fix LAMMPS wheel with CUDA wheels (#2887)

It's annoying that in the CUDA Python wheels, the library filename is
`libcudart.so.11.0`, instead of `libcudart.so` or `libcudart.so.11` or
`libcudart.11.0.xxx`. There are no symlinks in these CUDA Python wheels
on PyPI. Thus, in this case, when DeePMD-kit tries to load
`libcudart.so`, no file can be found.
Python interface doesn't have this issue, as it tries to use `glob` to
find detailed library names. For the LAMMPS executable called from
Python, `LD_LIBRARY_PATH` doesn't work with CUDA Python libraries (but
works with an installation that contains symlinks). Thus, in this PR,
`LD_PRELOAD` is set instead to preload CUDA symbols.
This fixes the docker images for the development branch.

Signed-off-by: Jinzhe Zeng <[email protected]>

deepmd/lmp.py

@@ -35,13 +35,15 @@ def get_env(paths: List[Optional[str]]) -> str:
     return ":".join(p for p in paths if p is not None)
 
 
-def get_library_path(module: str) -> List[str]:
+def get_library_path(module: str, filename: str) -> List[str]:
     """Get library path from a module.
 
     Parameters
     ----------
     module : str
         The module name.
+    filename : str
+        The library filename pattern.
 
     Returns
     -------
@@ -53,7 +55,8 @@ def get_library_path(module: str) -> List[str]:
     except ModuleNotFoundError:
         return []
     else:
-        return [str(Path(m.__file__).parent)]
+        libs = sorted(Path(m.__path__[0]).glob(filename))
+        return [str(lib) for lib in libs]
 
 
 if platform.system() == "Linux":
@@ -63,6 +66,13 @@ def get_library_path(module: str) -> List[str]:
 else:
     raise RuntimeError("Unsupported platform")
 
+if platform.system() == "Linux":
+    preload_env = "LD_PRELOAD"
+elif platform.system() == "Darwin":
+    preload_env = "DYLD_INSERT_LIBRARIES"
+else:
+    raise RuntimeError("Unsupported platform")
+
 tf_dir = tf.sysconfig.get_lib()
 op_dir = str((Path(__file__).parent / "lib").absolute())
 
@@ -71,37 +81,37 @@ def get_library_path(module: str) -> List[str]:
 if platform.system() == "Linux":
     cuda_library_paths.extend(
         [
-            *get_library_path("nvidia.cuda_runtime.lib"),
-            *get_library_path("nvidia.cublas.lib"),
-            *get_library_path("nvidia.cublas.lib"),
-            *get_library_path("nvidia.cufft.lib"),
-            *get_library_path("nvidia.curand.lib"),
-            *get_library_path("nvidia.cusolver.lib"),
-            *get_library_path("nvidia.cusparse.lib"),
-            *get_library_path("nvidia.cudnn.lib"),
+            *get_library_path("nvidia.cuda_runtime.lib", "libcudart.so*"),
+            *get_library_path("nvidia.cublas.lib", "libcublasLt.so*"),
+            *get_library_path("nvidia.cublas.lib", "libcublas.so*"),
+            *get_library_path("nvidia.cufft.lib", "libcufft.so*"),
+            *get_library_path("nvidia.curand.lib", "libcurand.so*"),
+            *get_library_path("nvidia.cusolver.lib", "libcusolver.so*"),
+            *get_library_path("nvidia.cusparse.lib", "libcusparse.so*"),
+            *get_library_path("nvidia.cudnn.lib", "libcudnn.so*"),
         ]
     )
 
+os.environ[preload_env] = get_env(
+    [
+        os.environ.get(preload_env),
+        *cuda_library_paths,
+    ]
+)
+
 # set LD_LIBRARY_PATH
 os.environ[lib_env] = get_env(
     [
         os.environ.get(lib_env),
         tf_dir,
         os.path.join(tf_dir, "python"),
         op_dir,
-        *cuda_library_paths,
     ]
 )
 
 # preload python library, only for TF<2.12
 if find_libpython is not None:
     libpython = find_libpython()
-    if platform.system() == "Linux":
-        preload_env = "LD_PRELOAD"
-    elif platform.system() == "Darwin":
-        preload_env = "DYLD_INSERT_LIBRARIES"
-    else:
-        raise RuntimeError("Unsupported platform")
     os.environ[preload_env] = get_env(
         [
             os.environ.get(preload_env),