• fixed warmup generation loop to complete the generation of tokens instead of generating just one token
• removed redundant generator object initialization in the benchmark loop. This improves (reduces) memory utilization
• fixed things related to printing memory usage

Issue: If multiple prompt lengths are used in benchmark arg, the script would print the same memory usage for all cases.
Solution: Need to reset monitoring and peak_gpu_memory variables in each benchmark loop and print at the end of the loop along with other metrics