Repository navigation

feat(cuda): ggml-CUDA EdgeTAM backend (NVIDIA fast-path) - #3

Draft

njayj wants to merge 16 commits into

claude/sam3cpp-coreml-trackerfrom

claude/cuda-nvidia-edgetam

Draft

feat(cuda): ggml-CUDA EdgeTAM backend (NVIDIA fast-path)#3
njayj wants to merge 16 commits into
claude/sam3cpp-coreml-trackerfrom
claude/cuda-nvidia-edgetam

Commits on Jun 23, 2026
Commits on Jul 4, 2026
- build(capi): SAM3_CAPI_SHARED C-ABI DLL target + force CUDA graphs on SAM3_CUDA
  
  njayj
  and
  claude
  committed
  2604544
- feat(cuda): guarded runtime fallback chain + ggml encoder-ahead producer
  
  njayj
  and
  claude
  committed
  a3194c7
Commits on Jul 7, 2026
- cuda: fall back to manual attention for head dims ggml-cuda flash-attn lacks
  
  Hunter-Dolan
  and
  claude
  committed
  d0d2c45
- cuda: stage-transport driver increment 1 — sam3_stage_feed + the 4 neck_trk edges
  
  Hunter-Dolan
  and
  claude
  committed
  356475e
- cuda: stage-transport increment 2 — constant device twins + weight-read caches
  
  Hunter-Dolan
  and
  claude
  committed
  d35fab4
Commits on Jul 8, 2026
- cuda: stage-transport increment 3 — fixed bank ring + device prompt assembly
  
  Hunter-Dolan
  and
  claude
  committed
  092b313
- cuda: stage-transport increment 4 — device memenc→perceiver chain
  
  Hunter-Dolan
  and
  claude
  committed
  e1b5665
- cuda: stage-transport increment 5 — device default on CUDA + async D2D feeds
  
  Hunter-Dolan
  and
  claude
  committed
  edb60a5
- cuda: increment 6a — cached propagate stage graph (CUDA-graph replay path)
  
  Hunter-Dolan
  and
  claude
  committed
  f5875db
- cuda: increment 6b — cached memory-encoder + perceiver stage graphs
  
  Hunter-Dolan
  and
  claude
  committed
  b8e7ecb
- cuda: increment 6c — cached image-encoder stage graph
  
  Hunter-Dolan
  and
  claude
  committed
  e865052
- cuda: increment 7 — device-resident encoder-ahead seam (18 fps threaded)
  
  Hunter-Dolan
  and
  claude
  committed
  56a27ec
Commits on Jul 11, 2026
- cuda: device frame preprocess (plan v1 step 5)
  Hunter-Dolan
  committed
  3772b35

Navigation Menu

Repository navigation

feat(cuda): ggml-CUDA EdgeTAM backend (NVIDIA fast-path) - #3

feat(cuda): ggml-CUDA EdgeTAM backend (NVIDIA fast-path)#3
njayj wants to merge 16 commits into
claude/sam3cpp-coreml-trackerfrom
claude/cuda-nvidia-edgetam

Commits on Jun 23, 2026

feat(build): SAM3_CUDA ggml-CUDA backend option (NVIDIA fast-path)

feat(cuda): select the ggml-CUDA backend in the model loader (CUDA->Vulkan->CPU)

feat(capi): expose active ggml backend name (anti silent-CPU-fallback)

Commits on Jul 4, 2026

build(capi): SAM3_CAPI_SHARED C-ABI DLL target + force CUDA graphs on SAM3_CUDA

feat(cuda): guarded runtime fallback chain + ggml encoder-ahead producer

Commits on Jul 7, 2026

cuda: fall back to manual attention for head dims ggml-cuda flash-attn lacks

cuda: stage-transport driver increment 1 — sam3_stage_feed + the 4 neck_trk edges

cuda: stage-transport increment 2 — constant device twins + weight-read caches

Commits on Jul 8, 2026

cuda: stage-transport increment 3 — fixed bank ring + device prompt assembly

cuda: stage-transport increment 4 — device memenc→perceiver chain

cuda: stage-transport increment 5 — device default on CUDA + async D2D feeds

cuda: increment 6a — cached propagate stage graph (CUDA-graph replay path)

cuda: increment 6b — cached memory-encoder + perceiver stage graphs

cuda: increment 6c — cached image-encoder stage graph

cuda: increment 7 — device-resident encoder-ahead seam (18 fps threaded)

Commits on Jul 11, 2026

cuda: device frame preprocess (plan v1 step 5)