fix(sglang): support SGLang 0.5.11+ (version detection + refactored leak check)

[RixinLiu]

Summary

Two general (non-AMD) compatibility defects in the SGLang integration, found while porting kvcached to AMD. Bug #2 affects SGLang 0.5.11+ (verified on the 0.5.12.post1 wheel); Bug #1 affects any build that doesn't expose sglang.__version__. The fix was additionally checked against 0.4.9 (the oldest version kvcached supports) and 0.5.10 (the last version before the break) to confirm no regression on the older, still-working layouts.

Bug #1 — version detection misses builds without a module-level __version__

VersionManager.detect_version() only reads module attributes (__version__, then version/VERSION/_version). Some SGLang builds (notably source builds) expose none of these on the sglang module, even though the installed-package metadata carries the version. When detection returns None, every SGLang patch bails and kvcached silently becomes a no-op:

[kvcached][WARNING] Could not detect version for sglang
[kvcached][WARNING] Failed to apply elastic_memory_pool / ...
[kvcached][WARNING] Failed to patch sglang: elastic_allocator, ...

Fix: fall back to importlib.metadata.version(library_name) when no module attribute is found.

Bug #2 — scheduler_memory_leak patch doesn't match SGLang's newer leak-check layout

kvcached maps physical KV pages lazily, so SGLang's static-pool invariant (total == available + in-use) doesn't hold for the KV pool; the scheduler_memory_leak patch exists to neutralize SGLang's leak detector for that pool. The old patch scanned for a single Scheduler method containing both "memory leak detected" and "token_to_kv_pool_allocator".

Through 0.5.10 that still worked (check_memory names the KV allocator). At 0.5.11+, SGLang routes the KV-pool leak through a generic _report_leak(pool_name, …) that names no pool — so no single method has both literals, the patch matches nothing, fails to apply, and SGLang's own detector crashes the scheduler:

[kvcached][WARNING] Failed to apply scheduler_memory_leak
ValueError: pool memory leak detected! [full] total=..., available=..., evictable=0 ...
Received sigquit from a child process.   # exit 137

Fix: select the Scheduler leak-raisers by source ("memory leak detected") but skip any check specific to req_to_token_pool — a pool kvcached does not manage, whose invariant still holds, and whose alarm must stay live so a genuine request-pool leak still surfaces. The KV/token-pool checks are kept (the old combined check names token_to_kv_pool_allocator; the new generic _report_leak names no pool and is only ever called for token pools). This covers the old single-method layout and the new split/generic layout without over-suppressing unrelated alarms.

Works across versions (verified)

Bug #2 only breaks on 0.5.11+. The two older versions below are regression checks, not bug repros — they confirm the new matcher still neutralizes the KV-pool check exactly like the original patch did:

SGLang	KV-cache leak check	request-pool check	result
0.4.9	check_memory (KV + request fused)	(same method)	KV check neutralized — matches original patch behavior
0.5.10	check_memory (KV)	separate _check_req_pool	KV check neutralized, request check left live
0.5.12	generic _report_leak (KV)	separate _check_req_pool	KV check neutralized, request check left live — fixes the crash the old patch couldn't

No regression on older versions (behavior matches the original patch where it worked); fixes 0.5.11+; and the request-pool alarm is preserved wherever the version's structure keeps it separate.

Manifestation

	exists in code?	manifests on 0.5.12 wheel?	manifests on no-__version__ build?
Bug #1	✅ yes	❌ no (__version__ present)	✅ yes
Bug #2	✅ yes	✅ yes (leak crash, exit 137)	✅ yes

Changes

• kvcached/integration/version_utils.py — importlib.metadata fallback in detect_version().
• kvcached/integration/sglang/patches.py — generalize the scheduler_memory_leak patch across SGLang's leak-check layouts, skipping the req_to_token_pool-specific check.

Commits:

• support SGLang 0.5.11-0.5.12 (version detection + refactored leak check)
• leave req_to_token_pool leak check intact

[Copilot]

Pull request overview

This pull request improves kvcached’s SGLang integration compatibility across multiple SGLang versions by (1) making version detection more robust when sglang.__version__ is missing and (2) refactoring the scheduler leak-check suppression logic to match the reorganized leak-check code paths introduced in SGLang 0.5.11+.

Changes:

• Add an importlib.metadata.version() fallback in VersionManager.detect_version() when no module-level version attribute is available.
• Generalize the scheduler_memory_leak patch to suppress the relevant leak-raiser methods across both pre-0.5.11 and 0.5.11+ SGLang layouts, while intentionally skipping request-pool-specific checks.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
kvcached/integration/version_utils.py	Adds a distribution-metadata fallback for version detection when module attributes are absent.
kvcached/integration/sglang/patches.py	Refactors the scheduler leak-check patch to match newer SGLang leak-check structure while preserving request-pool leak detection.

---

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