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: Windows torch 2.12 kernels from our prebuilt wheels (#13808)

Daniel Handac778ce2a9c
diff --git a/studio/backend/tests/test_kernel_install.py b/studio/backend/tests/test_kernel_install.py
index c3ab1ab074e..9f6b7c02135 100644
--- a/studio/backend/tests/test_kernel_install.py
+++ b/studio/backend/tests/test_kernel_install.py
@@ -99,7 +99,7 @@ def test_causal_conv1d_reuses_torch_2_10_wheel_on_2_11():
[
None,
_env("2.11.0+cpu", ""),
- _env("2.12.0+cu130", "13.0", platform_tag = "win_amd64"),
+ _env("2.9.0+cu130", "13.0", platform_tag = "win_amd64"),
],
)
def test_causal_conv1d_has_no_wheel_off_linux_cuda(env):
@@ -346,7 +346,12 @@ def test_console_entry_drops_an_unwritable_ssl_keylog_file(tmp_path):
"https://github.com/unslothai/unsloth/releases/download/prebuilt-wheels-cu13/"
"flash_attn-2.8.4+cu13torch2.10cxx11abiTRUE-cp313-cp313-win_amd64.whl",
),
- (_env("2.12.0+cu130", "13.0", platform_tag = "win_amd64"), None),
+ (
+ _env("2.12.0+cu130", "13.0", platform_tag = "win_amd64"),
+ "https://github.com/unslothai/unsloth/releases/download/prebuilt-wheels-cu13/"
+ "flash_attn-2.8.4+cu13torch2.12cxx11abiTRUE-cp313-cp313-win_amd64.whl",
+ ),
+ (_env("2.9.0+cu130", "13.0", platform_tag = "win_amd64"), None),
(_env("2.11.0+cpu", ""), None),
],
)
diff --git a/studio/backend/tests/test_wheel_utils_windows_kernels.py b/studio/backend/tests/test_wheel_utils_windows_kernels.py
index 60fea21e2e1..648f7cd38a8 100644
--- a/studio/backend/tests/test_wheel_utils_windows_kernels.py
+++ b/studio/backend/tests/test_wheel_utils_windows_kernels.py
@@ -47,9 +47,9 @@ def test_win_amd64_resolves_every_kernel_to_our_release(torch_mm):
assert kernel_install.MAMBA_SSM.wheel_url(e) == f"{OURS}mamba_ssm-2.3.2.post1{suffix}"
[email protected]("torch_mm", ["2.10", "2.11"])
[email protected]("torch_mm", ["2.10", "2.11", "2.12"])
@pytest.mark.parametrize("python_tag", ["cp311", "cp312", "cp313"])
-def test_win_amd64_torch_210_and_211_resolve_to_our_release(torch_mm, python_tag):
+def test_win_amd64_torch_210_to_212_resolve_to_our_release(torch_mm, python_tag):
# Existing installs keep their torch minor; upstream's torch2.10 wheels have no Windows asset.
e = env(torch_mm = torch_mm, python_tag = python_tag, abi_tag = python_tag)
suffix = f"+cu13torch{torch_mm}cxx11abiTRUE-{python_tag}-{python_tag}-win_amd64.whl"
@@ -68,7 +68,7 @@ def test_win_amd64_torch_210_and_211_resolve_to_our_release(torch_mm, python_tag
{"python_tag": "cp312", "abi_tag": "cp312"}, # 2.13 / 2.14 publish cp313 only
{"abi_tag": "cp313t"}, # a free-threaded build cannot load the GIL wheel
{"cxx11abi": "FALSE"}, # not what the Windows cu130 torch reports
- {"torch_mm": "2.12"}, # no wheel, and no upstream fallback
+ {"torch_mm": "2.12", "python_tag": "cp310", "abi_tag": "cp310"},
{"torch_mm": "2.11", "python_tag": "cp314", "abi_tag": "cp314"},
{"torch_mm": "2.11", "abi_tag": "cp313t"},
{"torch_mm": "2.10", "python_tag": "cp310", "abi_tag": "cp310"},
@@ -290,7 +290,7 @@ def test_cli_resolves_windows_kernels():
"-cp313-cp313-win_amd64.whl"
)
assert kernel_install.resolve_wheel_url("mamba_ssm", e).endswith("-cp313-cp313-win_amd64.whl")
- assert kernel_install.resolve_wheel_url("flash_attn", env(torch_mm = "2.12")) is None
+ assert kernel_install.resolve_wheel_url("flash_attn", env(torch_mm = "2.9")) is None
assert (
kernel_install.resolve_wheel_url("flash_attn", env(platform_tag = "macosx_14_0_arm64"))
is None
diff --git a/studio/backend/utils/wheel_utils.py b/studio/backend/utils/wheel_utils.py
index a87fdd56193..a1988d2b300 100644
--- a/studio/backend/utils/wheel_utils.py
+++ b/studio/backend/utils/wheel_utils.py
@@ -139,10 +139,11 @@ _UNSLOTH_PREBUILT_PLATFORMS: dict[str, tuple[str, frozenset[str] | None]] = {
"win_arm64": ("TRUE", frozenset({"cp311", "cp312", "cp313"})),
}
-# (platform, torch minor) cells outside _UNSLOTH_PREBUILT_TORCH_MM, with their interpreters. Upstream's torch2.10 wheels (and their 2.11 reuse above) have no Windows asset, so win_amd64 torch 2.10 and 2.11 are ours; an existing install keeps its minor (a fresh one takes 2.13 on cp313), hence three interpreters.
+# (platform, torch minor) cells outside _UNSLOTH_PREBUILT_TORCH_MM, with their interpreters. Upstream's torch2.10 wheels (and their 2.11 / 2.12 reuse above) have no Windows asset, so win_amd64 torch 2.10, 2.11 and 2.12 are ours; an existing install keeps its minor (a fresh one takes 2.13 on cp313), hence three interpreters.
_UNSLOTH_PREBUILT_EXTRA_CELLS: dict[tuple[str, str], frozenset[str]] = {
("win_amd64", "2.10"): frozenset({"cp311", "cp312", "cp313"}),
("win_amd64", "2.11"): frozenset({"cp311", "cp312", "cp313"}),
+ ("win_amd64", "2.12"): frozenset({"cp311", "cp312", "cp313"}),
}
# Windows on ARM has one CUDA torch, NVIDIA's 2.14.0+cu134 (CUDA 13.4, the first with a Windows ARM64 nvcc), and NVIDIA publishes flash-attn for it, so only these two are ours there.
diff --git a/tests/python/test_prebuilt_cuda_wheels.py b/tests/python/test_prebuilt_cuda_wheels.py
index 15cbb7f974a..80b28588a60 100644
--- a/tests/python/test_prebuilt_cuda_wheels.py
+++ b/tests/python/test_prebuilt_cuda_wheels.py
@@ -117,7 +117,7 @@ class TestWorkflowAndResolverAgree:
)
assert url == f"{OUR_BASE}{built}"
checked += 1
- assert checked == 18
+ assert checked == 27
extra = {
(platform, prebuilt_wheels.torch_minor(v))
for platform in prebuilt_wheels.PLATFORMS