docs: Feedback for “Benchmarking NVIDIA TensorRT-LLM”

[Bazza-63]

Would it be possible for you to compare TensorRT-LLM against MLC LLM with the TVM Unity compiler?

For me when using a custom, patched ROCm along with the appropriate flags turned on whilst compiling TVM Unity. I managed 120 tokens per second on Qwen 2.5 8B Q8_0. This was with an RX 7800 XT too.

If you're wondering where I got the custom ROCm from, you can find it here:

• https://github.com/lamikr/rocm_sdk_builder