On 7 April 2026, the PyTorch Foundation announced that Helion, a high-level domain-specific language for writing machine-learning kernels, has joined as a foundation-hosted project alongside DeepSpeed, Ray, and vLLM. The same week, the Hugging Face Kernels project added first-class support for Helion, letting developers package, autotune, and distribute Helion kernels through the Hugging Face Hub with a single get_kernel call.
Helion, contributed by Meta, compiles a Python-embedded DSL down to Triton and other backends. Its autotuner searches not only tile sizes but also lowering strategies — memory-access patterns, loop ordering, persistent versus looped reductions — that would require a full rewrite in Triton or CUDA. The Kernels project solves the distribution side: it standardizes source layout, enforces reproducible builds, and serves pre-built binaries for a compatibility matrix of PyTorch versions, CUDA/ROCm/XPU backends, and OSes.
Confirmed
- Helion is now a PyTorch Foundation project (Linux Foundation press release, 7 April 2026).
- Hugging Face Kernels supports Helion as a torch-noarch backend: the kernel source ships, Helion compiles and autotunes on the user's machine.
- Developers scaffold a Helion kernel with kernel-builder init --backends cuda rocm xpu, declare python-depends = ["helion"] in build.toml, and publish via kernel-builder build-and-upload.
- Pre-tuning workflow: decorate with @helion.aot_kernel(static_shapes=True), run a benchmark script across target shapes, then invoke python -m helion.autotuner.aot_runner --phase all --goal max_slowdown --threshold 1.01 --max-configs 8 to produce a decision-tree file (e.g. _helion_aot_vector_add_cuda_sm90.py) that ships alongside the kernel.
- Published examples include a vector-add kernel (sayakpaul/vector-add-helion), an attention kernel in HelionDSL/attention, and seven linear-attention variants in HelionDSL/linear-attention.
Unknown
- Independent replication of the reported speed-ups: the attention kernel beats PyTorch SDPA FLASH on 19/19 pre-tuned shapes (geomean 1.20×) and 9/10 held-out shapes (geomean 1.17×); linear-attention kernels beat FLA on all seven variants (geomean 1.41× device time, 1.33× end-to-end on pre-tuned shapes). These figures come from the project's own benchmarks on H100 and B200.
- Rollout timing for Windows support and for backends beyond CUDA/ROCm/XPU (TileIR is mentioned as "coming soon").
- Whether the trust_remote_code=True requirement for get_kernel will be relaxed once Helion kernels are in the official compatibility matrix.
- Exact cold-start latency reduction from shipping decision trees versus on-device autotuning (the blog notes autotuning "typically takes around 10 minutes" on first run).
Our take
Helion's entry into the PyTorch Foundation gives it a vendor-neutral governance path at the same moment Hugging Face Kernels solves the "dependency hell" of consuming custom kernels. The combination means a kernel author writes once in a PyTorch-like DSL, the autotuner explores thousands of Triton configurations, and the consumer pulls a ready-to-run artifact with one line of code. The missing piece for production adoption is independent verification of the speed claims across diverse model architectures and a clear SLA for the Hub's binary availability.
