[Inductor] [CPU] Model bench timeout in single thread secnario

[yudongsi]

🐛 Describe the bug

In TorchInductor CPU Performance Dashboard , We observed several models timeout in single thread secnario:

• hf_GPT2_large
• hf_T5_base
• AlbertForMaskedLM
• AlbertForQuestionAnswering

Saw in raw log:

cpu  eval  hf_GPT2_large                       TIMEOUT
cpu  eval  hf_T5_base                          TIMEOUT
cpu  eval  AlbertForMaskedLM                   TIMEOUT
cpu  eval  AlbertForQuestionAnswering          TIMEOUT

Versions

Repro CMD:

python -m torch.backends.xeon.run_cpu --core_list 0 --ncores_per_instance 1 benchmarks/dynamo/torchbench.py --performance --float32 -dcpu -n50 --inductor  --no-skip --dashboard --only hf_GPT2_large  --cold_start_latency --batch_size 1 --threads 1

python -m torch.backends.xeon.run_cpu --core_list 0 --ncores_per_instance 1 benchmarks/dynamo/torchbench.py --performance --float32 -dcpu -n50 --inductor  --no-skip --dashboard --only hf_T5_base  --cold_start_latency --batch_size 1 --threads 1

python -m torch.backends.xeon.run_cpu --core_list 0 --ncores_per_instance 1 benchmarks/dynamo/huggingface.py --performance --float32 -dcpu -n50 --inductor  --no-skip --dashboard --only AlbertForMaskedLM  --cold_start_latency --batch_size 1 --threads 1


python -m torch.backends.xeon.run_cpu --core_list 0 --ncores_per_instance 1 benchmarks/dynamo/huggingface.py --performance --float32 -dcpu -n50 --inductor  --no-skip --dashboard --only AlbertForQuestionAnswering  --cold_start_latency --batch_size 1 --threads 1

cc @ezyang @soumith @msaroufim @wconstab @ngimel @bdhirsh