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 tests: keep GGUF loads off the host's Metal budget unless a test asks (#15910)

Daniel Handa8ec323b19c
diff --git a/studio/backend/tests/conftest.py b/studio/backend/tests/conftest.py
index 0308c38c46d..d96adf5c9fd 100644
--- a/studio/backend/tests/conftest.py
+++ b/studio/backend/tests/conftest.py
@@ -590,17 +590,43 @@ def _hf_cache_is_empty(_empty_hf_hub_cache, monkeypatch):
monkeypatch.setattr(constants, "HF_HUB_CACHE", _empty_hf_hub_cache)
+_REAL_APPLE_METAL_BUDGET = None
+
+
@pytest.fixture(autouse = True)
def _no_live_metal_wired_ceiling(monkeypatch):
"""Keep Metal context verdicts off the host's live GPU memory."""
+ import inspect
+
from core.inference.llama_cpp import LlamaCppBackend
+ global _REAL_APPLE_METAL_BUDGET
+ if _REAL_APPLE_METAL_BUDGET is None:
+ _REAL_APPLE_METAL_BUDGET = inspect.getattr_static(
+ LlamaCppBackend, "_apple_metal_memory_budget_bytes"
+ )
monkeypatch.setattr(
LlamaCppBackend, "_apple_metal_wired_ceiling_bytes", staticmethod(lambda: 0)
)
monkeypatch.setattr(
LlamaCppBackend, "_apple_metal_wired_headroom_bytes", staticmethod(lambda: None)
)
+ # On a Mac the real budget is non-zero, which sends every GGUF load down the Metal
+ # context-fit path and adds its "cannot size the KV cache" notice to loads that tests
+ # written against a CPU or CUDA host expect to be silent. Tests of the budget itself,
+ # or of the Metal path, set it explicitly or ask for real_apple_metal_budget.
+ monkeypatch.setattr(
+ LlamaCppBackend, "_apple_metal_memory_budget_bytes", staticmethod(lambda: 0)
+ )
+
+
+def real_apple_metal_budget(monkeypatch):
+ """The real budget probe, for the tests of the probe itself."""
+ from core.inference.llama_cpp import LlamaCppBackend
+ monkeypatch.setattr(
+ LlamaCppBackend, "_apple_metal_memory_budget_bytes", _REAL_APPLE_METAL_BUDGET
+ )
@pytest.fixture(autouse = True)
diff --git a/studio/backend/tests/test_health_reports_unified_memory.py b/studio/backend/tests/test_health_reports_unified_memory.py
index 238facd5af8..faeb2daf660 100644
--- a/studio/backend/tests/test_health_reports_unified_memory.py
+++ b/studio/backend/tests/test_health_reports_unified_memory.py
@@ -107,7 +107,7 @@ def test_an_unauthenticated_reply_carries_neither(monkeypatch):
assert "apple_silicon" not in body
-def test_the_gate_is_the_one_the_metal_budget_uses(monkeypatch):
+def test_the_gate_is_the_one_the_metal_budget_uses(monkeypatch, real_apple_metal_budget):
"""Not a second definition of Apple Silicon. If these ever diverge, the UI starts
describing a memory model the loader is not enforcing."""
import inspect
diff --git a/studio/backend/tests/test_llama_cpp_context_fit.py b/studio/backend/tests/test_llama_cpp_context_fit.py
index e5825bffc8a..6ab55a361ea 100644
--- a/studio/backend/tests/test_llama_cpp_context_fit.py
+++ b/studio/backend/tests/test_llama_cpp_context_fit.py
@@ -667,6 +667,7 @@ def _install_fake_psutil(
monkeypatch.setitem(sys.modules, "psutil", fake)
[email protected]("real_apple_metal_budget")
class TestAppleUnifiedMemoryBudget:
def test_zero_off_apple_silicon(self, monkeypatch):
import platform as _platform
diff --git a/studio/backend/tests/test_loads_stay_off_the_host_metal_budget.py b/studio/backend/tests/test_loads_stay_off_the_host_metal_budget.py
new file mode 100644
index 00000000000..d849f1da4cd
--- /dev/null
+++ b/studio/backend/tests/test_loads_stay_off_the_host_metal_budget.py
@@ -0,0 +1,25 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""The suite runs on a macos-15 leg too. There the real unified-memory budget is non-zero, and
+every GGUF load a test drives takes the Metal context-fit path, which the tests were not
+written for. conftest pins the budget to 0; these check the pin holds on any host."""
+
+import utils.hardware as hardware
+import utils.hardware.hardware as hardware_impl
+from core.inference.llama_cpp import LlamaCppBackend
+
+
+def _host_is_apple_silicon(monkeypatch):
+ monkeypatch.setattr(hardware, "is_apple_silicon", lambda: True)
+ monkeypatch.setattr(hardware_impl, "_metal_recommended_working_set_bytes", lambda: 16 * 1024**3)
+
+
+def test_a_load_on_apple_silicon_sees_no_metal_budget_by_default(monkeypatch):
+ _host_is_apple_silicon(monkeypatch)
+ assert LlamaCppBackend._apple_metal_memory_budget_bytes() == 0
+
+
+def test_the_real_probe_is_back_when_a_test_asks_for_it(monkeypatch, real_apple_metal_budget):
+ _host_is_apple_silicon(monkeypatch)
+ assert LlamaCppBackend._apple_metal_memory_budget_bytes() > 0