引入上下文长度无关的 decode CUDA Graph 路径 - #28

Merged

kuma-loong merged 22 commits into

mainfrom

refactor/new-agents-md

Aug 26, 2026

Commits on Aug 26, 2026

perf: precapture bounded decode cuda graphs
kuma-loong
committed
55d8239
perf: capture larger cuda graph shapes first
kuma-loong
committed
798dbbf
feat: restore batch-only decode graph providers
kuma-loong
committed
e6ef95a
feat: add typed batch-only decode graph contract
kuma-loong
committed
014d524
fix: use live decode graph runner during startup
kuma-loong
committed
6dac096
fix: inspect captured graphs through live runner
kuma-loong
committed
67b05d1
fix: read current decode graph probe summary
kuma-loong
committed
f09ef4c
test: close batch-only stage two coverage gaps
kuma-loong
committed
84dda47
feat: productionize batch-only decode providers
kuma-loong
committed
f997ede
feat: productionize batch-only tilelang mla scores
kuma-loong
committed
dcef2db
fix: support strided tilelang mla score views
kuma-loong
committed
01afe24
fix: preserve sparse mla score contracts
kuma-loong
committed
491116b
test: cover sparse decode topology partitioning
kuma-loong
committed
cc3192d
test: cover tilelang mla replay through 64k
kuma-loong
committed
cb22fed
test: cover sparse graph path transitions
kuma-loong
committed
5f8d962
refactor: normalize batch-only decode providers
kuma-loong
committed
378d177
fix: pin sglang kernel abi
kuma-loong
committed
add354a
fix: require native multiprocessor count for gemma4
kuma-loong
committed
7c6e11a
docs: add batch-only operator review rules
kuma-loong
committed
86d98fe
refactor: remove redundant decode score method set
kuma-loong
committed
e47c689
feat: support deltakv batch-only decode graphs
kuma-loong
committed
6f7b847
feat: default decode graphs to batch-only
kuma-loong
committed
4aada47

Navigation Menu

引入上下文长度无关的 decode CUDA Graph 路径 - #28

引入上下文长度无关的 decode CUDA Graph 路径#28
kuma-loong merged 22 commits into
mainfrom
refactor/new-agents-md

Commits on Aug 26, 2026

perf: precapture bounded decode cuda graphs

perf: capture larger cuda graph shapes first

feat: restore batch-only decode graph providers

feat: add typed batch-only decode graph contract

fix: use live decode graph runner during startup

fix: inspect captured graphs through live runner

fix: read current decode graph probe summary

test: close batch-only stage two coverage gaps

feat: productionize batch-only decode providers

feat: productionize batch-only tilelang mla scores

fix: support strided tilelang mla score views

fix: preserve sparse mla score contracts

test: cover sparse decode topology partitioning

test: cover tilelang mla replay through 64k

test: cover sparse graph path transitions

refactor: normalize batch-only decode providers

fix: pin sglang kernel abi

fix: require native multiprocessor count for gemma4

docs: add batch-only operator review rules

refactor: remove redundant decode score method set

feat: support deltakv batch-only decode graphs

feat: default decode graphs to batch-only