feat: add MindIE-SD as optional NPU attention and compilation backend#1004
Open
blian6 wants to merge 2 commits into
Open
feat: add MindIE-SD as optional NPU attention and compilation backend#1004blian6 wants to merge 2 commits into
blian6 wants to merge 2 commits into
Conversation
added 2 commits
May 10, 2026 11:32
> > Add MindIE-SD as a high-priority optional attention backend for Ascend NPU, > with automatic device-aware backend selection and compilation support. > > Modified files (8): > - attention/backends/register.py: add _MINDIESD_LASER enum > - attention/backends/npu.py: _mindiesd_laser_attention backend, > bridging to mindiesd.layers.attention_forward > - kernels/backend.py: KernelBackend.MINDIESD enum > - kernels/ops.py: MINDIESD kernel routing with PT fallback > - caching/cache_interface.py: auto-select attention backend in > enable_cache(); auto-enable MindieSDBackend compile on NPU > - compile/utils.py: skip CUDA-specific inductor configs on NPU > - envs.py: CACHE_DIT_ENABLE_MINDIESD_ATTN and > CACHE_DIT_FORCE_DISABLE_MINDIESD_COMPILE_CONFIG env vars > - _utils/utils.py: add _mindiesd_laser to CLI choices > > New file (1): > - _utils/backend_selector.py: BackendSelector.auto_select() > > Verified on Ascend 910B with FLUX.1-dev 512x512 and 1024x1024. > MindIE-SD backend matches native_npu performance with no regression, > and compile fusion provides +6% additional speedup at 1024x1024.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Add MindIE-SD as a high-priority optional attention backend for Ascend NPU devices,
with automatic device-aware backend selection and compilation support. MindIE-SD
uses laser_attention / fused_attn_score to accelerate attention operations on Ascend
hardware, and is auto-enabled when the
mindiesdpackage is importable.Key Design Decisions
import mindiesdwrapped intry/except Exception, gracefully falls back to_native_npuCACHE_DIT_FORCE_DISABLE_MINDIESD_COMPILE_CONFIG=1to disableCACHE_DIT_ENABLE_MINDIESD_ATTNselects attention backend ("1"=laser,"0"=native_npu); default is"1"(laser)enable_cacheflow (attention → parallelism → quantization → compile)Verification
_mindiesd_laser REGISTERED,KernelBackend.MINDIESD supportedCACHE_DIT_ENABLE_MINDIESD_ATTN=0/1/laser,CACHE_DIT_FORCE_DISABLE_MINDIESD_COMPILE_CONFIG=0/1all effective_native_npu: 0.693s/step_mindiesd_laser: 0.687s/step (laser activated, S=4608)_mindiesd_laser+MindieSDBackendcompile: 0.652s/step (+6% over baseline)