Add CPU fallback for triton EDT on macOS - #510

Open

msmackle wants to merge 14 commits into

facebookresearch:mainfrom

msmackle:fix/triton-macos-fallback

Commits on Apr 2, 2026

Add CPU fallback for triton EDT on macOS

mason-apmc
and
claude
committed
7c04b4b
Fix hardcoded device='cuda' in PositionEmbeddingSine precompute

mason-apmc
and
claude
committed
2a537d0
Fix all remaining hardcoded device='cuda' for macOS compatibility

mason-apmc
and
claude
committed
433d815
Fix hardcoded .cuda() calls in io_utils.py for macOS/CPU support

mason-apmc
and
claude
committed
5b6e96a
Add CPU/MPS fallback for addmm_act fused CUDA kernel

mason-apmc
and
claude
committed
86c0112
Fix bfloat16 matmul on MPS; add MPS to io_utils device selection

mason-apmc
and
claude
committed
265767d

Commits on Apr 3, 2026

Use float32 image tensors on non-CUDA devices to match model dtype

mason-apmc
and
claude
committed
d7aef74
Guard explicit .to(bfloat16) casts behind CUDA availability check

mason-apmc
and
claude
committed
88d3910
Fix addmm_act: don't cast to bfloat16 before the non-CUDA branch check

mason-apmc
and
claude
committed
555c78b
Remove MPS from _DEVICE in io_utils — must match model device (cpu)

mason-apmc
and
claude
committed
4523d18
Guard pin_memory().to() calls behind CUDA check for macOS compatibility

mason-apmc
and
claude
committed
abf441a
Guard CUDA autocast and storage_device in Sam3TrackerPredictor for non-CUDA

mason-apmc
and
claude
committed
8a67090
Replace hardcoded .cuda() calls with .to(device) for non-CUDA compatibility

mason-apmc
and
claude
committed
48d2eaa

Commits on Apr 10, 2026

Fix empty-batch crash in connected_components_cpu

mason-apmc
and
claude
committed
56c4875