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

Python function

replicate_experts

replicate_experts()

max.pipelines.lib.eplb_rebalance.replicate_experts(weight, num_phy)

source

Replicate num_log logical experts to num_phy physical slots.

Greedy: at each step we duplicate whichever logical expert currently has the highest per-replica load.

Returns:

[X, num_phy] — logical expert id of each physical slot. rank: [X, num_phy] — replica rank within its logical expert. logcnt: [X, num_log] — number of replicas per logical expert.

Return type:

phy2log

Parameters:

• weight (ndarray[tuple[Any, ...], dtype[float64]])
• num_phy (int)