docs(sglang): correct session radix availability - #11241
Conversation
Signed-off-by: Ishan Dhanani <ishandhanani@gmail.com>
|
/ok-to-test e3a988e |
WalkthroughDocumentation and inline comments across CLAUDE.md, several docs files, and SGLang launch scripts are updated to remove claims that SGLang supports passive session-aware radix KV ownership, replacing them with descriptions of priority-based KV eviction and clarifying that session headers remain used for tracing and router affinity only. ChangesSession Identity Documentation Correction
Estimated code review effort: 1 (Trivial) | ~5 minutes Related PRs: None identified. Suggested labels: documentation Suggested reviewers: None identified. 🐰 A rabbit hops through docs so fine, 🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
Comment |
Signed-off-by: Ishan Dhanani <ishandhanani@gmail.com>
Correct the SGLang agent documentation and launcher commentary to match the capabilities shipped in SGLang 0.5.14. This prevents users from expecting passive session-radix behavior that the release cannot provide.
CLOSES: DYN-3400
How This Was Implemented
session_paramsrequires an explicitly opened session and is not a passive KV ownership tag.Walkthrough
Mental model
SGLang 0.5.14 accepts
session_params, but its scheduler routes those requests through the explicit session controller and rejects unknown IDs. The passive session-radix implementation and first-class session identity landed separately from the v0.5.14 release tag, so Dynamo's disabled compatibility gate must not be described as an active feature.The updated guidance distinguishes three behaviors: session headers remain available for tracing and router affinity,
--radix-eviction-policy prioritycontrols eviction ordering only, andsession_paramsbelongs to SGLang's explicit session lifecycle.Boundaries and limitations
This PR does not change runtime request handling or enable session-radix ownership. Keeping the current runtime no-op avoids forwarding arbitrary header IDs as
session_params.id, which would make SGLang 0.5.14 reject requests whose sessions were not opened first.Validation
pre-commit run --files <changed files>fern checkfern docs broken-linksbash -n examples/backends/sglang/launch/agg_agent.sh examples/backends/sglang/launch/_test_agg.shgit diff --check origin/main...HEADSummary by CodeRabbit