Expand description
Canonical op names. Free-form &'static str is accepted everywhere,
but sticking to these keeps the report groupable.
Constantsยง
- ENGINE_
PREFILL_ BATCH - An engine-level capability rather than a per-tensor kernel: does this engine have a batched prefill path at all?
- FFN_
SWIGLU - Whole gate/up/down SwiGLU fused into one dispatch.
- GEMM_
PREFILL - A real batched GEMM over
batchactivations (prefill). - MATVEC
- One activation against a whole weight matrix (decode).
- MATVEC_
MULTI - Several independent matvecs sharing one activation (fused q/k/v).