pub fn wide_prefill_kernel(gpu: &dyn GpuBackend) -> (KernelHandle, u32)Expand description
2026-09-25: The head_dim > 256 prefill kernel and the BR its grid must be
built for. qwen3_attention::init takes the handle from here and
prefill_attention takes the BR from here, so the two cannot disagree; a
grid built for the wrong BR would not fail, it would cover the wrong
q-tiles.
The tensor-core attn_prefill_512tc (BR=32) when the target carries it;
otherwise, or under METRALE_ATTN_512_TC=0, the scalar attn_prefill_512
(BR=16), whose handle is zero when the target lacks it too.