Remove state init

[grimoire]

• fill conv state in model when forward
• update gate to ignore init state of gdr
• remove init cache, it take too much times
• l2 norm before repeat interleave

[Copilot]

Pull request overview

This PR removes explicit state-cache initialization for the GatedDelta/SSM path by making the model handle “init state” behavior during forward, and optimizes the kv head replication path by applying Q/K L2-normalization before repeat_interleave to reduce overhead.

Changes:

• Add init-state metadata (is_init, is_init_token) to GatedDeltaMeta, zero conv initial states on init, and mask GDR gate for init tokens.
• Move kv_ratio replication logic into GatedDelta (and add a helper that normalizes before replication).
• Remove StateCacheEngine.init_caches and its call site during model forward.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

File	Description
lmdeploy/pytorch/nn/gated_delta.py	Adds init-token handling and moves kv replication + (optional) Q/K L2-norm before replication into the GatedDelta wrapper.
lmdeploy/pytorch/models/qwen3_5.py	Stops repeating Q/K in the model and passes kv_ratio into GatedDelta.
lmdeploy/pytorch/engine/model_agent/agent.py	Removes the state cache initialization call during forward.
lmdeploy/pytorch/engine/cache_engine.py	Removes StateCacheEngine.init_caches implementation.

---

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

[RunningLeon]

will it work for dp>1?