Merge branch 'main' into xpu-skill

Author: sayakpaul

.github/workflows/build_kernel_macos.yaml

@@ -28,7 +28,7 @@ jobs:
       # For now we only test that there are no regressions in building macOS
       # kernels. Also run tests once we have a macOS runner.
       - name: Build relu kernel
-        run: ( cd examples/kernels/relu && nix build .\#redistributable.torch210-metal-aarch64-darwin -L )
+        run: ( cd examples/kernels/relu && nix build .\#redistributable.torch211-metal-aarch64-darwin -L )
 
       - name: Build relu metal cpp kernel
-        run: ( cd examples/kernels/relu-metal-cpp && nix build .\#redistributable.torch210-metal-aarch64-darwin -L )
+        run: ( cd examples/kernels/relu-metal-cpp && nix build .\#redistributable.torch211-metal-aarch64-darwin -L )

.github/workflows/build_kernel_rocm.yaml

@@ -34,7 +34,7 @@ jobs:
       # For now we only test that there are no regressions in building ROCm
       # kernels. Also run tests once we have a ROCm runner.
       - name: Build relu kernel
-        run: ( cd examples/kernels/relu && nix build .\#redistributable.torch210-cxx11-rocm70-x86_64-linux -L )
+        run: ( cd examples/kernels/relu && nix build .\#redistributable.torch211-cxx11-rocm71-x86_64-linux -L )
 
       - name: Build relu kernel (compiler flags)
-        run: ( cd examples/kernels/relu-compiler-flags && nix build .\#redistributable.torch210-cxx11-rocm70-x86_64-linux )
+        run: ( cd examples/kernels/relu-compiler-flags && nix build .\#redistributable.torch211-cxx11-rocm71-x86_64-linux )

.github/workflows/build_kernel_xpu.yaml

@@ -34,13 +34,13 @@ jobs:
       # For now we only test that there are no regressions in building XPU
       # kernels. Also run tests once we have a XPU runner.
       - name: Build relu kernel
-        run: ( cd examples/kernels/relu && nix build .\#redistributable.torch210-cxx11-xpu20253-x86_64-linux -L )
+        run: ( cd examples/kernels/relu && nix build .\#redistributable.torch211-cxx11-xpu20253-x86_64-linux -L )
 
       - name: Build relu tvm-ffi kernel
         run: ( cd examples/kernels/relu-tvm-ffi && nix build .\#redistributable.tvm-ffi01-xpu20253-x86_64-linux -L )
 
       - name: Build relu kernel (compiler flags)
-        run: ( cd examples/kernels/relu-compiler-flags && nix build .\#redistributable.torch210-cxx11-xpu20253-x86_64-linux )
+        run: ( cd examples/kernels/relu-compiler-flags && nix build .\#redistributable.torch211-cxx11-xpu20253-x86_64-linux )
 
       - name: Build cutlass-gemm kernel
-        run: ( cd examples/kernels/cutlass-gemm && nix build .\#redistributable.torch210-cxx11-xpu20253-x86_64-linux -L )
+        run: ( cd examples/kernels/cutlass-gemm && nix build .\#redistributable.torch211-cxx11-xpu20253-x86_64-linux -L )

.github/workflows/test_kernels.yaml

@@ -26,8 +26,8 @@ jobs:
     strategy:
       max-parallel: 4
       matrix:
-        python-version: ["3.10", "3.12"]
-        torch-version: ["2.10.0", "2.11.0"]
+        python-version: ["3.10", "3.14"]
+        torch-version: ["2.11.0", "2.12.0"]
 
     env:
       UV_PYTHON_PREFERENCE: only-managed

docs/source/builder/build-variants.md

@@ -7,27 +7,21 @@ available. This list will be updated as new PyTorch versions are released.
 
 ## CPU aarch64-darwin
 
-- `torch210-cpu-aarch64-darwin`
 - `torch211-cpu-aarch64-darwin`
 - `torch212-cpu-aarch64-darwin`
 
 ## Metal aarch64-darwin
 
-- `torch210-metal-aarch64-darwin`
 - `torch211-metal-aarch64-darwin`
 - `torch212-metal-aarch64-darwin`
 
 ## CPU aarch64-linux
 
-- `torch210-cxx11-cpu-aarch64-linux`
 - `torch211-cxx11-cpu-aarch64-linux`
 - `torch212-cxx11-cpu-aarch64-linux`
 
 ## CUDA aarch64-linux
 
-- `torch210-cxx11-cu126-aarch64-linux`
-- `torch210-cxx11-cu128-aarch64-linux`
-- `torch210-cxx11-cu130-aarch64-linux`
 - `torch211-cxx11-cu126-aarch64-linux`
 - `torch211-cxx11-cu128-aarch64-linux`
 - `torch211-cxx11-cu130-aarch64-linux`
@@ -37,15 +31,11 @@ available. This list will be updated as new PyTorch versions are released.
 
 ## CPU x86_64-linux
 
-- `torch210-cxx11-cpu-x86_64-linux`
 - `torch211-cxx11-cpu-x86_64-linux`
 - `torch212-cxx11-cpu-x86_64-linux`
 
 ## CUDA x86_64-linux
 
-- `torch210-cxx11-cu126-x86_64-linux`
-- `torch210-cxx11-cu128-x86_64-linux`
-- `torch210-cxx11-cu130-x86_64-linux`
 - `torch211-cxx11-cu126-x86_64-linux`
 - `torch211-cxx11-cu128-x86_64-linux`
 - `torch211-cxx11-cu130-x86_64-linux`
@@ -55,16 +45,13 @@ available. This list will be updated as new PyTorch versions are released.
 
 ## ROCm x86_64-linux
 
-- `torch210-cxx11-rocm70-x86_64-linux`
-- `torch210-cxx11-rocm71-x86_64-linux`
 - `torch211-cxx11-rocm71-x86_64-linux`
 - `torch211-cxx11-rocm72-x86_64-linux`
 - `torch212-cxx11-rocm71-x86_64-linux`
 - `torch212-cxx11-rocm72-x86_64-linux`
 
 ## XPU x86_64-linux
 
-- `torch210-cxx11-xpu20253-x86_64-linux`
 - `torch211-cxx11-xpu20253-x86_64-linux`
 - `torch212-cxx11-xpu20253-x86_64-linux`
 

docs/source/builder/writing-kernels.md

@@ -351,6 +351,47 @@ def mykernel(x: torch.Tensor, out: Optional[torch.Tensor] = None) -> torch.Tenso
     return out
 ```
 
+## Registering Torch operators
+
+You may want to register Torch ops from your kernel's Python code or
+register fake ops for `torch.compile` support. It is important to register
+such ops in the namespace that kernel-builder makes for your kernel
+build. This is required for compliant kernels to ensure that multiple
+versions of the same kernel can be loaded at the same time without
+namespace conflicts.
+
+You can use the `add_op_namespace_prefix` to prefix an op name with the
+correct prefix. So for instance, replace
+
+```python
+@torch.library.register_fake("relu::relu_fwd")
+def relu_fwd_fake(input: torch.Tensor) -> torch.Tensor:
+    return torch.empty_like(input)
+```
+
+by
+
+```python
+from ._ops import add_op_namespace_prefix
+
+@torch.library.register_fake(add_op_namespace_prefix("relu_fwd"))
+def relu_fwd_fake(input: torch.Tensor) -> torch.Tensor:
+    return torch.empty_like(input)
+```
+
+As mentioned in the above, the `_ops` module is generated by kernel-builder.
+
+kernel-builder uses a hook to reject incorrect usage of Torch op registration
+functions. However, it can only catch direct use of certain `torch.library`
+decorators. For instance, the hook would not reject the following decorator,
+so it should be seen as a last-resort check if human review failed:
+
+```python
+@some_indirection_for_register_fake("relu::relu_fwd")
+def relu_fwd_fake(input: torch.Tensor) -> torch.Tensor:
+    return torch.empty_like(input)
+```
+
 ## Kernel tests
 
 Kernel tests are stored in the `tests` directory. Since running all

docs/source/installation.md

@@ -17,3 +17,9 @@ or if you want the latest version from the `main` branch:
 ```bash
 pip install "kernels[benchmark] @ git+https://github.com/huggingface/kernels#subdirectory=kernels"
 ```
+
+> [!IMPORTANT] 
+> We strongly recommend not using a free-threaded Python build yet.
+These builds are not only experimental, but do not support the stable ABI
+on Python versions before 3.15. Kernels are compiled with the stable ABI
+to support a wide range of Python versions.

examples/kernels/flake.nix

@@ -84,7 +84,7 @@
           drv = sys: out: out.packages.${sys}.default;
           torchVersions = _defaultVersions: [
             {
-              torchVersion = "2.10";
+              torchVersion = "2.11";
               cudaVersion = "12.8";
               systems = [
                 "x86_64-linux"

examples/kernels/relu-specific-torch/flake.nix

@@ -15,7 +15,7 @@
       path = ./.;
       torchVersions = defaultVersions: [
         {
-          torchVersion = "2.10";
+          torchVersion = "2.11";
           cudaVersion = "12.8";
           systems = [
             "x86_64-linux"

examples/kernels/relu-torch-bounds/build.toml

@@ -11,8 +11,8 @@ backends = [
 repo-id = "kernels-test/relu-torch-bounds"
 
 [torch]
-minver = "2.10"
-maxver = "2.10"
+minver = "2.11"
+maxver = "2.11"
 src = [
     "torch-ext/torch_binding.cpp",
     "torch-ext/torch_binding.h",