[0z5a]

Summary

scripts/run_examples.sh always enables --features reference-cpu,experimental-tune. reference-cpu pulls in candle, and candle's gemm-f16 uses FP16 arithmetic instructions (fmla v0.8h, v1.8h, v2.h[0]). The generic aarch64 target does not enable those. On aarch64, the whole cutile-examples package therefore fails to build, and every example fails, not only the two that need the CPU reference:

error: instruction requires: fullfp16
    --> …/gemm-common-0.19.0/src/simd.rs:1994:18

cutile-examples/Cargo.toml already documents the problem (DGX Spark) and suggests RUSTFLAGS="-C target-cpu=native" for manual runs, but the runner script does not apply it.

Changes

run_examples.sh, on aarch64 when the features include reference-cpu:

• if /proc/cpuinfo lists asimdhp, append -C target-feature=+fp16 to RUSTFLAGS, keeping any value the caller set;
• otherwise, drop reference-cpu and skip the two examples that require it (async_gemm, flash_attention).

Nothing changes on x86_64.

Why +fp16 rather than -C target-cpu=native:

• It enables only the extension candle needs, so the rest of the build stays generic, and the binaries run on any aarch64 CPU with FP16.
• The fallback keeps the other examples running on CPUs without FP16.

Testing

GH200 (aarch64, Neoverse V2), CUDA 13.4 tile toolchain:

cargo build --examples --features reference-cpu,experimental-tune                         → error: instruction requires: fullfp16
RUSTFLAGS="-C target-feature=+fp16" cargo build --examples --features reference-cpu,…      → ok
cargo build --examples --features experimental-tune                                       → ok

scripts/run_examples.sh (asimdhp present):              ✓ 25 examples passed, 3 skipped
scripts/run_examples.sh with the asimdhp check forced off (fallback path):
  Skipping example: async_gemm
  Skipping example: flash_attention
  ✓ 23 examples passed, 5 skipped

The 3 skips in both runs are mxfp8, nvfp4 and nvfp4_moe, which the examples themselves skip below sm_100. Before this change, the same machine reported 28 examples failed and 0 passed.

• bash -n scripts/run_examples.sh
• examples run on GH200

Checklist

• All commits signed off (git commit -s)
• SPDX headers on new source files -- no new files

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.