Consolidate TP support and validate vLLM and SGLang - #5

Draft

rchalamala wants to merge 119 commits into

mainfrom

cursor/vllm-tp-consolidation-5d04

Commits on Jul 23, 2026

Commits on Jul 24, 2026

docs: plan multi-shape SGLang replay

cursoragent
and
rchalamala
committed
feat: add SGLang shape replay state

cursoragent
and
rchalamala
committed
fix: load shape replay state tests without foundry package init

cursoragent
and
rchalamala
committed
feat: persist SGLang shape state manifest

cursoragent
and
rchalamala
committed
feat: adapt SGLang FlashInfer shape state

cursoragent
and
rchalamala
committed
fix: harden SGLang shape activation ownership

cursoragent
and
rchalamala
committed
fix: harden SGLang shape activation cleanup

cursoragent
and
rchalamala
committed
feat: add shape-aware SGLang graph relocation

cursoragent
and
rchalamala
committed
fix: harden FlashInfer graph relocation validation

cursoragent
and
rchalamala
committed
fix: make FlashInfer relocation atomic

cursoragent
and
rchalamala
committed
feat: materialize SGLang shape replay capsules

cursoragent
and
rchalamala
committed
fix: harden SGLang shape capsule lifecycle

cursoragent
and
rchalamala
committed
feat: gate SGLang multi-shape replay

cursoragent
and
rchalamala
committed
fix: require explicit batch schedule in multi-shape hook

cursoragent
and
rchalamala
committed
test: verify every SGLang replay shape

cursoragent
and
rchalamala
committed
test: harden SGLang replay shape validation harness

cursoragent
and
rchalamala
committed
fix: adopt padded replay metadata for SGLang

cursoragent
and
rchalamala
committed
test: parse SGLang replay from decode graph key size

cursoragent
and
rchalamala
committed
test: require eager decode evidence for max+1 fallback

cursoragent
and
rchalamala
committed
test: force deterministic eager fallback decode sampling

cursoragent
and
rchalamala
committed
docs: normalize task report list formatting

cursoragent
and
rchalamala
committed
docs: record SGLang multi-shape milestones

cursoragent
and
rchalamala
committed
chore: remove tracked SDD scratch reports

cursoragent
and
rchalamala
committed
test: batch graph-shape exercises in one generate request

cursoragent
and
rchalamala
committed
test: exercise graph shapes largest-to-smallest

cursoragent
and
rchalamala
committed
test: reject duplicate graph exercise batch keys

cursoragent
and
rchalamala
committed