[Diffusion] Tighten runtime cache compatibility guards

Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>

docs/user_guide/diffusion/offloader/distributed_layerwise_offload.md

@@ -197,7 +197,7 @@ must enter each collective.
   online quantization methods require no-AllGather until their runtime layouts
   are validated.
 - Runtime-cache v1 rejects all quantized, non-contiguous, aliased/tied,
-  device-only, HSDP, CFG-parallel, and PP layouts.
+  device-only, HSDP, expert-parallel, CFG-parallel, and PP layouts.
 - Resident leading layers require `--dlo-no-use-allgather` and a model
   `OffloadPlan` that declares eligible `resident_dit_paths`.
 - DP concurrency requires an explicit, identical inference-step count.