[FEATURE] Run Wan2.2 pipelines on Ascend NPU - #42

Open

yx0716 wants to merge 10 commits into

Tele-AI:mainfrom

yx0716:feat/npu-adapt

Commits on Sep 11, 2026

fix(models): gate channels_last_3d activations to CUDA devices
yx0716
committed
cf3248c
fix(distributed): default device mesh to the detected platform
yx0716
committed
e9fd116
fix(worker): marshal parallel-worker queue tensors through CPU off CUDA
yx0716
committed
7d5c8f4
refactor(ops): move fused FP8 QKV Triton kernels into kernel.triton
yx0716
committed
74dd3ab
fix(distributed): allocate pipeline P2P buffers on the platform device
yx0716
committed
b6dea52
fix(worker): keep marshalled queue results device-agnostic
yx0716
committed
00a2124
feat(examples): auto-detect device in the wan22 5B T2V example
yx0716
committed
1b96425
fix(models): accept batched tensors in wan22 VAE parallel decode
yx0716
committed
edf966d
fix(worker): isolate HCCL socket ranges per parallel worker group
yx0716
committed
d34c842
docs: register Ascend NPU in the platform, installation, and wan_video guides
yx0716
committed
54a2b94