Expert-parallel size that does not divide the physical expert count fails with a bare AssertionError naming num_redundant_experts

On a 3x A100-SXM4-80GB with --ep-size 3 and a MoE model with 256 physical experts: 'AssertionError: n_physical_experts=256 must be divisible by ep_size=3. Adjust num_redundant_experts.' The engine aborts during model construction. The message's suggested remedy (num_redundant_experts) is not the constraint the user is hitting - 256 simply does not divide by 3.

Expert parallelism partitions MoE experts across ranks, so the physical expert count must factor into ep_size. The assertion was originally written to guide the redundant-expert (EPLB) logic and was left carrying a remedy that only applies when the count divides unevenly. Being raised as a bare assert rather than a ValueError, it surfaces with no indication of what the valid ep_size values are.

Choose ep_size as a divisor of n_physical_experts. The accompanying PR (#52446) converts the bare assert into a ValueError; the separate head-divisibility message improvement (#52535) is a different code path and does not cover this.

Linked to mgpu-tp-size-must-divide-attention-heads: when a user picks a TP/ep size from the physical GPU count, both checks can fire and it is easy to fix only one.