[ascend]adapt for s1-pro dp*tp+ep

[yao-fengchen]

No description provided.

[Copilot]

Pull request overview

This PR adapts the dlinfer (Ascend) backend to support dp*tp+ep parallelism configurations. It makes two changes:

1. Adds FopeRotaryEmbeddingImpl (Fractional Rotary Positional Embedding) support to the dlinfer rotary embedding builder, reusing the default backend's implementation.
2. Enables group-based router support (n_groups) in DlinferSoftmaxTopKImpl by removing the prior NotImplementedError and instead forwarding n_groups to moe_metadata during the forward pass.

Changes:

• Add FopeRotaryEmbeddingImpl support for RopeType.Fope in the dlinfer rotary embedding builder.
• Remove the NotImplementedError for group routing in DlinferSoftmaxTopKImpl, storing n_groups and propagating it to moe_metadata during inference.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
lmdeploy/pytorch/backends/dlinfer/rotary_embedding.py	Imports FopeRotaryEmbeddingImpl from the default backend and adds a RopeType.Fope case to DlinferRotaryEmbeddingBuilder.build, mirroring the default builder.
lmdeploy/pytorch/backends/dlinfer/moe.py	Replaces the NotImplementedError for grouped routing with a stored n_groups field and assigns it to moe_metadata.router_n_groups in forward.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.