[chensuyue]

This pull request updates the logic in the benchmark/lm_eval/run_lm_eval.sh script to improve how attention backend and quantization settings are configured for different GPU architectures and evaluation formats. The changes make the script more robust and flexible when running benchmarks on various GPUs, especially in handling FP8 and QDQ plugin scenarios.

Improvements to attention backend and quantization logic:

• The script now detects the GPU compute capability and applies the QDQ plugin and Marlin optimizations only on GPUs other than SM 10.x, moving this logic earlier and making it more robust. [1] [2]
• The attention backend (ATTN_BACKEND) is now resolved based on whether FP8 is used for attention or KV cache, with clearer precedence and configuration.
• The script ensures that the correct environment variables and CLI arguments are set for the chosen backend, handling both TRITON_ATTN and FLASHINFER cases, and only quantizes Q when attention runs in FP8.
• Redundant and misplaced code for QDQ plugin activation was removed from the end of the script, as this logic is now handled earlier.