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: use the Vulkan llama.cpp bundle, not CPU, on Windows NVIDIA drivers below CUDA 12.4 (#13129)

Daniel Han236dc94c33df
diff --git a/studio/install_llama_prebuilt.py b/studio/install_llama_prebuilt.py
index 4f773cc71b3..a8214ce9d0e 100644
--- a/studio/install_llama_prebuilt.py
+++ b/studio/install_llama_prebuilt.py
@@ -4219,14 +4219,29 @@ def resolve_release_asset_choice(
and selection_host.driver_cuda_version[0] >= _MIN_CUDA_MAJOR
):
# #12842: the Windows source fallback needs a toolkit this driver runs, which
- # winget rarely offers, so setup would fail. CPU until the driver is updated.
- cpu_choice = published_asset_choice_for_kind(release, "windows-cpu")
- if cpu_choice is not None:
+ # winget rarely offers, so setup would fail. Vulkan still runs on the NVIDIA
+ # GPU (the reporter's Vulkan build worked); CPU only after it. Not under a CUDA
+ # mask: Vulkan ignores it, so it could take a card the caller hid.
+ vulkan_choice = (
+ published_asset_choice_for_kind(release, "windows-vulkan", host = host)
+ if masked_host is None
+ else None
+ )
+ choices = [
+ choice
+ for choice in (
+ vulkan_choice,
+ published_asset_choice_for_kind(release, "windows-cpu"),
+ )
+ if choice is not None
+ ]
+ if choices:
+ bundle = "Vulkan" if vulkan_choice is not None else "CPU"
log(
- f"{cuda_driver_floor_message(selection_host)} Installing the CPU "
+ f"{cuda_driver_floor_message(selection_host)} Installing the {bundle} "
"bundle for now."
)
- return apply_approved_hashes([cpu_choice], checksums)
+ return apply_approved_hashes(choices, checksums)
torch_preference = detect_torch_cuda_runtime_preference(
selection_host, gpu_hidden_by_mask = masked_host is not None
)
diff --git a/tests/studio/install/test_cuda_driver_floor_12842.py b/tests/studio/install/test_cuda_driver_floor_12842.py
index e54d9a29596..513535bcb29 100644
--- a/tests/studio/install/test_cuda_driver_floor_12842.py
+++ b/tests/studio/install/test_cuda_driver_floor_12842.py
@@ -203,37 +203,72 @@ def test_a_driver_below_the_floor_stops_the_release_walk(monkeypatch, system):
assert len(walked) == 1
-def _release_with_windows_cpu():
- cpu = m.parse_published_artifact(
- {
- "asset_name": f"app-{TAG}-windows-x64-cpu.zip",
- "install_kind": "windows-cpu",
- "bundle_profile": "windows-cpu-x64",
- "runtime_line": None,
- "coverage_class": None,
- "rank": 1000,
- }
- )
- assert cpu is not None
+def _release_with(*kinds):
+ rows = {
+ "windows-vulkan": ("windows-vulkan-x64", 60),
+ "windows-cpu": ("windows-cpu-x64", 1000),
+ }
+ extra = []
+ for kind in kinds:
+ profile, rank = rows[kind]
+ artifact = m.parse_published_artifact(
+ {
+ "asset_name": f"app-{TAG}-{kind.replace('windows-', 'windows-x64-')}.zip",
+ "install_kind": kind,
+ "bundle_profile": profile,
+ "runtime_line": None,
+ "coverage_class": None,
+ "rank": rank,
+ }
+ )
+ assert artifact is not None
+ extra.append(artifact)
return m.PublishedReleaseBundle(
repo = RELEASE.repo,
release_tag = RELEASE.release_tag,
upstream_tag = RELEASE.upstream_tag,
- assets = {**RELEASE.assets, cpu.asset_name: f"https://example.com/{cpu.asset_name}"},
- artifacts = [*RELEASE.artifacts, cpu],
+ assets = {
+ **RELEASE.assets,
+ **{a.asset_name: f"https://example.com/{a.asset_name}" for a in extra},
+ },
+ artifacts = [*RELEASE.artifacts, *extra],
)
-class TestWindowsBelowTheFloorGetsTheCpuBundle:
- def _choices(self, monkeypatch, driver):
+class TestWindowsBelowTheFloorGetsTheVulkanBundle:
+ def _choices(
+ self,
+ monkeypatch,
+ driver,
+ kinds = ("windows-vulkan", "windows-cpu"),
+ ):
monkeypatch.setattr(m, "apply_approved_hashes", lambda attempts, _checksums: attempts)
monkeypatch.setattr(m, "github_release_assets", lambda _repo, _tag: {})
return m.resolve_release_asset_choice(
- host("Windows", driver), "b11443", _release_with_windows_cpu(), None
+ host("Windows", driver), "b11443", _release_with(*kinds), None
)
- def test_a_12_2_driver_gets_the_cpu_bundle(self, monkeypatch):
+ def test_a_12_2_driver_gets_vulkan_then_cpu(self, monkeypatch):
choices = self._choices(monkeypatch, (12, 2))
+ assert [c.install_kind for c in choices] == ["windows-vulkan", "windows-cpu"]
+ assert choices[0].name == f"app-{TAG}-windows-x64-vulkan.zip"
+
+ def test_a_masked_gpu_stays_off_vulkan(self, monkeypatch):
+ # Vulkan ignores CUDA_VISIBLE_DEVICES, so a hidden card must not be reached through it.
+ monkeypatch.setattr(m, "apply_approved_hashes", lambda attempts, _checksums: attempts)
+ masked = m.dataclasses_replace(
+ host("Windows", (12, 2)),
+ compute_caps = [],
+ visible_cuda_devices = "-1",
+ has_usable_nvidia = False,
+ )
+ choices = m.resolve_release_asset_choice(
+ masked, "b11443", _release_with("windows-vulkan", "windows-cpu"), None
+ )
+ assert [c.install_kind for c in choices] == ["windows-cpu"]
+
+ def test_without_a_vulkan_bundle_it_gets_the_cpu_bundle(self, monkeypatch):
+ choices = self._choices(monkeypatch, (12, 2), kinds = ("windows-cpu",))
assert [c.install_kind for c in choices] == ["windows-cpu"]
@pytest.mark.parametrize("driver", [(12, 4), (13, 0)])