Llama: Custom Profiler

[kurtis-b-1]

Custom profiler for performance measurements of Llama 3.2-1B inference. A logging function is executed at the call and return of each function. Initially tried with cProfile, but the measurements of PyXRT execution was missing (possibly cProfile excludes I/O like the NPU). This method allows for deterministic profiling of inference, though profiling overhead is higher compared to statistical profiler. But this method is apt for getting a comprehensive picture of the inference run, as one can check the time of each AIE and CPU operator relative to the inference metrics such as TTFT. The overhead of profiling shouldn't skew the relative contribution of each operator to the latency of inference.

This branch is based on the branch in this PR: #8

Added

• Script for generating CSV based on log

Changed

• Attach custom profiler to system during inference if a CLI argument is passed

Removed

PR Merge Checklist

1. The PR is rebased on the latest devel commit and pointing to devel.
2. Your PR reviewed and approved.
3. All checks are passing.