[Ascend NPU] Improve torchbenchmark results

[shink]

Some benchmarks failed on NPU, there are still something we can do to improve them.

Torchbenchmark issues tracker

CV models

• vision_maskrcnn

Installtation problems

• fastNLP_Bert
• opacus_cifar10 (Network problem)
• torch_multimodal_clip (Network problem)

OOM

• llava

NPU out of memory. Tried to allocate 174.00 MiB (NPU 0; 29.50 GiB total capacity; 28.26 GiB already allocated; 28.26 GiB current active; 34.72 MiB free; 29.10 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation.

Runtime Error

• sam_fast: No module named 'triton'
• timm_efficientdet: not_implemented
• simple_gpt: not_implemented
• simple_gpt_tp_manual: not_implemented