Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio: exclude NVIDIA NPU from GPU memory accounting (#15334)

Daniel Han5c2b1fb12a59
diff --git a/studio/backend/tests/test_dgx_spark_gpu_inventory.py b/studio/backend/tests/test_dgx_spark_gpu_inventory.py
index 91288723b63..a0c3c94116d 100644
--- a/studio/backend/tests/test_dgx_spark_gpu_inventory.py
+++ b/studio/backend/tests/test_dgx_spark_gpu_inventory.py
@@ -18,6 +18,36 @@ SPARK_TOTAL_BYTES = 124609 * MIB
SPARK_TOTAL_GB = round(SPARK_TOTAL_BYTES / GIB, 2)
+def test_spark_npu_is_not_a_cuda_gpu_even_when_smi_lists_it(monkeypatch):
+ monkeypatch.setattr(
+ nvidia.gpu_query,
+ "run_nvidia_smi",
+ lambda *a, **k: types.SimpleNamespace(
+ returncode = 0,
+ stdout = "0, NVIDIA RTX Spark N1X, 8128\n1, NVIDIA NPU, [N/A]\n",
+ ),
+ )
+ result = nvidia.get_backend_visible_gpu_info([0, 1], None)
+ assert result["available"] is True
+ assert [(d["index"], d["name"]) for d in result["devices"]] == [(0, "NVIDIA RTX Spark N1X")]
+ assert len(nvidia.get_physical_gpu_inventory()["devices"]) == 1
+
+
+def test_npu_exclusion_keeps_real_gpus_with_unknown_memory_and_physical_ids(monkeypatch):
+ monkeypatch.setattr(
+ nvidia.gpu_query,
+ "run_nvidia_smi",
+ lambda *a, **k: types.SimpleNamespace(
+ returncode = 0,
+ stdout = "0, NVIDIA NPU, [N/A]\n1, NVIDIA GB10, [N/A]\n2, NVIDIA H100, 81920\n",
+ ),
+ )
+ rows = nvidia._query_gpu_inventory("test")
+ assert [d["index"] for d in rows] == [1, 2]
+ assert rows[0]["memory_total_gb"] is None
+ assert rows[1]["memory_total_gb"] == 80
+
+
class _SparkProps:
"""cudaDeviceProp as torch surfaces it for a GB10."""
diff --git a/studio/backend/tests/test_system_vulkan_gpu_info.py b/studio/backend/tests/test_system_vulkan_gpu_info.py
index 78394b761c0..c77a15c2408 100644
--- a/studio/backend/tests/test_system_vulkan_gpu_info.py
+++ b/studio/backend/tests/test_system_vulkan_gpu_info.py
@@ -12,6 +12,48 @@ def _shared_setup_1(monkeypatch):
monkeypatch.setattr(main, "_system_gpu_cache", None)
+def test_system_monitor_ignores_spark_npu_telemetry(monkeypatch):
+ import utils.hardware as hardware
+ from utils.hardware import nvidia
+ from core.inference.llama_cpp import LlamaCppBackend
+
+ monkeypatch.setattr(
+ nvidia.gpu_query,
+ "run_nvidia_smi",
+ lambda *a, **k: SimpleNamespace(
+ returncode = 0,
+ stdout = "0, NVIDIA RTX Spark N1X, 46477\n1, NVIDIA NPU, [N/A]\n",
+ ),
+ )
+ monkeypatch.setattr(
+ hardware,
+ "get_backend_visible_gpu_info",
+ lambda: {**nvidia.get_backend_visible_gpu_info([0, 1], None), "backend": "cuda"},
+ )
+ monkeypatch.setattr(
+ hardware,
+ "get_visible_gpu_utilization",
+ lambda: {
+ "backend": "cuda",
+ "devices": [
+ {"index": 0, "vram_total_gb": 45.39, "vram_used_gb": 20.5},
+ {"index": 1, "vram_total_gb": None, "vram_used_gb": None},
+ ],
+ },
+ )
+ monkeypatch.setattr(hardware, "get_vulkan_inference_gpu_info", lambda: None)
+ monkeypatch.setattr(hardware, "get_cross_vendor_inference_gpu_info", lambda: None)
+ monkeypatch.setattr(LlamaCppBackend, "_is_vulkan_backend", staticmethod(lambda: False))
+ monkeypatch.setattr(main, "_system_gpu_cache", None)
+
+ gpu, inference_gpu = main._get_cached_system_gpu_info(SimpleNamespace(debug = lambda *a: None))
+ # Both monitor views read this inventory and sum its per-device usage.
+ assert len(gpu["devices"]) == 1
+ assert gpu["devices"][0]["index"] == 0
+ assert sum(d["vram_used_gb"] for d in gpu["devices"]) == 20.5
+ assert inference_gpu["devices"] == gpu["devices"]
+
+
def test_system_gpu_info_preserves_vulkan_visibility_metrics(monkeypatch):
import utils.hardware as hardware
diff --git a/studio/backend/utils/hardware/nvidia.py b/studio/backend/utils/hardware/nvidia.py
index db49d8104f5..2b73fcd97a4 100644
--- a/studio/backend/utils/hardware/nvidia.py
+++ b/studio/backend/utils/hardware/nvidia.py
@@ -543,6 +543,9 @@ def _query_gpu_inventory(caller: str) -> Any:
continue
# Rejoin in case the GPU name contains commas
name = parts[1] if len(parts) == 3 else ", ".join(parts[1:-1])
+ # RTX Spark's MCDM NPU: listed by nvidia-smi, not a CUDA GPU, no memory metrics.
+ if name.casefold() == "nvidia npu":
+ continue
# A capacity this build of nvidia-smi will not report ("[N/A]", as _parse_smi_value already recognises) is a missing metric, not a missing card. Dropping the row hid the GPU from the whole inventory, and the Linux procfs fallback does not run either since it answers for a query that FAILED rather than one that came back short: the host lost its mismatch and its repair guidance over an unknown size.
mem_total_mb = _parse_smi_value(parts[-1])
rows.append(