Skip to content

Write-back policy fix for unified tree - #31845

Merged
xiezhq-hermann merged 4 commits into
mainfrom
unified_tree_write_back_fix
Jul 24, 2026
Merged

xiezhq-hermann merged 4 commits into
mainfrom
unified_tree_write_back_fix

Conversation

@xiezhq-hermann

@xiezhq-hermann xiezhq-hermann commented Jul 20, 2026

Copy link
Copy Markdown
Collaborator

Motivation

We priorly introduced a few fixes for write-back policy with HiRadixCache: #28614. This PR ports this to UnifiedRadixCache as well as new unit tests.

Modifications

Accuracy Tests

Speed Tests and Profiling

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.

CI States

Latest PR Test (Base): ❌ Run #29986203847
Latest PR Test (Extra): ✅ Run #29986203687

@gemini-code-assist

Copy link
Copy Markdown
Contributor

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

@xiezhq-hermann xiezhq-hermann self-assigned this Jul 20, 2026
@xiezhq-hermann xiezhq-hermann added hicache Hierarchical Caching for SGLang run-ci run-ci-extra labels Jul 20, 2026
@hzh0425

hzh0425 commented Jul 21, 2026

Copy link
Copy Markdown
Collaborator

/rerun-group /radix_cache/unified_radix_tree/

@github-actions

github-actions Bot commented Jul 21, 2026

Copy link
Copy Markdown
Contributor

Results for /rerun-group /radix_cache/unified_radix_tree/:

🚀 4-gpu-h100 (4 tests): ✅ View workflow run

cd test/ && python3 registered/radix_cache/unified_radix_tree/test_unified_radix_cache_hicache_pp_kl.py
cd test/ && python3 registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_cp.py
cd test/ && python3 registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_dsv4.py
cd test/ && python3 registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_mamba.py

🚀 8-gpu-h200 (2 tests): ❌ View workflow run

cd test/ && python3 registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_dsv4_pp.py
cd test/ && python3 registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_nightly.py

🚀 2-gpu-h100 (2 tests): ✅ View workflow run

cd test/ && python3 registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_full.py
cd test/ && python3 registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_swa.py

@xiezhq-hermann
xiezhq-hermann merged commit 5da0b6e into main Jul 24, 2026
332 of 346 checks passed
@xiezhq-hermann
xiezhq-hermann deleted the unified_tree_write_back_fix branch July 24, 2026 16:58
Jialin added a commit to Jialin/sglang that referenced this pull request Jul 24, 2026
…l-project#31845)

Trunk's write-back fix lands the host-lock pin (already carried by the
split) plus its companion: when a write-back D-leaf's D->H backup fails
under host memory pressure, drop the unbacked subtree (declining if any
node is locked) so device eviction keeps making progress. Port the
companion into the split structure: the subtree walk, layer release, and
unbacked-leaf deletion live on UnifiedTreeCore (frees flow through the
device/host free dicts); the controller invokes the fallback when the
deferred write-back BackupKV comes back unwritten. Ports trunk's two
write-back regression tests adapted to split node handles.
Jialin added a commit to Jialin/sglang that referenced this pull request Jul 25, 2026
…l-project#31845)

Trunk's write-back fix lands the host-lock pin (already carried by the
split) plus its companion: when a write-back D-leaf's D->H backup fails
under host memory pressure, drop the unbacked subtree (declining if any
node is locked) so device eviction keeps making progress. Port the
companion into the split structure: the subtree walk, layer release, and
unbacked-leaf deletion live on UnifiedTreeCore (frees flow through the
device/host free dicts); the controller invokes the fallback when the
deferred write-back BackupKV comes back unwritten. Ports trunk's two
write-back regression tests adapted to split node handles.
jinzhenfan pushed a commit to jinzhenfan/sglang that referenced this pull request Jul 29, 2026
jakki-amd pushed a commit to jakki-amd/sglang that referenced this pull request Sep 9, 2026
Atituiset pushed a commit to Atituiset/sglang that referenced this pull request Sep 10, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants