Repository navigation

feat: multi-GPU replicas with workload-aware placement (0.12.0) - #38

Merged

amrit110 merged 7 commits into

mainfrom

feat/multi-gpu-engine

Oct 18, 2026

Merged

feat: multi-GPU replicas with workload-aware placement (0.12.0)#38
amrit110 merged 7 commits into
mainfrom
feat/multi-gpu-engine

Commits on Oct 17, 2026
- feat: multi-GPU replicas (MultiDeviceEngine) + workload benchmark
  
  amrit110
  and
  claude
  committed
  0bc6651
- docs(roadmap): PyTorch-level first, custom kernels only where profiles demand
  
  amrit110
  and
  claude
  committed
  24dd3d2
- docs: multi-GPU README section, 0.12.0 changelog (unreleased), bump version
  
  amrit110
  and
  claude
  committed
  7ceebb9
- bench: profiling harness (#31), copy bandwidth (#39), GKE GPU devbox
  
  amrit110
  and
  claude
  committed
  4c00471
- fix(multidevice): review findings on #38
  
  amrit110
  and
  claude
  committed
  b4bfa30
- fix(backends): initialize CUDA before device-explicit memory stats
  
  amrit110
  and
  claude
  committed
  81e7797
- perf(tabpfn): share the model across contexts (~42x less GPU per context); spill without weights; warm up weights at start
  
  amrit110
  and
  claude
  committed
  87e5180