Skip to content

[UnifiedTree]: Support HiCache For DeepSeek_V4#24691

Open
hzh0425 wants to merge 2 commits intomainfrom
hybrid_tree/deepseek_v4_hicache_integrate
Open

[UnifiedTree]: Support HiCache For DeepSeek_V4#24691
hzh0425 wants to merge 2 commits intomainfrom
hybrid_tree/deepseek_v4_hicache_integrate

Conversation

@hzh0425
Copy link
Copy Markdown
Collaborator

@hzh0425 hzh0425 commented May 8, 2026

Motivation

Follow up swa hicache #23391 , this pr implement HiCache for V4 by leveraging UnifiedTree's SWA HiCache capabilities, combined with the Shadow Radix mechanism.

Collaborate with @ispobock @huangtingwei9988

Roadmap: #20415

Modifications

Accuracy Tests

image

Speed Tests and Profiling

image

Checklist

Review and Merge Process

  1. Ping Merge Oncalls to start the process. See the PR Merge Process.
  2. Get approvals from CODEOWNERS and other reviewers.
  3. Trigger CI tests with comments or contact authorized users to do so.
    • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
  4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

hzh0425 and others added 2 commits May 8, 2026 11:14
Co-authored-by: ispobock <ispobaoke@gmail.com>
Co-authored-by: 晟海 <huangtingwei.htw@antgroup.com>
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Code Review

This pull request introduces a sidecar pool mechanism to allow memory pools to share indices from source pools, replacing the previous shared anchor logic. It specifically enables DeepSeek V4 HiCache support by adding new host pool types and stack assembly logic. Feedback highlights a critical bug where the sidecar resolution logic only supports KV as a source, which will cause crashes for DeepSeek V4 state pools that derive from SWA. A redundant enum member in the storage definitions was also noted.

Comment thread python/sglang/srt/mem_cache/hybrid_cache/hybrid_cache_controller.py
Comment thread python/sglang/srt/mem_cache/hicache_storage.py
Comment thread python/sglang/srt/mem_cache/hicache_storage.py
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

deepseek hicache Hierarchical Caching for SGLang high priority run-ci

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants