Don't set $TPU_LIBRARY_PATH during import

pytorch · Oct 11, 2023 · 84a783d · 84a783d
1 parent 418c751
commit 84a783d
Show file tree

Hide file tree

Showing 4 changed files with 14 additions and 8 deletions.
diff --git a/torch_xla/__init__.py b/torch_xla/__init__.py
@@ -72,27 +72,30 @@ def _aws_ec2_inf_trn_init():
 
 
 def _setup_tpu_vm_library_path() -> bool:
-  """Returns true if $TPU_LIBRARY is set or can be inferred.
+  """Returns true if $TPU_LIBRARY_PATH is set or can be inferred.
 
   We load libtpu.so in the following order of precedence:
 
   1. User-set $TPU_LIBRARY_PATH
   2. libtpu.so included in torch_xla/lib
   3. libtpu-nightly pip package
+
+  Sets $PTXLA_TPU_LIBRARY_PATH if path is inferred by us to prevent conflicts
+  with other frameworks. This env var will be removed in a future version.
   """
-  if 'TPU_LIBRARY_PATH' in os.environ:
+  if 'TPU_LIBRARY_PATH' in os.environ or 'PTXLA_TPU_LIBRARY_PATH' in os.environ:
     return True
 
   module_path = os.path.dirname(__file__)
   bundled_libtpu_path = os.path.join(module_path, 'lib/libtpu.so')
   if os.path.isfile(bundled_libtpu_path) and not os.getenv('TPU_LIBRARY_PATH'):
     logger.info('Using bundled libtpu.so (%s)', bundled_libtpu_path)
-    os.environ['TPU_LIBRARY_PATH'] = bundled_libtpu_path
+    os.environ['PTXLA_TPU_LIBRARY_PATH'] = bundled_libtpu_path
     return True
 
   try:
     import libtpu
-    libtpu.configure_library_path()
+    os.environ['PTXLA_TPU_LIBRARY_PATH'] = libtpu.get_library_path()
     return True
   except ImportError:
     return False

diff --git a/torch_xla/csrc/runtime/env_vars.cc b/torch_xla/csrc/runtime/env_vars.cc
@@ -14,6 +14,7 @@ const char* const kEnvPjRtTpuMaxInflightComputations =
 const char* const kEnvPjrtAsyncCpuClient = "PJRT_CPU_ASYNC_CLIENT";
 const char* const kEnvPjrtAsyncGpuClient = "PJRT_GPU_ASYNC_CLIENT";
 const char* const kEnvTpuLibraryPath = "TPU_LIBRARY_PATH";
+const char* const kEnvInferredTpuLibraryPath = "PTXLA_TPU_LIBRARY_PATH";
 const char* const kEnvXpuLibraryPath = "XPU_LIBRARY_PATH";
 const char* const kEnvNeuronLibraryPath = "NEURON_LIBRARY_PATH";
 const char* const kEnvPjrtDistServiceAddr = "PJRT_DIST_SERVICE_ADDR";

diff --git a/torch_xla/csrc/runtime/env_vars.h b/torch_xla/csrc/runtime/env_vars.h
@@ -24,6 +24,7 @@ extern const char* const kEnvPjRtTpuMaxInflightComputations;
 extern const char* const kEnvPjrtAsyncCpuClient;
 extern const char* const kEnvPjrtAsyncGpuClient;
 extern const char* const kEnvTpuLibraryPath;
+extern const char* const kEnvInferredTpuLibraryPath;
 extern const char* const kEnvXpuLibraryPath;
 extern const char* const kEnvNeuronLibraryPath;
 extern const char* const kEnvPjrtDistServiceAddr;

diff --git a/torch_xla/csrc/runtime/pjrt_computation_client.cc b/torch_xla/csrc/runtime/pjrt_computation_client.cc
@@ -109,10 +109,11 @@ PjRtComputationClient::PjRtComputationClient() {
     client_ = std::move(xla::GetTfrtCpuClient(async, cpu_device_count).value());
   } else if (device_type == "TPU" || device_type == "TPU_C_API") {
     TF_VLOG(1) << "Initializing TFRT TPU client...";
-    XLA_CHECK_OK(
-        pjrt::LoadPjrtPlugin(
-            "tpu", sys_util::GetEnvString(env::kEnvTpuLibraryPath, "libtpu.so"))
-            .status());
+    // Prefer $TPU_LIBRARY_PATH if set
+    auto tpu_library_path = sys_util::GetEnvString(
+        env::kEnvTpuLibraryPath,
+        sys_util::GetEnvString(env::kEnvInferredTpuLibraryPath, "libtpu.so"));
+    XLA_CHECK_OK(pjrt::LoadPjrtPlugin("tpu", tpu_library_path).status());
     tsl::Status tpu_status = pjrt::InitializePjrtPlugin("tpu");
     XLA_CHECK(tpu_status.ok());
     client_ = std::move(xla::GetCApiClient("TPU").value());