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 sd.cpp: pin the prebuilt with the aarch64 K-quant GEMM and macOS Accelerate routing (#17489)

Daniel Han8da0386bce25
diff --git a/studio/backend/tests/test_diffusion_more_families.py b/studio/backend/tests/test_diffusion_more_families.py
index 072651fc340..3eaab1e8b3a 100644
--- a/studio/backend/tests/test_diffusion_more_families.py
+++ b/studio/backend/tests/test_diffusion_more_families.py
@@ -761,7 +761,7 @@ def test_the_pinned_prebuilt_is_one_that_can_load_qwen_image_21():
module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(module)
- assert module.DEFAULT_TAG == "master-982-0518b34-ueb0e729", (
+ assert module.DEFAULT_TAG == "master-982-0518b34-u67cba2c", (
"the pinned prebuilt must be one built from a tree carrying Qwen-Image-2.1, FLUX.3 and MOSS-Image; "
f"{module.DEFAULT_TAG} is not"
)
diff --git a/studio/backend/tests/test_sd_cpp_args.py b/studio/backend/tests/test_sd_cpp_args.py
index 50d12748908..0ab5e365a0a 100644
--- a/studio/backend/tests/test_sd_cpp_args.py
+++ b/studio/backend/tests/test_sd_cpp_args.py
@@ -1061,8 +1061,9 @@ def _all_builders(binary, extra_args = None):
@pytest.mark.parametrize(
"tag, expected",
[
- ("master-982-0518b34-ueb0e729", True), # the pinned bundle
- ("master-982-0518b34-uba6ac30", True), # the previous pin
+ ("master-982-0518b34-u67cba2c", True), # the pinned bundle
+ ("master-982-0518b34-ueb0e729", True), # the previous pin
+ ("master-982-0518b34-uba6ac30", True), # an older pin
("master-982-0518b34-uaba3a1c", True), # an older pin
("master-982-0518b34-ud331c66", True), # an older pin
("master-968-8cd731e-ubdb47d1", True), # the pin before it
@@ -1130,7 +1131,7 @@ def _cpu_install(
binary = root / "bin" / "sd-cli"
binary.write_text("")
(root / ".unsloth-sd-cpp-install.json").write_text(
- json.dumps({"tag": "master-982-0518b34-ueb0e729", "accelerator": accelerator})
+ json.dumps({"tag": "master-982-0518b34-u67cba2c", "accelerator": accelerator})
)
for fn in (
_sd_cpp_args_module._installed_sd_cpp_tag,
diff --git a/studio/backend/tests/test_sd_cpp_cuda_cpu_torch.py b/studio/backend/tests/test_sd_cpp_cuda_cpu_torch.py
index 43c424aeb75..d551e44404b 100644
--- a/studio/backend/tests/test_sd_cpp_cuda_cpu_torch.py
+++ b/studio/backend/tests/test_sd_cpp_cuda_cpu_torch.py
@@ -247,29 +247,29 @@ def test_selection_installs_the_cuda_build(monkeypatch):
_PINNED_RELEASE_ASSETS = [
- "sd-master-982-0518b34-ueb0e729-bin-Linux-Ubuntu-22.04-x86_64-cuda12.zip",
- "sd-master-982-0518b34-ueb0e729-bin-Linux-Ubuntu-22.04-x86_64-cuda13.zip",
- "sd-master-982-0518b34-ueb0e729-bin-Linux-Ubuntu-22.04-x86_64-vulkan.zip",
- "sd-master-982-0518b34-ueb0e729-bin-Linux-Ubuntu-22.04-x86_64.zip",
- "sd-master-982-0518b34-ueb0e729-bin-Linux-Ubuntu-24.04-aarch64.zip",
- "sd-master-982-0518b34-ueb0e729-bin-win-cpu-arm64.zip",
- "sd-master-982-0518b34-ueb0e729-bin-win-cpu-x64.zip",
- "sd-master-982-0518b34-ueb0e729-bin-win-cuda12-x64.zip",
- "sd-master-982-0518b34-ueb0e729-bin-win-cuda13-x64.zip",
- "sd-master-982-0518b34-ueb0e729-bin-win-vulkan-x64.zip",
+ "sd-master-982-0518b34-u67cba2c-bin-Linux-Ubuntu-22.04-x86_64-cuda12.zip",
+ "sd-master-982-0518b34-u67cba2c-bin-Linux-Ubuntu-22.04-x86_64-cuda13.zip",
+ "sd-master-982-0518b34-u67cba2c-bin-Linux-Ubuntu-22.04-x86_64-vulkan.zip",
+ "sd-master-982-0518b34-u67cba2c-bin-Linux-Ubuntu-22.04-x86_64.zip",
+ "sd-master-982-0518b34-u67cba2c-bin-Linux-Ubuntu-24.04-aarch64.zip",
+ "sd-master-982-0518b34-u67cba2c-bin-win-cpu-arm64.zip",
+ "sd-master-982-0518b34-u67cba2c-bin-win-cpu-x64.zip",
+ "sd-master-982-0518b34-u67cba2c-bin-win-cuda12-x64.zip",
+ "sd-master-982-0518b34-u67cba2c-bin-win-cuda13-x64.zip",
+ "sd-master-982-0518b34-u67cba2c-bin-win-vulkan-x64.zip",
]
@pytest.mark.parametrize(
"system,machine,major,expected",
[
- ("Windows", "AMD64", "13", "sd-master-982-0518b34-ueb0e729-bin-win-cuda13-x64.zip"),
- ("Windows", "AMD64", "12", "sd-master-982-0518b34-ueb0e729-bin-win-cuda12-x64.zip"),
+ ("Windows", "AMD64", "13", "sd-master-982-0518b34-u67cba2c-bin-win-cuda13-x64.zip"),
+ ("Windows", "AMD64", "12", "sd-master-982-0518b34-u67cba2c-bin-win-cuda12-x64.zip"),
(
"Linux",
"x86_64",
"13",
- "sd-master-982-0518b34-ueb0e729-bin-Linux-Ubuntu-22.04-x86_64-cuda13.zip",
+ "sd-master-982-0518b34-u67cba2c-bin-Linux-Ubuntu-22.04-x86_64-cuda13.zip",
),
("Linux", "aarch64", "13", None),
("Windows", "ARM64", "13", None),
@@ -292,8 +292,8 @@ def test_the_pinned_release_has_the_cuda_bundle_it_routes_to(system, machine, ma
@pytest.mark.parametrize(
"machine,expected",
[
- ("ARM64", "sd-master-982-0518b34-ueb0e729-bin-win-cpu-arm64.zip"),
- ("AMD64", "sd-master-982-0518b34-ueb0e729-bin-win-cpu-x64.zip"),
+ ("ARM64", "sd-master-982-0518b34-u67cba2c-bin-win-cpu-arm64.zip"),
+ ("AMD64", "sd-master-982-0518b34-u67cba2c-bin-win-cpu-x64.zip"),
],
)
@pytest.mark.parametrize("accelerator", ["auto", "cpu"])
diff --git a/studio/backend/tests/test_sd_cpp_vulkan_cpu_torch.py b/studio/backend/tests/test_sd_cpp_vulkan_cpu_torch.py
index 9ebc8bab42d..f904350c0cd 100644
--- a/studio/backend/tests/test_sd_cpp_vulkan_cpu_torch.py
+++ b/studio/backend/tests/test_sd_cpp_vulkan_cpu_torch.py
@@ -353,14 +353,14 @@ def test_selection_installs_the_vulkan_build(monkeypatch):
_PINNED_RELEASE_ASSETS = [
- "sd-master-982-0518b34-ueb0e729-bin-Linux-Ubuntu-22.04-x86_64-cuda12.zip",
- "sd-master-982-0518b34-ueb0e729-bin-Linux-Ubuntu-22.04-x86_64-vulkan.zip",
- "sd-master-982-0518b34-ueb0e729-bin-Linux-Ubuntu-22.04-x86_64.zip",
- "sd-master-982-0518b34-ueb0e729-bin-Linux-Ubuntu-24.04-aarch64.zip",
- "sd-master-982-0518b34-ueb0e729-bin-win-cpu-arm64.zip",
- "sd-master-982-0518b34-ueb0e729-bin-win-cpu-x64.zip",
- "sd-master-982-0518b34-ueb0e729-bin-win-cuda12-x64.zip",
- "sd-master-982-0518b34-ueb0e729-bin-win-vulkan-x64.zip",
+ "sd-master-982-0518b34-u67cba2c-bin-Linux-Ubuntu-22.04-x86_64-cuda12.zip",
+ "sd-master-982-0518b34-u67cba2c-bin-Linux-Ubuntu-22.04-x86_64-vulkan.zip",
+ "sd-master-982-0518b34-u67cba2c-bin-Linux-Ubuntu-22.04-x86_64.zip",
+ "sd-master-982-0518b34-u67cba2c-bin-Linux-Ubuntu-24.04-aarch64.zip",
+ "sd-master-982-0518b34-u67cba2c-bin-win-cpu-arm64.zip",
+ "sd-master-982-0518b34-u67cba2c-bin-win-cpu-x64.zip",
+ "sd-master-982-0518b34-u67cba2c-bin-win-cuda12-x64.zip",
+ "sd-master-982-0518b34-u67cba2c-bin-win-vulkan-x64.zip",
]
@@ -370,9 +370,9 @@ _PINNED_RELEASE_ASSETS = [
(
"Linux",
"x86_64",
- "sd-master-982-0518b34-ueb0e729-bin-Linux-Ubuntu-22.04-x86_64-vulkan.zip",
+ "sd-master-982-0518b34-u67cba2c-bin-Linux-Ubuntu-22.04-x86_64-vulkan.zip",
),
- ("Windows", "AMD64", "sd-master-982-0518b34-ueb0e729-bin-win-vulkan-x64.zip"),
+ ("Windows", "AMD64", "sd-master-982-0518b34-u67cba2c-bin-win-vulkan-x64.zip"),
("Linux", "aarch64", None),
("Windows", "ARM64", None),
],
diff --git a/studio/install_sd_cpp_prebuilt.py b/studio/install_sd_cpp_prebuilt.py
index b1323396f71..741c15bf0fd 100644
--- a/studio/install_sd_cpp_prebuilt.py
+++ b/studio/install_sd_cpp_prebuilt.py
@@ -41,7 +41,7 @@ UPSTREAM_FALLBACK_REPO = "leejet/stable-diffusion.cpp"
# moss_uncond_skip* sample args; master-968 predates MOSS-Image (master-975, split-QKV GGUF loading in the same
# release); master-951 before it predated FLUX.3 Image Klein (master-968), and master-813 Qwen-Image-2.1 (master-883)
# and its reference-alpha fix (master-896). CUDA bundles carry no CUDA runtime.
-DEFAULT_TAG = "master-982-0518b34-ueb0e729"
+DEFAULT_TAG = "master-982-0518b34-u67cba2c"
REPO = DEFAULT_REPO