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.

Build Windows x86_64 prebuilt CUDA wheels for torch 2.12 (#13807)

Daniel Hand6fc96b9cda2
diff --git a/.github/scripts/prebuilt_wheels.py b/.github/scripts/prebuilt_wheels.py
index d89436c4c41..89f61432a8a 100644
--- a/.github/scripts/prebuilt_wheels.py
+++ b/.github/scripts/prebuilt_wheels.py
@@ -80,9 +80,9 @@ PLATFORMS = {
"abi": "TRUE",
"label": " / win_amd64",
"pytorch_nvcc": "C:/ccache/ccache.exe C:/cuda/bin/nvcc.exe",
- # On top of TORCH_VERSIONS: upstream's torch2.10 wheels (which Linux also reuses for 2.11)
- # have no Windows build.
- "extra_torches": ("2.10.0", "2.11.0"),
+ # On top of TORCH_VERSIONS: upstream's torch2.10 wheels (which Linux also reuses for 2.11
+ # and 2.12) have no Windows build.
+ "extra_torches": ("2.10.0", "2.11.0", "2.12.0"),
},
# Upstream's aarch64 wheels stop at torch 2.10. ubuntu-22.04-arm keeps the glibc 2.35 floor
# the x86_64 build has, and the build asserts it.
@@ -130,11 +130,12 @@ def _platform(platform: str, key: str):
return PLATFORMS[platform].get(key, _PLATFORM_DEFAULTS.get(key))
-# torch 2.10 and 2.11 pin triton==3.6.0, 2.13 triton==3.7.1 and 2.14 triton~=3.8.0 (Linux-only
-# markers); exact pins keep the Windows smoke test reproducible.
+# torch 2.10 and 2.11 pin triton==3.6.0, 2.12.0 triton==3.7.0, 2.13 triton==3.7.1 and 2.14
+# triton~=3.8.0 (Linux-only markers); exact pins keep the Windows smoke test reproducible.
TRITON_WINDOWS = {
"2.10": "3.6.0.post26",
"2.11": "3.6.0.post26",
+ "2.12": "3.7.0.post26",
"2.13": "3.7.1.post27",
"2.14": "3.8.0.post29",
}
@@ -227,8 +228,8 @@ def _platform_torches(platform: str) -> tuple[str, ...]:
)
-# Accepted from a dispatch; each platform builds only its own subset (2.10.0 and 2.11.0 are
-# win_amd64 only).
+# Accepted from a dispatch; each platform builds only its own subset (2.10.0, 2.11.0 and 2.12.0
+# are win_amd64 only).
ALLOWED_TORCH = tuple(
sorted(
{torch for platform in PLATFORMS for torch in _platform_torches(platform)},
@@ -402,7 +403,7 @@ def build_matrix(
if not include:
raise SystemExit(
"no cells: none of the chosen platforms builds that package / torch combination "
- "(torch 2.10.0 and 2.11.0 are win_amd64 only; win_arm64 builds causal-conv1d and "
+ "(torch 2.10.0 to 2.12.0 are win_amd64 only; win_arm64 builds causal-conv1d and "
"mamba-ssm for torch 2.14.0 only)."
)
return include
diff --git a/.github/workflows/prebuilt-cuda-wheels.yml b/.github/workflows/prebuilt-cuda-wheels.yml
index 7a90cbe4afb..4b8eda42235 100644
--- a/.github/workflows/prebuilt-cuda-wheels.yml
+++ b/.github/workflows/prebuilt-cuda-wheels.yml
@@ -29,7 +29,7 @@ on:
workflow_dispatch:
inputs:
torch_versions:
- description: 'torch versions to build for, comma separated (2.13.0, 2.14.0; 2.10.0 and 2.11.0 build on win_amd64 only)'
+ description: 'torch versions to build for, comma separated (2.13.0, 2.14.0; 2.10.0, 2.11.0 and 2.12.0 build on win_amd64 only)'
type: string
default: '2.13.0,2.14.0'
python_versions:
diff --git a/tests/python/test_prebuilt_cuda_wheels.py b/tests/python/test_prebuilt_cuda_wheels.py
index fa147d06dc5..b52321e4261 100644
--- a/tests/python/test_prebuilt_cuda_wheels.py
+++ b/tests/python/test_prebuilt_cuda_wheels.py
@@ -121,7 +121,7 @@ class TestWorkflowAndResolverAgree:
assert wheel_utils._UNSLOTH_PREBUILT_VERSIONS[spec["dist"]] == spec["version"], package
def test_every_extra_cell_the_workflow_builds_is_what_the_resolver_asks_for(self):
- """win_amd64 torch 2.10 / 2.11: outside TORCH_VERSIONS, so the class above does not see them."""
+ """win_amd64 torch 2.10 to 2.12: outside TORCH_VERSIONS, so the class above does not see them."""
checked = 0
for (platform, torch_mm), pythons in wheel_utils._UNSLOTH_PREBUILT_EXTRA_CELLS.items():
built_torches = prebuilt_wheels._platform_torches(platform)
@@ -341,7 +341,7 @@ class TestMatrix:
[
{"packages": "flash-attn; rm -rf /"},
{"packages": "torchvision"},
- {"torches": "2.12.0"},
+ {"torches": "2.9.0"},
{"torches": "$(whoami)"},
{"pythons": "3.9"},
{"pythons": "3.13 && curl evil"},
diff --git a/tests/python/test_prebuilt_cuda_wheels_windows.py b/tests/python/test_prebuilt_cuda_wheels_windows.py
index a3a4355012b..4f4008ba754 100644
--- a/tests/python/test_prebuilt_cuda_wheels_windows.py
+++ b/tests/python/test_prebuilt_cuda_wheels_windows.py
@@ -142,7 +142,7 @@ class TestMatrix:
assert cell["smoke_extra"] == cell["smoke_exclude"] == ""
def test_every_torch_minor_has_a_triton_windows_pin(self):
- """torch 2.10 and 2.11 pin triton 3.6.0, 2.13 3.7.1 and 2.14 3.8.x; triton-windows follows."""
+ """torch 2.10 / 2.11 pin triton 3.6.0, 2.12 3.7.0, 2.13 3.7.1, 2.14 3.8.x; triton-windows follows."""
minors = {
prebuilt_wheels.torch_minor(v)
for platform, target in prebuilt_wheels.PLATFORMS.items()
@@ -152,10 +152,11 @@ class TestMatrix:
assert set(prebuilt_wheels.TRITON_WINDOWS) == minors
assert prebuilt_wheels.TRITON_WINDOWS["2.10"].startswith("3.6.0.post")
assert prebuilt_wheels.TRITON_WINDOWS["2.11"].startswith("3.6.0.post")
+ assert prebuilt_wheels.TRITON_WINDOWS["2.12"].startswith("3.7.0.post")
assert prebuilt_wheels.TRITON_WINDOWS["2.13"].startswith("3.7.1.post")
assert prebuilt_wheels.TRITON_WINDOWS["2.14"].startswith("3.8.0.post")
- @pytest.mark.parametrize("torch_version", ["2.10.0", "2.11.0"])
+ @pytest.mark.parametrize("torch_version", ["2.10.0", "2.11.0", "2.12.0"])
def test_win_amd64_only_torches_are_opt_in(self, torch_version):
torch_mm = prebuilt_wheels.torch_minor(torch_version)
cells = prebuilt_wheels.build_matrix(
@@ -167,7 +168,8 @@ class TestMatrix:
for cell in cells:
assert f"+cu13torch{torch_mm}cxx11abiTRUE-" in cell["wheel_name"]
if cell["package"] == "mamba-ssm":
- assert cell["smoke_extra"] == "triton-windows==3.6.0.post26"
+ triton = prebuilt_wheels.TRITON_WINDOWS[torch_mm]
+ assert cell["smoke_extra"] == f"triton-windows=={triton}"
for platform in ("linux_x86_64", "linux_aarch64", "win_arm64"):
with pytest.raises(SystemExit):
prebuilt_wheels.build_matrix(torches = torch_version, platforms = platform)
@@ -185,6 +187,7 @@ class TestMatrix:
assert prebuilt_wheels._platform_torches("win_amd64") == (
"2.10.0",
"2.11.0",
+ "2.12.0",
"2.13.0",
"2.14.0",
"2.15.0",
@@ -355,7 +358,7 @@ class TestNotes:
for index, (package, torch, python) in enumerate(
[
(pkg, t, py)
- for t in ("2.10.0", "2.11.0")
+ for t in ("2.10.0", "2.11.0", "2.12.0")
for py in ("3.11", "3.12", "3.13")
for pkg in prebuilt_wheels.SPECS
]
@@ -364,8 +367,8 @@ class TestNotes:
]
notes = prebuilt_wheels.render_notes(entries, tag = "t", repo = "r")
assert (
- "built for PyTorch 2.10 and 2.11 on Python 3.11, 3.12 and 3.13; 2.13 and 2.14 on "
- "Python 3.13." in notes
+ "built for PyTorch 2.10, 2.11 and 2.12 on Python 3.11, 3.12 and 3.13; 2.13 and 2.14 "
+ "on Python 3.13." in notes
)
def test_platforms_with_different_coverage_are_not_merged(self):