[None][fix] Scope KVCM warmup capacity constraints to DeepSeek V4 - #19213

Open

yizhang-nv wants to merge 15 commits into

NVIDIA:mainfrom

yizhang-nv:codex/fix-kvcm-v2-init-warmup-budget

Commits on Oct 4, 2026

[None][fix] Respect KVCM V2 initialization and warmup budgets
yizhang-nv
committed
209b7b9
[None][test] Preserve full and mixed attention constraint coverage
yizhang-nv
committed
b1d7f2b
[None][fix] Limit initialization floor changes to self attention
yizhang-nv
committed
26033aa
[None][test] Cover multimodal examples with KVCM V2 budget fix
yizhang-nv
committed
8aee228
[None][fix] Estimate V2 cache constraints and query warmup capacity
yizhang-nv
committed
06a47f0
[None][fix] Keep V2 budget fix focused on existing CI failures
yizhang-nv
committed
8662968
[None][fix] Restore warmup capacity query after short request allocation
yizhang-nv
committed
edaa0fe
[None][test] Cover insufficient CUDA graph warmup capacity
yizhang-nv
committed
e3eaecf
[None][fix] Keep maximum-length warmup constraints specific to DeepSeek V4
yizhang-nv
committed
e4820bd
[None][test] Keep KVCM constraint test changes minimal
yizhang-nv
committed
42f0f00
[None][fix] Preserve generic context warmup constraints
yizhang-nv
committed
3ca96e8
[None][fix] Avoid counting the generation dummy input twice
yizhang-nv
committed
2876088
[None][test] Cover DeepSeek V4 warmup constraint configuration
yizhang-nv
committed
cf8bd33
[None][fix] Preserve V4 estimation batch limits and warmup coverage
yizhang-nv
committed
141fc0b
[None][test] Complete V4 sparse offload cache fixture initialization
yizhang-nv
committed
45fc4ac