mlcommons
Information redundancy
There is too much information from pytorch ET plus, including function call relationships with too many details, but in fact we only focus on computing, memory access and communication. Does such a large amount of information run counter to the design philosophy of chakra?
Thank you for raising a concern, @XcodeRole. Could you please specify which information appears redundant? The Chakra working group welcomes any suggestions for improved design and engineering. If you are referring to the function call relationships, this information is crucial for one of our key downstream use cases, namely simulation. We must maintain the function call relationships to replay operators in a simulator.
In Chakra ET, nodes are represented on a per-function-call basis. Many functions have an execution time of 0 microseconds because they do not perform computations themselves; instead, they invoke computational functions. However, these still occupy a node in ET, leading to a very large and redundant call graph.
Yes, it makes sense if that is the case. Do you have any example traces that exhibit such behaviors? I had an opportunity to collect LLAMA traces, but I did not observe any 0-microsecond operators. This might be due to rounding, but I believe the measurements are accurate because I have validated them, as shown in the following PR.
- https://github.com/mlcommons/chakra/pull/47
You responded very quickly, which was a pleasant surprise. However, the debug code you recently merged is inconsistent with the version I am using. There were some bugs in the version I was working with, and I had made my own modifications to fix them. Therefore, it will take me some time to rerun tests to ensure the changes are compatible with the updated code.