Edit: this dump was updated with this PR: #1726. Running ttrt run out.ttnn --memory --save-artifacts will output memory report file.

@tapspatel This would probably be a good introductory issue for anyone onboarding onto the explorer project. I can create a more detailed roadmap if anyone new is going to take up this issue.

@vprajapati-tt we discussed this briefly offline. I think that there is a very detailed memory report in ttnn-visualizer and that we might not need to recreate this feature in explorer.

Using ttnn-visualizer reports from ttrt execution is going to be a requirement by the perf team, so i'd like to test this flow out first and if it's not sufficient we should discuss adding this to explorer. Opened an issue here #2110

Yeah, for detailed memory reports I definitely suggest using ttnn-visualizer. That being said, this visualization provides a single attribute to be viewed after execution for a holistic view of DRAM and L1 usage (as a percentage). I agree, let's test this out and see what the flow looks like.

@vprajapati-tt Yes, i was referencing a more detailed view of the memory. Having a single percentage is fine now.

@jgrimTT Here's a plan to action for taking this on as a first issue!

1. Visualizing the holistic memory usage for DRAM and L1 requires polling for the information first. This will be done through the execution command on tt_adapter. You will need to find where the "3 files produced" are stored relative to the ttrt_artifacts directory and create some helper function in tt_adapter.runner.ModelRunner that get_memory_usage given a model_path.
2. Once execution is finished, a convert call will be placed. Try and mirror how get_perf_trace gets called in tt_adapter:main:TTAdapter:convert and passed into the build_graph function.
3. In tt_adapter:mlir:build_graph you can then use the given memory_usage data to create a new attribute for each op based on the location in the memory_usage that contains a percentage of MEMORY_USED/TOTAL_MEMORY. Do this for both L1 and DRAM.
4. The frontend has an unique visualizer for these percentages (Feature: memory info model-explorer#23). Enable the memory visualization on the percentage using add_to_dataclass in tt_adapter.util.

This should provide a brief overview of where to look in the code but I may have omitted some more detailed semantics. Feel free to ping for help, but this should be the rough roadmap to implement this feature :)