Repository navigation

[FlowANN step 3/12] Add optional GDRCopy queue transport - #2979

Open

ZhaoHaoRu wants to merge 38 commits into

NVIDIA:mainfrom

ZhaoHaoRu:zhr/flowann-v3-03-gdrcopy

Open

[FlowANN step 3/12] Add optional GDRCopy queue transport#2979
ZhaoHaoRu wants to merge 38 commits into
NVIDIA:mainfrom
ZhaoHaoRu:zhr/flowann-v3-03-gdrcopy

Commits on Oct 29, 2026
- [FlowANN step 1/12] Refactor CAGRA traversal to accept graph providers
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  4a99eaa
- [FlowANN step 2/12] Add graph storage and the CUDA-copy queue runtime
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  b29b06d
- [FlowANN step 3/12] Add optional GDRCopy queue transport
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  befcb9e
Commits on Nov 2, 2026
- Fix provider candidate tails and preserve resident invalid IDs
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  282a915
- Merge FlowANN step 1 review fixes into step 2
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  afe8a2c
- Add regressions for graph replacement and context moves
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  a75bd4c
- Keep graph replacement transactional and sessions stable across context moves
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  cf42977
- Complete invalid queue requests and unblock GPU waiters on poller failure
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  7d46ec7
- Merge FlowANN step 2 review fixes into step 3
  ZhaoHaoRu
  committed
  eb61427
- Fix GDRCopy page-sized mappings and installed runtime discovery
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  ae20f4b
- Rate-limit emergency host flag reads to preserve queue throughput
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  2628058
- Merge FlowANN step 2 review fixes into step 3
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  e3298c5
- Use the native CUDA stream handle in the regression kernel launch
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  742c612
- Merge FlowANN step 2 review fixes into step 3
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  b128ea3
- Defer emergency host reads until normal queue completion can progress
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  641d2d1
- Merge FlowANN step 2 review fixes into step 3
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  a3f0d3c
- Format FlowANN review corrections with repository tools
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  15ae502
- Format FlowANN review corrections with repository tools
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  9d94fc9
- Use the shared RAPIDS formatting configuration for CMake
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  b2d14ab
- Limit CMake changes to the GDRCopy runtime path and FlowANN source includes
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  a5f4081
- Merge NVIDIA cuVS main into FlowANN step 1
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  adbc73f
- Merge FlowANN step 1 review fixes into step 2
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  231345e
- Merge FlowANN step 2 review fixes into step 3
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  13eb469
- Adapt FlowANN step 2 to upstream dataset data and manifest views
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  73285d9
- Merge FlowANN step 2 review fixes into step 3
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  f33bc19
- Construct empty padded views with the upstream dataset factory
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  59c8127
- Merge FlowANN step 2 review fixes into step 3
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  c43edc8
- Use the upstream row-major view for cosine norm reduction
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  d06aaa4
- Merge FlowANN step 2 review fixes into step 3
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  e204115
- Preserve output-width invalid IDs for mapped providers and remove a redundant barrier
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  1353ed7
- Merge FlowANN step 1 review fixes into step 2
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  6ac3fc5
- Cover abort draining, concurrent lease failures and recovery after invalid requests
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  95d1dd5
- Merge FlowANN step 2 review fixes into step 3
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  0f7db48
- Latch queue aborts and report each lease failure while allowing invalid-request recovery
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  ca636a2
- Define relocatable GDRCopy runtime deployment paths for shared and static C API linkage
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  16a0e93
- Format poller failure reporting with the repository C++ style
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  12844a5
- Merge FlowANN step 2 review fixes into step 3
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  53320fb
- Honor external GDRCopy runtime paths in standalone static C builds
  
  ZhaoHaoRu
  and
  He-Jingkai
  committed
  725aa90

Commits

Commits on Oct 29, 2026

Commits on Nov 2, 2026