For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).

Mojo function

router_gate_use_mixed_gemv

def router_gate_use_mixed_gemv(m: Int) -> Bool

Returns whether runtime m should take the fused mixed router GEMV.

Only tiny-M (decode) rows use the single-launch mixed GEMV. m == 0 (graph-capture warmup) takes neither path, and large m (prefill) must fall back to cast + fp32 matmul to preserve the baseline's performance.

Args:

m (Int): The runtime row count of the router-gate activation.

Returns:

Bool: True if the fused mixed GEMV should be launched for this m.