[rocprofiler-compute] Only depend on amdsmi in profile phase (#2044)

* Only depepnd on amdsmi in profile phase

* amdsmi interface tests should have common prefix for easier testing

projects/rocprofiler-compute/src/utils/amdsmi_interface.py

@@ -34,18 +34,35 @@ from utils.logger import (
     console_warning,
 )
 
-sys.path.insert(0, os.getenv("ROCM_PATH", "/opt/rocm") + "/share/amd_smi")
+_amdsmi_module = None
 
-try:
+
-    import amdsmi
+# Ignore undefined name amdsmi since it's dynamically imported
-except ImportError as e:
+def import_amdsmi_module() -> "amdsmi":  # noqa: F821
-    console_warning(f"Unhandled import error: {e}")
+    """
-    console_error("Failed to import the amdsmi Python library.")
+    Dynamically import the amdsmi module because we only
+    want profile time dependency on amdsmi.
+    Uses global cache to avoid repeated imports.
+    """
+    global _amdsmi_module
+
+    if not _amdsmi_module:
+        sys.path.insert(0, os.getenv("ROCM_PATH", "/opt/rocm") + "/share/amd_smi")
+        try:
+            import amdsmi
+
+            _amdsmi_module = amdsmi
+        except ImportError as e:
+            console_warning(f"Unhandled import error: {e}")
+            console_error("Failed to import the amdsmi Python library.")
+
+    return _amdsmi_module
 
 
 @contextmanager
 def amdsmi_ctx() -> Iterator[None]:
     """Context manager to initialize and shutdown amdsmi."""
+    amdsmi = import_amdsmi_module()
     try:
         amdsmi.amdsmi_init()
         yield
@@ -58,8 +75,10 @@ def amdsmi_ctx() -> Iterator[None]:
             console_warning(f"amd-smi shutdown failed: {e}")
 
 
-def get_device_handle() -> "amdsmi.ProcessorHandle | None":
+# Ignore undefined name amdsmi since it's dynamically imported
+def get_device_handle() -> "amdsmi.ProcessorHandle | None":  # noqa: F821
     """Get the first AMD device handle."""
+    amdsmi = import_amdsmi_module()
     try:
         devices = amdsmi.amdsmi_get_processor_handles()
         if len(devices) == 0:
@@ -74,6 +93,7 @@ def get_device_handle() -> "amdsmi.ProcessorHandle | None":
 
 def get_mem_max_clock() -> float:
     """Get the maximum memory clock of the device."""
+    amdsmi = import_amdsmi_module()
     try:
         return amdsmi.amdsmi_get_clock_info(
             get_device_handle(), amdsmi.AmdSmiClkType.GFX
@@ -85,6 +105,7 @@ def get_mem_max_clock() -> float:
 
 def get_gpu_model() -> str:
     """Get the GPU model name."""
+    amdsmi = import_amdsmi_module()
     try:
         gpu_model_info = (
             # board -> product_name
@@ -103,6 +124,7 @@ def get_gpu_model() -> str:
 
 def get_gpu_vbios_part_number() -> str:
     """Get the GPU VBIOS part number."""
+    amdsmi = import_amdsmi_module()
     try:
         vbios_part_number = amdsmi.amdsmi_get_gpu_vbios_info(get_device_handle())[
             "part_number"
@@ -116,6 +138,7 @@ def get_gpu_vbios_part_number() -> str:
 
 def get_gpu_compute_partition() -> str:
     """Get the GPU compute partition."""
+    amdsmi = import_amdsmi_module()
     try:
         compute_partition = amdsmi.amdsmi_get_gpu_compute_partition(get_device_handle())
         console_debug(f"GPU Compute Partition: {compute_partition}")
@@ -127,6 +150,7 @@ def get_gpu_compute_partition() -> str:
 
 def get_gpu_memory_partition() -> str:
     """Get the GPU memory partition."""
+    amdsmi = import_amdsmi_module()
     try:
         memory_partition = amdsmi.amdsmi_get_gpu_memory_partition(get_device_handle())
         console_debug(f"GPU Memory Partition: {memory_partition}")
@@ -138,6 +162,7 @@ def get_gpu_memory_partition() -> str:
 
 def get_amdgpu_driver_version() -> str:
     """Get the AMDGPU driver version."""
+    amdsmi = import_amdsmi_module()
     try:
         driver_info = amdsmi.amdsmi_get_gpu_driver_info(get_device_handle())
         driver_version = driver_info["driver_version"]
@@ -150,6 +175,7 @@ def get_amdgpu_driver_version() -> str:
 
 def get_gpu_vram_size() -> int:
     """Get the GPU VRAM size in MB."""
+    amdsmi = import_amdsmi_module()
     try:
         vram_info = amdsmi.amdsmi_get_gpu_vram_info(get_device_handle())
         vram_size = str(int(vram_info["vram_size"]) * 1024)  # MB -> KB

projects/rocprofiler-compute/tests/test_utils.py

@@ -8552,7 +8552,9 @@ def test_list_metrics(binary_handler_analyze_rocprof_compute, capsys):
 
 
 def test_amdsmi_ctx():
-    from utils.amdsmi_interface import amdsmi_ctx
+    from utils.amdsmi_interface import amdsmi_ctx, import_amdsmi_module
+
+    _ = import_amdsmi_module()
 
     with mock.patch("amdsmi.amdsmi_init") as amdsmi_init_mock:
         with mock.patch("amdsmi.amdsmi_shut_down") as amdsmi_shutdown_mock:
@@ -8561,8 +8563,10 @@ def test_amdsmi_ctx():
             amdsmi_shutdown_mock.assert_called_once()
 
 
-def test_get_device_handle():
+def test_amdsmi_get_device_handle():
-    from utils.amdsmi_interface import get_device_handle
+    from utils.amdsmi_interface import get_device_handle, import_amdsmi_module
+
+    _ = import_amdsmi_module()
 
     with mock.patch("amdsmi.amdsmi_get_processor_handles") as device_handles_mock:
         device_handles_mock.return_value = [12345]
@@ -8576,8 +8580,10 @@ def test_get_device_handle():
         assert handle is None
 
 
-def test_get_mem_max_clock():
+def test_amdsmi_get_mem_max_clock():
-    from utils.amdsmi_interface import get_mem_max_clock
+    from utils.amdsmi_interface import get_mem_max_clock, import_amdsmi_module
+
+    _ = import_amdsmi_module()
 
     with mock.patch("amdsmi.amdsmi_get_processor_handles") as device_handles_mock:
         device_handles_mock.return_value = [12345]
@@ -8588,8 +8594,10 @@ def test_get_mem_max_clock():
             assert clk == 100
 
 
-def test_get_gpu_model():
+def test_amdsmi_get_gpu_model():
-    from utils.amdsmi_interface import get_gpu_model
+    from utils.amdsmi_interface import get_gpu_model, import_amdsmi_module
+
+    _ = import_amdsmi_module()
 
     with mock.patch("amdsmi.amdsmi_get_processor_handles") as device_handles_mock:
         device_handles_mock.return_value = [12345]
@@ -8610,8 +8618,10 @@ def test_get_gpu_model():
             assert model == "N/A"
 
 
-def test_get_gpu_vbios_part_number():
+def test_amdsmi_get_gpu_vbios_part_number():
-    from utils.amdsmi_interface import get_gpu_vbios_part_number
+    from utils.amdsmi_interface import get_gpu_vbios_part_number, import_amdsmi_module
+
+    _ = import_amdsmi_module()
 
     with mock.patch("amdsmi.amdsmi_get_processor_handles") as device_handles_mock:
         device_handles_mock.return_value = [12345]
@@ -8630,8 +8640,10 @@ def test_get_gpu_vbios_part_number():
             assert part_number == "N/A"
 
 
-def test_get_gpu_compute_partition():
+def test_amdsmi_get_gpu_compute_partition():
-    from utils.amdsmi_interface import get_gpu_compute_partition
+    from utils.amdsmi_interface import get_gpu_compute_partition, import_amdsmi_module
+
+    _ = import_amdsmi_module()
 
     with mock.patch("amdsmi.amdsmi_get_processor_handles") as device_handles_mock:
         device_handles_mock.return_value = [12345]
@@ -8651,8 +8663,10 @@ def test_get_gpu_compute_partition():
             assert partition == "N/A"
 
 
-def test_get_gpu_memory_partition():
+def test_amdsmi_get_gpu_memory_partition():
-    from utils.amdsmi_interface import get_gpu_memory_partition
+    from utils.amdsmi_interface import get_gpu_memory_partition, import_amdsmi_module
+
+    _ = import_amdsmi_module()
 
     with mock.patch("amdsmi.amdsmi_get_processor_handles") as device_handles_mock:
         device_handles_mock.return_value = [12345]