Referenced script:

import os
from pathlib import Path

import torch
import torch.nn.functional as F
import torch.utils.benchmark as benchmark
from kernels import get_local_kernel

REPO = Path("/home/jiqing/qrdgvgjak43jp7w90ckydgz3righnyi2-activation-torch-ext")

act = get_local_kernel(REPO, backend="cpu")
print(f"loaded: {act.__name__}  silu_and_mul={act.silu_and_mul}\n")


def kernel(x):
    d = x.shape[-1] // 2
    out = torch.empty(x.shape[:-1] + (d,), dtype=x.dtype)
    act.silu_and_mul(out, x.contiguous())
    return out


def ref(x):
    d = x.shape[-1] // 2
    return F.silu(x[..., :d]) * x[..., d:]


def run_correctness():
    print("=== Correctness ===")
    torch.manual_seed(0)
    dtypes = {torch.bfloat16: 2e-2, torch.float16: 1e-3, torch.float32: 1e-5}
    shapes = [(7, 512), (83, 512), (2048, 512), (2048, 13824), (1, 4096)]
    ok_all = True
    for dt, tol in dtypes.items():
        for ntok, d in shapes:
            x = torch.randn(ntok, 2 * d, dtype=dt)
            got, exp = kernel(x), ref(x)
            diff = (got.float() - exp.float()).abs().max().item()
            ok = torch.allclose(got, exp, atol=tol, rtol=tol)
            ok_all &= ok
            print(f"  {str(dt):17s} ntok={ntok:5d} d={d:6d}  "
                  f"max_diff={diff:.3e}  {'OK' if ok else 'FAIL'}")
    print("ALL CORRECT\n" if ok_all else "SOME FAILED\n")
    return ok_all


def run_perf():
    print("=== Performance (bf16, torch.utils.benchmark) ===")
    nt = torch.get_num_threads()
    print(f"  threads={nt}  OMP_NUM_THREADS={os.environ.get('OMP_NUM_THREADS')}")
    torch.manual_seed(0)
    for ntok, d in [(2048, 512), (2048, 4096), (2048, 13824), (8192, 4096)]:
        x = torch.randn(ntok, 2 * d, dtype=torch.bfloat16)
        out = torch.empty(ntok, d, dtype=torch.bfloat16)

        tk = benchmark.Timer(
            stmt="act.silu_and_mul(out, x)",
            globals={"act": act, "x": x, "out": out}, num_threads=nt,
        ).blocked_autorange(min_run_time=2.0)
        tr = benchmark.Timer(
            stmt="out.copy_(ref(x))",
            globals={"ref": ref, "x": x, "out": out}, num_threads=nt,
        ).blocked_autorange(min_run_time=2.0)

        k, r = tk.median * 1e6, tr.median * 1e6
        print(f"  ntok={ntok:5d} d={d:6d}  kernel={k:8.1f}us  "
              f"torch={r:8.1f}us  speedup={r / k:5.2f}x")


if __name__ == "__main__":
    ok = run_correctness()
    run_perf()
    raise SystemExit(0 if ok else 1)

This kernel was written following the
cpu-kernels skill
in kernel-builder, which guides the runtime-dispatch + AVX512 + ATen-fallback design.

Sorry, this slipped through the cracks.

/kernel-bot security-and-build activation

Hi @danieldk . Is it okay to merge the PR? Thanks!

/kernel-bot security-and-build activation

Build request processed.

Command: /kernel-bot security-and-build activation
Mode: security audit + build
Target branch: pr-942
PR head SHA: 6d6a63001968edd59c4332560ddc6d72acbb1aa4
Workflows: build.yaml, build-mac.yaml, build-windows.yaml

Dispatched (3):

• activation (build-mac.yaml): https://github.com/huggingface/kernels-community/actions/runs/28848521084
• activation (build-windows.yaml): https://github.com/huggingface/kernels-community/actions/runs/28848521961
• activation (build.yaml): https://github.com/huggingface/kernels-community/actions/runs/28848522939

Dispatched (1):

• security (security-audit.yml): https://github.com/huggingface/kernels-community/actions/runs/28848523887

🔒 Security Audit — findings detected

An automated, AI-generated security review of this PR flagged the following. These may include false positives — review each item before acting.

I've reviewed the diff thoroughly. This is a CPU backend addition for the silu_and_mul activation kernel, following the same established pattern already present in the repo (e.g. rmsnorm/rmsnorm_cpu/ uses the identical CPU_KERNEL define, AVX512 impl, and CPU feature detection structure).

Assessment:

• Memory safety: The AVX512 core reads in_data + token_idx*(2*d) + offset (and b_ptr = a_ptr + d), all bounded by elements_to_process ≤ d - offset per token, and writes only to out_data[i]. Access stays within the declared [tokens, 2*d] input / [tokens, d] output shapes. The vectorized loop guard j <= elements_to_process - kVb uses signed int64_t, so no unsigned wraparound when the remainder is small.
• Math matches the name: Output is silu(input[...,:d]) * input[...,d:] — no writes to a second buffer, no scanning of unrelated memory.
• /proc/cpuinfo read is a standard, benign CPU-feature probe (AVX512_BF16 detection); it changes only which compute path is taken, not control flow that leaks data. No network, subprocess, eval/exec, dlopen, or filesystem writes anywhere.
• Op registration is correct: defined and impl'd under TORCH_EXTENSION_NAME for torch::kCPU, no aten:: override or cross-kernel namespace hijack.
• build.toml only adds a cpu backend depending on torch, with standard AVX512 compiler flags; no repo-id change, no new URLs, no fetch-at-build, no unfamiliar deps.
• No obfuscation, homoglyphs, encoded blobs, or checked-in binaries.

NO_FINDINGS

Audited head SHA 6d6a63001968edd59c4332560ddc6d72acbb1aa4 · audit run

/kernel-bot build activation

Build request processed.

Command: /kernel-bot build activation
Mode: build only
Target branch: pr-942
PR head SHA: 7176bf18da9097eb9313946b90c4b42ab7f8e710
Workflows: build.yaml, build-mac.yaml, build-windows.yaml

Dispatched (3):

• activation (build-mac.yaml): https://github.com/huggingface/kernels-community/actions/runs/28850351544
• activation (build-windows.yaml): https://github.com/huggingface/kernels-community/actions/runs/28850352513
• activation (build.yaml): https://github.com/huggingface/kernels-community/actions/runs/28850353668

Thanks a lot!

/kernel-bot merge-and-upload activation

Build request processed.

Command: /kernel-bot merge-and-upload activation
Mode: merge, build and upload
Target branch: ``
PR head SHA: 7176bf18da9097eb9313946b90c4b42ab7f8e710
Workflows: `build.yaml, build-mac.yaml, build-windows.yaml`

Merge result: PR is already merged. Continuing with build/upload.

Dispatched (3):

• activation (build-mac.yaml): https://github.com/huggingface/kernels-community/actions/runs/28852397427
• activation (build-windows.yaml): https://github.com/huggingface/kernels-community/actions/runs/28852398581
• activation (build.yaml): https://github.com/huggingface/kernels-community/actions/runs/28852399731