Hey - hitting a hard install failure on an RTX 5090 (Blackwell / sm_120, Windows, CUDA 12.8, torch 2.7.1). Install and Update both abort at the WanGP runtime verification step with a CUDA crash:
CUDA error (.../xformers/.../flash-attention/hopper/flash_fwd_launch_template.h:188): invalid argument
Traced it to scripts/runtime_verify.py - the verify() step runs xformers.ops.memory_efficient_attention(...) because xformers.ops is in the wangp recipe's verificationImports. The pinned xformers==0.0.31.post1 dispatches that call to the flash-attention Hopper kernels, which are invalid on Blackwell, and it's a hard CUDA abort - it can't be caught in Python, so it kills the whole install and no runtime profile gets written.
The thing is, the runtime itself is fine on Blackwell - the recipe already ships SageAttention 2.2.0+cu128 in acceleratorPackages and excludes flash-attn, and SDPA works too. So only the verification probe is stricter than the actual runtime. The app runs great on a 5090 once it's past that check (verified: full UI + all /api/v1 endpoints serving, GPU engaged, no runtime attention errors).
Suggested fix: make that xformers.ops probe Blackwell-aware - skip it on sm_120 (or force a non-flash backend like cutlass/SDPA for the probe), since the real runtime doesn't rely on the flash path there anyway. Happy to test a patch.
Hey - hitting a hard install failure on an RTX 5090 (Blackwell / sm_120, Windows, CUDA 12.8, torch 2.7.1). Install and Update both abort at the WanGP runtime verification step with a CUDA crash:
CUDA error (.../xformers/.../flash-attention/hopper/flash_fwd_launch_template.h:188): invalid argument
Traced it to scripts/runtime_verify.py - the verify() step runs xformers.ops.memory_efficient_attention(...) because xformers.ops is in the wangp recipe's verificationImports. The pinned xformers==0.0.31.post1 dispatches that call to the flash-attention Hopper kernels, which are invalid on Blackwell, and it's a hard CUDA abort - it can't be caught in Python, so it kills the whole install and no runtime profile gets written.
The thing is, the runtime itself is fine on Blackwell - the recipe already ships SageAttention 2.2.0+cu128 in acceleratorPackages and excludes flash-attn, and SDPA works too. So only the verification probe is stricter than the actual runtime. The app runs great on a 5090 once it's past that check (verified: full UI + all /api/v1 endpoints serving, GPU engaged, no runtime attention errors).
Suggested fix: make that xformers.ops probe Blackwell-aware - skip it on sm_120 (or force a non-flash backend like cutlass/SDPA for the probe), since the real runtime doesn't rely on the flash path there anyway. Happy to test a patch.