[Feature][KVCache] Implement Cache Manager V1 with multi-GPU support and MTP KV Cache#7097

Open

kevincheng2 wants to merge 17 commits intoPaddlePaddle:developfrom

kevincheng2:feature/cache_manager_v1_rebase

Commits on Apr 1, 2026

[Feature][KVCache] Support cache manager v1 architecture

kevincheng2
and
claude
committed
Update cache manager and related modules

kevincheng2
and
claude
committed
chore: update cache_manager and related modules

kevincheng2
and
claude
committed
fix: add node to evictable set in complete_swap_to_device

kevincheng2
and
claude
committed
feat: update cache manager v1 and related modules

kevincheng2
and
claude
committed
feat(cache): add cache controller v1 implementation
kevincheng2
committed
feat(cache_manager): update cache manager v1
kevincheng2
committed
fix(cache_manager): 修复 swap_cache H2D/D2H 方向的 block_ids 逻辑并清理 ForwardMeta

kevincheng2
and
claude
committed
feat(cache_manager): refactor cache manager v1 and optimize swap ops
kevincheng2
committed
[KVCache][MTP] 支持 cache_manager_v1 下的 MTP KV Cache 初始化及多模态 hash
kevincheng2
committed
fix(cache_manager): multi-GPU fix, mm hash boundary fix, and remove batch ops

kevincheng2
and
claude
committed
[BugFix][KVCache] fix List import and move write_policy normalization to CacheManager

kevincheng2
and
claude
committed
[BugFix][KVCache] fix pre-commit code style issues

kevincheng2
and
claude
committed
[Feature][KVCache] update cache_manager_v1 modules

kevincheng2
and
claude
committed
[Feature][KVCache] add BatchRequest.from_tasks and refactor worker task parsing

kevincheng2
and
claude
committed
[Feature][KVCache] add NUMA affinity for host cache and skip swap cache tests

kevincheng2
and
claude
committed
[BugFix][KVCache] fix unittest failures for cache_manager_v1

kevincheng2
and
claude
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Feature][KVCache] Implement Cache Manager V1 with multi-GPU support and MTP KV Cache#7097

[Feature][KVCache] Implement Cache Manager V1 with multi-GPU support and MTP KV Cache#7097
kevincheng2 wants to merge 17 commits intoPaddlePaddle:developfrom
kevincheng2:feature/cache_manager_v1_rebase

Commits on Apr 1, 2026

[Feature][KVCache] Support cache manager v1 architecture

Update cache manager and related modules

chore: update cache_manager and related modules

fix: add node to evictable set in complete_swap_to_device

feat: update cache manager v1 and related modules

feat(cache): add cache controller v1 implementation

feat(cache_manager): update cache manager v1

fix(cache_manager): 修复 swap_cache H2D/D2H 方向的 block_ids 逻辑并清理 ForwardMeta

feat(cache_manager): refactor cache manager v1 and optimize swap ops

[KVCache][MTP] 支持 cache_manager_v1 下的 MTP KV Cache 初始化及多模态 hash

fix(cache_manager): multi-GPU fix, mm hash boundary fix, and remove batch ops

[BugFix][KVCache] fix List import and move write_policy normalization to CacheManager

[BugFix][KVCache] fix pre-commit code style issues

[Feature][KVCache] update cache_manager_v1 modules

[Feature][KVCache] add BatchRequest.from_tasks and refactor worker task parsing

[Feature][KVCache] add NUMA affinity for host cache and skip swap cache tests

[BugFix][KVCache] fix unittest failures for cache_manager_v1