kernelmeter 0.4.0__tar.gz → 0.4.2__tar.gz

{kernelmeter-0.4.0/src/kernelmeter.egg-info → kernelmeter-0.4.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: kernelmeter
-Version: 0.4.0
+Version: 0.4.2
 Summary: Query every CUDA device attribute without profiling a kernel, and benchmark your kernels against the hardware's speed of light.
 Author: nuemaan
 License: MIT

{kernelmeter-0.4.0 → kernelmeter-0.4.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "kernelmeter"
-version = "0.4.0"
+version = "0.4.2"
 description = "Query every CUDA device attribute without profiling a kernel, and benchmark your kernels against the hardware's speed of light."
 readme = "README.md"
 license = { text = "MIT" }

{kernelmeter-0.4.0 → kernelmeter-0.4.2}/src/kernelmeter/__init__.py

@@ -8,7 +8,7 @@ from .extras import DeviceExtras
 from .occupancy import Occupancy
 from .peaks import Peaks
 
-__version__ = "0.4.0"
+__version__ = "0.4.2"
 
 __all__ = [
     "BenchResult",

{kernelmeter-0.4.0 → kernelmeter-0.4.2}/src/kernelmeter/attrs.py

@@ -159,9 +159,16 @@ KNOWN_ATTRS: dict[int, str] = {
     146: "host_alloc_dma_buf_supported",
     147: "only_partial_host_native_atomic_supported",
     148: "atomic_reduction_supported",
-    # 149 is CU_DEVICE_ATTRIBUTE_MAX, a sentinel rather than a real
-    # attribute, so it stops here. Anything the driver adds beyond this is
-    # still reported generically as attribute_<id> by the probe below.
+    149: "d3d12_cig_streams_supported",
+    150: "dma_buf_mmap_supported",
+    151: "logical_endpoint_unicast_supported",
+    152: "logical_endpoint_multicast_supported",
+    153: "logical_endpoint_counted_ops_supported",
+    154: "logical_endpoint_unicast_access_on_owner_device_supported",
+    # CU_DEVICE_ATTRIBUTE_MAX (155 as of CUDA 13.x) is a sentinel that
+    # moves up with each toolkit release, not a real attribute. Names stop
+    # at the last defined value; anything newer the driver reports is still
+    # surfaced generically as attribute_<id> by the probe below.
 }
 
 

{kernelmeter-0.4.0 → kernelmeter-0.4.2}/src/kernelmeter/nvml.py

@@ -35,7 +35,9 @@ ARCH_NAMES = {
     10: "Blackwell",
 }
 
-# nvmlBrandType_t (common entries)
+# nvmlBrandType_t. Modern data-center cards report NVML_BRAND_NVIDIA (14)
+# rather than the older TESLA (2), so the 12-16 entries matter on recent
+# drivers -- without them a current Tesla/A/H-series card comes back blank.
 BRAND_NAMES = {
     0: "Unknown",
     1: "Quadro",
@@ -49,6 +51,11 @@ BRAND_NAMES = {
     9: "NVIDIA vCS",
     10: "NVIDIA vWS",
     11: "NVIDIA Cloud Gaming",
+    12: "Quadro RTX",
+    13: "NVIDIA RTX",
+    14: "NVIDIA",
+    15: "GeForce RTX",
+    16: "Titan RTX",
 }
 
 

{kernelmeter-0.4.0 → kernelmeter-0.4.2/src/kernelmeter.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: kernelmeter
-Version: 0.4.0
+Version: 0.4.2
 Summary: Query every CUDA device attribute without profiling a kernel, and benchmark your kernels against the hardware's speed of light.
 Author: nuemaan
 License: MIT

{kernelmeter-0.4.0 → kernelmeter-0.4.2}/tests/test_attrs.py

@@ -20,8 +20,8 @@ def test_unsupported_ids_are_skipped(fake_driver):
 def test_unknown_but_supported_ids_get_generic_names(fake_driver):
     dev = fake_driver.device(0)
     result = attrs.query_all(fake_driver, dev)
-    # id 155 succeeds in the fake but has no name in our table
-    assert result["attribute_155"] == 7
+    # id 160 succeeds in the fake but has no name in our table
+    assert result["attribute_160"] == 7
 
 
 def test_cuda12_range_names(fake_driver):
@@ -29,13 +29,16 @@ def test_cuda12_range_names(fake_driver):
     result = attrs.query_all(fake_driver, dev)
     assert result["numa_id"] == -1
     assert result["gpu_pci_device_id"] == 0x1EB810DE
-    # last named driver attribute before the CU_DEVICE_ATTRIBUTE_MAX sentinel
     assert result["atomic_reduction_supported"] == 1
+    # a CUDA 13.x attribute past the old 0.3.1 table
+    assert result["dma_buf_mmap_supported"] == 1
 
 
 def test_max_sentinel_is_not_named():
-    # 149 is CU_DEVICE_ATTRIBUTE_MAX, not a real attribute
-    assert 149 not in attrs.KNOWN_ATTRS
+    # 155 is CU_DEVICE_ATTRIBUTE_MAX as of CUDA 13.x, not a real attribute
+    assert 155 not in attrs.KNOWN_ATTRS
+    # the last real attribute we name
+    assert attrs.KNOWN_ATTRS[154] == "logical_endpoint_unicast_access_on_owner_device_supported"
 
 
 def test_device_metadata(fake_driver):

{kernelmeter-0.4.0 → kernelmeter-0.4.2}/tests/test_extras.py

@@ -10,7 +10,7 @@ def _fake_nvml():
 def test_from_nvml_builds_extras():
     ex = extras.from_nvml(_fake_nvml(), 0)
     assert ex.architecture == "Turing"
-    assert ex.brand == "Tesla"
+    assert ex.brand == "NVIDIA"
     assert ex.num_gpu_cores == 2560
     assert ex.memory_total_bytes == 15843721216
     assert ex.pcie_gen_current == 3

{kernelmeter-0.4.0 → kernelmeter-0.4.2}/tests/test_nvml.py

@@ -51,7 +51,7 @@ class FakeNvmlLib:
         return NVML_SUCCESS
 
     def nvmlDeviceGetBrand(self, handle, ptr):
-        ptr._obj.value = 2  # Tesla
+        ptr._obj.value = 14  # NVML_BRAND_NVIDIA, what modern T4 drivers report
         return NVML_SUCCESS
 
     def nvmlDeviceGetNumGpuCores(self, handle, ptr):
@@ -119,7 +119,7 @@ def test_static_device_facts():
     n = nvml.Nvml(lib=FakeNvmlLib())
     h = n.device(0)
     assert nvml.ARCH_NAMES[n.architecture(h)] == "Turing"
-    assert nvml.BRAND_NAMES[n.brand(h)] == "Tesla"
+    assert nvml.BRAND_NAMES[n.brand(h)] == "NVIDIA"  # value 14, the modern range
     assert n.num_gpu_cores(h) == 2560
     total, free, used = n.memory_info(h)
     assert total == 15843721216