Feat. MindIE-SD compilation support.

[zhangtao0408]

Test Env

device	Ascend910B3
torch	2.8.0
torch-npu	2.8.0
CANN	8.3.RC2
Device	910B3x1

Test Codes

1. Prepare Env

git clone --branch master --single-branch https://gitcode.com/Ascend/MindIE-SD.git
cd ./MindIE-SD
pip install wheel && python3 setup.py bdist_wheel
pip install ./dist/*.whl --force-reinstall --no-deps

2. Run test

FLUX_PATH=/PATH/TO/FLUX.1-dev

# Non-Compilation
python3 -m cache_dit.generate flux --model-path $FLUX_PATH

# Compilation
python3 -m cache_dit.generate flux --model-path $FLUX_PATH --compile

Results

1. Performance

Model	Batch Size	Resolution	Compile	E2E Time(s)
flux.1-dev	1	1024x1024	❌	14.09
flux.1-dev	1	1024x1024	✅	12.85

2. Output

2.1. No Compile

2.2. Compiled

[DefTruth]

@zhangtao0408 Hi~ thanks for working on this! Could please also share the results of distributed inference? w/ or w/o MindIE-SD Compilation. For example:

torchrun --nproc_per_node=4 -m cache_dit.generate flux --parallel ulysses
torchrun --nproc_per_node=4 -m cache_dit.generate flux --parallel ulysses --compile

[DefTruth]

rename to _maybe_xxx since we don't want users use this help func directly.

[DefTruth]

We should check the platform before patching. The patching should only be applied while the platform is NPU, otherwise, skip this patch.

Suggested change

	# MindIESD Backend Available
	# MindIESD Backend Available
	if current_platform.device_type != "npu":
	return # DO NOTHING