fix potential issues in allreduce fusion kernel and ut

[yilin-void]

potential issues:

1. The float val[4] is not aligned to 16 bytes, which may result in a misaligned address when reinterpret_cast val to float4*.
2. There is a typo in the Lamport kernel when handling -0 in the data.
3. FP4 quant is not supported when sm < 100.

[yilin-void]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #4860 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #4860 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #3518 completed with status: 'FAILURE'

[yilin-void]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #4947 [ run ] triggered by Bot

[yilin-void]

/bot kill

[tensorrt-cicd]

PR_Github #4964 [ kill ] triggered by Bot

[tensorrt-cicd]

PR_Github #4947 [ run ] completed with state ABORTED

[yilin-void]

/bot run

[tensorrt-cicd]

PR_Github #4964 [ kill ] completed with state SUCCESS
Successfully killed previous jobs for commit 92249f1

[tensorrt-cicd]

PR_Github #4967 [ run ] triggered by Bot

[hyukn]

Some questions about these potential issues, not sure if I understand them correctly:
1 will not cause illegal memory access, but some precision issue? Because the data is misaligned in the wrong order.
2 will cause overflow/precision issues?
3 will not be a true issue because we have high level path to guarantee that this kernel will not be triggered before Blackwell?

[tensorrt-cicd]

PR_Github #4967 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #3606 completed with status: 'FAILURE'

[yilin-void]

hi @hyukn ,

1 will not cause illegal memory access, but some precision issue? Because the data is misaligned in the wrong order.

This does not cause any problems at present, but customers have reported that this will cause compute-sanitizer to prompt misaligned address.

2 will cause overflow/precision issues?

If the last 8 bytes of a consecutive 16-byte segment in the actual data, when interpreted as two floats, happen to be [-0.f, 0.f], it will cause a hang. However, this bug has not been observed to be triggered so far.

3 will not be a true issue because we have high level path to guarantee that this kernel will not be triggered before Blackwell?

This issue is only relevant to the cpp unit test, in actual use, we will not use fp4 on pre-Blackwell.

[yilin-void]

/bot run

[hyukn]

Thanks a lot for the detailed explanation. LGTM.

[tensorrt-cicd]

PR_Github #5006 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #5006 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #3639 completed with status: 'FAILURE'

[yilin-void]

/bot run

[tensorrt-cicd]

PR_Github #5119 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #5119 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #3731 completed with status: 'FAILURE'

[yilin-void]

/bot run

[tensorrt-cicd]

PR_Github #5319 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #5319 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #3879 completed with status: 'FAILURE'

[yilin-void]

/bot run

[tensorrt-cicd]

PR_Github #5676 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #5676 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #4147 completed with status: 'SUCCESS'