[TRTLLM-4971]: Use safe deserialization in ParallelConfig

[yibinl-nvidia]

Use safe deserialization in ParallelConfig

Description

Remove unsafe pickle.load usage in ParallelConfig by using a safe deserialization method. During the tests, I realized that the current location of serialization.py under executor directory will cause a circular import of the files that import ParallelConfig, so I move the serialization.py outside of executor directory.

Test Coverage

Added a new test test_parallel_config_serialization to check the correctness of safe serialization / deserialization of ParallelConfig.

[yibinl-nvidia]

/bot run --add-multi-gpu-test --disable-fail-fast

[tensorrt-cicd]

PR_Github #6334 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #6334 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #4624 completed with status: 'FAILURE'

[yibinl-nvidia]

/bot run --add-multi-gpu-test --disable-fail-fast

[tensorrt-cicd]

PR_Github #6350 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #6350 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #4637 completed with status: 'FAILURE'

[yibinl-nvidia]

/bot run --add-multi-gpu-test --disable-fail-fast

[tensorrt-cicd]

PR_Github #6397 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #6397 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #4677 completed with status: 'FAILURE'

[yibinl-nvidia]

/bot run --add-multi-gpu-test --disable-fail-fast

[tensorrt-cicd]

PR_Github #6501 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #6501 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #4757 completed with status: 'SUCCESS'

[yibinl-nvidia]

@kaiyux Who do you think would be the best reviewer for this PR? I checked contribution history to ParallelConfig and your name pops up :)

[kaiyux]

@yuxianq Since it's related to auto parallel, can you help take a look?

[yibinl-nvidia]

/bot run --add-multi-gpu-test --disable-fail-fast

[tensorrt-cicd]

PR_Github #7243 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #7243 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #5245 completed with status: 'SUCCESS'

[yibinl-nvidia]

/bot run --add-multi-gpu-test --disable-fail-fast

[tensorrt-cicd]

PR_Github #8981 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #8981 [ run ] completed with state FAILURE
/LLM/main/L0_MergeRequest_PR pipeline #6556 completed with status: 'FAILURE'

[yibinl-nvidia]

/bot run --add-multi-gpu-test --disable-fail-fast

[tensorrt-cicd]

PR_Github #9043 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #9043 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #6610 completed with status: 'FAILURE'

[yibinl-nvidia]

/bot run --add-multi-gpu-test --disable-fail-fast

[tensorrt-cicd]

PR_Github #9224 [ run ] triggered by Bot

[yibinl-nvidia]

/bot kill

[tensorrt-cicd]

PR_Github #9258 [ kill ] triggered by Bot

[tensorrt-cicd]

PR_Github #9258 [ kill ] completed with state SUCCESS
Successfully killed previous jobs for commit d0caf72

[yibinl-nvidia]

/bot run --add-multi-gpu-test

[tensorrt-cicd]

PR_Github #9261 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #9261 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #6795 completed with status: 'SUCCESS'
Pipeline passed with automatic retried tests. Check the rerun report for details.

[yibinl-nvidia]

@kaiyux @Superjomn This MR relocates serialization.py to a better location and apply safe deserialization to ParallelConfig. I would expect very few updates to the BASE_AUTOPP_CLASSES (unlike IPC) so we should be safe to do this change. Could you review and merge this PR? Thanks.