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

AGENTS.md

@@ -0,0 +1,29 @@
+## Workflow
+
+- install: `docker build -t lidar-lidar .` (deps baked into image)
+- build: `docker build -t lidar-lidar .`
+- test all: `./run.sh --test`
+- test file: `docker run --rm lidar-lidar python3 -m pytest -v --pyargs lidar_pipeline.tests.<module>`
+- test case: `docker run --rm lidar-lidar python3 -m pytest -v --pyargs lidar_pipeline.tests.<module>::<TestClass>::<test_method>`
+- lint: not configured
+- format: not configured
+- after every edit: `./run.sh --test`
+- debug: `./run.sh --debug` (file:line logging); container shell: `docker run --rm -it -v $(pwd)/input:/data/input -v $(pwd)/output:/data/output --entrypoint bash lidar-lidar`
+
+## Conventions
+
+- **Bilingual naming**: all code identifiers are English; every user-facing string, log message, argparse help, and comment is French.
+- **Adding a visualization requires 3 edits**: (1) `generate_X()` in `visualizations.py`, (2) entry in `VIZ_STEPS` in `pipeline.py`, (3) entry in `COLORMAPS` in `rendering.py`. Missing any one breaks the pipeline.
+- **`generate_*` signature is strict**: `(dem_file, basename, vis_dir, resolution, shared=None)` returning `Path` on success, `None` on failure. IGN overlays (`ortho`, `topo`) omit `shared`.
+- **Return `None` on failure, never raise**: `dtm.py`, `visualizations.py`, and `ign.py` all return `None` to let the pipeline continue. Raising aborts the entire file.
+- **Logger is always `logging.getLogger("lidar")`**, never `__name__`. All modules route through this single logger so worker processes can configure it.
+- **Filename special-cases** in `_expected_output_path()`: `pos_open` → `positive_openness`, `neg_open` → `negative_openness`, `hillshade` → `hillshade_multi`.
+- **Default output is AVIF**, not WebP. Use `--format webp` for WebP. Quality default is 98.
+- **Tests use lazy imports inside each test function**, never at module top, to avoid importing CuPy/GDAL at import time.
+- **`_`-prefixed names are critical private**: `_create_ground_pipeline`, `_fallback_to_smrf`, `_fill_nans`, `_init_gpu`, `_process_file_standalone` — do not call from outside their module.
+
+## Commit & Pull Request Guidelines
+
+Commits use imperative tense, short single-line subjects (~60–80 chars), no prefixes or scopes. Compound commits are common — multiple related changes joined by commas or "and". Examples: ``Fix multi-GPU with lazy CuPy init + rendering improvements``, ``Add multi-resolution support and remove PDF generation``, ``Fix corrupted COPC detection, add CSF→SMRF fallback, improve MSRM colormap, add SVF and anisotropic openness``.
+
+No PR template, no CI pipeline, no issue tracker. This is a standalone Docker project with no formal PR process.

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
@@ -58,14 +58,22 @@ 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)
+        # Try GPUs in order of capability.
-        best = gpus[0]
+        # CuPy 13.4 + CUDA 11.8 JIT works for sm_89 (RTX 40xx).
-        _best_gpu_id = best[0]
+        # sm_120 (RTX 50xx) is NOT supported by any nvcc yet (CUDA ≤ 12.9).
-        _gpu_name = best[1]
+        for gpu in gpus:
-        _gpu_mem_gb = best[3] // 1024
+            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
 
@@ -102,20 +110,20 @@ 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()
-
         _xp = _real_cupy
         _cp = _real_cupy
         _cp_ndimage = _real_cupy_ndimage