mirror of
https://github.com/hanzoai/ml.git
synced 2026-11-19 23:50:51 +00:00
The ROCm test lane could never compile: `rocm` maps to hanzo-cubecl-hip, but two tests imported `cubecl::hip`, which that feature never enables. `cargo test --features rocm` failed at E0432 before running anything, so the gate that was supposed to prove ROCm works was itself dead. Both sites now import the runtime the feature actually provides, matching the third site that was already correct. hello_kernel restated rms_norm's body, so the example proved a copy rather than the kernel that ships. It now imports hanzo_kernel::norm::rms_norm_run and runs it over a generic Runtime, with the backend chosen by cfg: one source, one launch contract, and Target::of reports the architecture it actually lowered to instead of the one we hoped for. Inputs are deterministic so the checksum is comparable across boxes -- the same number on AMD, NVIDIA and Apple is the falsifiable claim. rms_norm_ref stays the oracle and is pure host code, so the comparison holds on backends where the cpu runtime cannot build.