pub fn wide_prefill_kernel(gpu: &dyn GpuBackend) -> (KernelHandle, u32)Expand description
The HDIM>256 prefill kernel: its module/entry name and the BR its grid must
be built for. ONE reader, because the name is chosen in qwen3_attention::init
and the grid here, and a mismatch is silent.
Default is the tensor-core instantiation (BR=32). ATLAS_ATTN_512_TC=0
selects the scalar reference (BR=16) — kept reachable because it is the
oracle the TC path was validated against (cosine 0.999998, 64.7x faster on
S=1024/4q/2kv/causal).