feat: NPU operator adaptations for MindIE-SD#243

Open

Chitandaaaaa wants to merge 29 commits into

modelscope:v1from

Chitandaaaaa:feat/npu-adaptations

Commits on Apr 22, 2026

feat: add NPU FastGELUMLP (task 1 of 4)
hammer
committed
feat: add NPU RMSNorm wrapper (task 2 of 4)
hammer
committed
feat: add MINDIE attention backend for NPU (task 3 of 4)
hammer
committed

Commits on Apr 30, 2026

Commits on May 6, 2026

fix: unsqueeze gates to [B,1,D] for broadcast with 3D tensors
hammer
committed
fix: align img modulation with txt under zero_cond_t to fix 2*B vs B crash
hammer
committed
fix: only auto-switch to MINDIE on NPU when user doesn't specify attn_type
hammer
committed

Commits on May 15, 2026

Commits on May 19, 2026

Commits on May 20, 2026

feat: add single-block profiling hook point for NPU performance analysis
hammer
committed

Commits on May 21, 2026

Commits on May 25, 2026

feat: add NPU profiling hook for multi-worker distributed inference
hammer
committed

Commits on May 29, 2026

add progress logging for CPU-to-NPU weight transfer in load_model_weights
hammer
committed
fix: stream safetensor shards into modules to reduce Host RAM peak

hammer
and
cursoragent
committed
fix: resolve int profiler config to torch_npu enums in multi-worker setup

hammer
and
cursoragent
committed
fix: lazy-resolve profiler enums for torch_npu version differences

hammer
and
cursoragent
committed
perf: speed up weight loading with auto bulk vs streaming

hammer
and
cursoragent
committed