Commit History
-
Implement attention slicing: chunk SDPA calls along the batch*heads dim since torch 2.8s fused kernels reject this shape on this GPU, forcing an O(seq_len^2 * batch) math fallback that OOMs even a 96GB card
Verified signature63d710a -
Diagnose why SDPA auto-select avoids flash/efficient backends: try FLASH_ATTENTION explicitly and log the actual error, plus dtype/contiguity/mask info
Verified signature4df1266 -
Add memory diagnostics at each attention call to see if VRAM usage is growing unbounded or just spiking once
Verified signature4ba1e6a -
Revert forced EFFICIENT_ATTENTION backend (unsupported for these tensor shapes on this GPU); back to auto-selected SDPA
Verified signature42c2660 -
Force SDPA's memory-efficient backend for PSHuman's multiview attention: rule out a math-fallback OOM on Blackwell
Verified signatureff5e65e
-
Keep PSHuman's custom multiview attention processors but swap their inner xformers.ops.memory_efficient_attention call for torch's native scaled_dot_product_attention (xformers' Hopper kernel crashes on Blackwell/sm_120)
Verified signature2080275
-
Migrated from GitHub
Verified signature2252f3d