Refactor/npu transformer cleanup - #272

Open

fxp19950422 wants to merge 20 commits into

modelscope:v1from

fxp19950422:refactor/npu-transformer-cleanup

Commits on Aug 21, 2026

feat(npu): add Ascend platform abstraction and MindIE attention

authored and
Super User
committed
e9dd2b0
feat(npu): enable Ulysses SP with CFG parallel on Ascend

authored and
Super User
committed
a030fcb
feat(npu): fuse Qwen DiT addcmul, MindIE RoPE, and LN modulate

authored and
Super User
committed
8d1302c
Add AscendLongContextAttention: NPU/MindIE long-context attention

root
authored and
Super User
committed
bfa1663
feat(npu): long-context attention, fused ops, unified platform abstraction

gaoyuanyuanqiqi
authored and
Super User
committed
7b8f81e
feat(npu): extract AscendLongContextAttention to standalone file
Super User
committed
6704256
feat(npu): add attention factory function for platform-aware creation
Super User
committed
dc3e080
test(npu): add unit tests for attention factory
Super User
committed
405e697
feat(npu): add unified platform ops interface for fused operators
Super User
committed
b49b54a
test(npu): add unit tests for platform ops
Super User
committed
2254061
refactor(npu): replace NPU hardcoded branches with unified platform ops and attention factory
Super User
committed
7ca6b99
test(npu): add NPU multi-card parallel tests (4-card and 8-card Ulysses SP)
Super User
committed
6c4cc32
test(npu): add NPU single-card integration tests for all Qwen Image scenes
Super User
committed
33cdd49

Commits on Aug 22, 2026

docs(npu): torch.compile FFN exploration - no benefit on current NPU stack
Super User
committed
2e5f17a
docs(npu): add NPU performance analysis report
Super User
committed
050fb35
docs(npu): add NPU adaptation technical documentation
Super User
committed
cc717c9
perf(multicard): add GPU/NPU multi-card scaling profiling data
Super User
committed
aa567d4
perf(multicard): verify CFG parallel P0 optimization (+102% on 8-card NPU)
Super User
committed
e5d744f
bench: add PR#270 vs refactored same-methodology NPU comparison
Super User
committed
84acce6

Commits on Aug 27, 2026

fix(attention): add attn_type guard to factory dispatch and honor causal in MindIE backend
Super User
committed
a837164