Queries regarding using astra-sim

[Annanya]

Hi Folks,

I am part of the Tesla AI team and we are working on simulating language models on custom hardware and topology. I came across your project and think that it can be really useful for us but had some doubts:

1. Would it be possible to generate trace files for the simulation that is done to get a better understanding how exactly did the simulation run? Like some kind of logging of events on a single node.  
2. Does the project support pipelining between compute ops and collectives.
3. Also how easy is it to generate a chakra trace from some arbitrary dependency graph.

Thanks
Annanya      

[Joongun Park]

Hi Annanya,

Apologies for the delayed response.

1. The current public version of ASTRA-Sim consumes traces in a graph format but does not generate them. However, with some code modifications, trace generation should be possible.

2. Yes—since ASTRA-Sim consumes a dependency graph, compute and collective operations without dependencies will be pipelined or executed concurrently.

3. There are a few ways to generate dependency graphs:

   • Manually from structured text

   • Synthesizing LLM traces with arbitrary parallelism

   • Collecting traces from actual model executions

You can find demos and an overview from the MICRO 2024 tutorial here: https://astra-sim.github.io/tutorials/micro-2024.

Thank you for your interest!

Best regards,
Joongun

2025년 6월 13일 (금) 오후 5:54, Annanya <anna...@gmail.com>님이 작성:

--
You received this message because you are subscribed to the Google Groups "ASTRA-sim Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to astrasim-user...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/astrasim-users/CAFWoR%3DUf-%3DpdfMsM2HRobQqcBBqqToUZMKxORR%3DxgPWah7bFsA%40mail.gmail.com.
For more options, visit https://groups.google.com/d/optout.