Skip to content

A GPU task manager for Windows, Linux, and WSL: per-process VRAM, GPU activity, and LLM model/KV-cache insights (Windows AMD and NVIDIA).

Python

Merge main into skip-nvml-non-nvidia

Scott Hanselman43ec72bfbf0e
diff --cc src/gputaskman/collector.py
index 6528b69,ea15405..1c73c86
--- a/src/gputaskman/collector.py
+++ b/src/gputaskman/collector.py
@@@ -11,9 -12,10 +12,10 @@@ import psuti
from .amdsmi import AmdSmi, _bdf
from .adl import Adl
-from .dxgi import VENDOR_NVIDIA, Adapter
+from .dxgi import VENDOR_AMD, VENDOR_NVIDIA, Adapter
from .handy import HandyEvidence, inspect_handy
from .llm import LlmView, ServerProber, build_view, gb, ollama_log_path
+ from . import nvml as nvml_mod
from .nvml import Nvml, NvDevice
from .pdh import ProcGpu
from .procinfo import AI_KINDS, ProcDetail, ProcInspector, classify
diff --cc tests/test_core.py
index 9fe8b84,4a1e709..db352b4
--- a/tests/test_core.py
+++ b/tests/test_core.py
@@@ -249,27 -250,25 +250,46 @@@ class MemoryTests(unittest.TestCase)
Console(file=output, width=160).print(GpuTaskManager().render_device(s))
self.assertIn("could not be matched", output.getvalue())
+ def test_nvml_is_swept_only_for_an_nvidia_adapter(self):
+ amd_telemetry = NvDevice(index=0, source="AMD ADL", util_gpu=5)
+ for vendor, swept, expected in ((0x1002, False, amd_telemetry), (0x8086, False, None),
+ (0x10DE, True, None)):
+ with self.subTest(vendor=hex(vendor)):
+ c = Collector.__new__(Collector)
+ c.adapter = Adapter(1, "GPU", vendor, 16 * GiB, 0, False, (1, 0, 0))
+ c.adapters = [c.adapter]
+ c.gpu_index, c.scan_budget, c.host, c.host_status = 0, None, None, ""
+ c.nvml = Mock(ok=True)
+ c.nvml.sample.return_value = []
+ c.adl, c.amdsmi = Mock(), Mock()
+ c.adl.sample.return_value = amd_telemetry
+ c.counters = Mock(outside={})
+ c.counters.sample.return_value = ({1: {}}, {1: AdapterGpu(dedicated=0, shared=0)})
+ c.inspector, c.prober = Mock(), Mock()
+ with patch("gputaskman.collector.BACKEND", "wddm"):
+ s = c.sample()
+ self.assertEqual(c.nvml.sample.called, swept)
+ self.assertIs(s.nv, expected)
+
+ def test_lost_nvml_telemetry_is_explained_not_silent(self):
+ c = Collector.__new__(Collector)
+ c.adapter = Adapter(1, "RTX 3060", 0x10DE, 12 * GiB, 0, False, (5, 0, 0))
+ c.adapters = [c.adapter]
+ c.gpu_index, c.scan_budget, c.host, c.host_status = 0, None, None, ""
+ c.nvml = Mock(ok=True, driver="", cuda_version="")
+ c.nvml.sample.return_value = []
+ c.counters = Mock(outside={})
+ c.counters.sample.return_value = ({1: {}}, {1: AdapterGpu(dedicated=GiB, shared=0)})
+ c.inspector, c.prober = Mock(), Mock()
+ with patch("gputaskman.collector.BACKEND", "wddm"), \
+ patch("gputaskman.collector.sys.platform", "win32"), \
+ patch("gputaskman.nvml.LOST", {"00000000:05:00.0"}):
+ s = c.sample()
+ self.assertIsNone(s.nv)
+ self.assertTrue(s.nv_lost)
+ self.assertIn("restart gputaskman", s.counter_status)
+ self.assertEqual(s.vram_used, GiB)
+
def test_windows_matching_counter_zero_is_not_unavailable(self):
c = Collector.__new__(Collector)
c.adapter = Adapter(1, "GPU", 0x10DE, 16 * GiB, 0, False)