mem_gb = mem_mi // 1024
+            HAS_GPU = True
+            return _best_gpu_id
+
+        _gpu_reason = "aucun GPU compatible (sm_120+ non supporté par nvcc/CuPy)"
 
     except (FileNotFoundError, subprocess.TimeoutExpired, Exception):
         return None
@@ -102,19 +110,19 @@ def _init_gpu():
         return
 
     try:
+        # MUST set CUDA_VISIBLE_DEVICES before importing CuPy.
+        # If we don't, CuPy creates its CUDA context on device 0 (sm_89)
+        # and pre-compiled kernels won't work on device 1 (sm_120).
+        os.environ['CUDA_VISIBLE_DEVICES'] = str(_best_gpu_id)
+
         import cupy as _real_cupy
         import cupyx.scipy.ndimage as _real_cupy_ndimage
 
-        # Select the target GPU using Device API
-        with _real_cupy.cuda.Device(_best_gpu_id):
-            # Warm-up: verify kernel execution works on this GPU
-            _test = _real_cupy.array([1.0, 2.0, 3.0], dtype=_real_cupy.float32)
-            _result = _real_cupy.sum(_test * _test)
-            _ = _result.get()
-            del _test, _result
-
-        # Make this device the default for all subsequent operations
-        _real_cupy.cuda.Device(_best_gpu_id).use()
+        # Warm-up: verify kernel execution works on this GPU
+        _test = _real_cupy.array([1.0, 2.0, 3.0], dtype=_real_cupy.float32)
+        _result = _real_cupy.sum(_test * _test)
+        _ = _result.get()
+        del _test, _result
 
         _xp = _real_cupy
         _cp = _real_cupy