You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I've attached the llama prefill model given from @dgolubovicTT. We have mechanisms in MLIR itself to run golden comparisons if the golden data exists. Part of the bringup is being able to generate the golden intermediates/output for each of the ttir ops in the llama model.
I'm requesting for there to be some support to dump the goldens as .pt tensors for each op in the ttir graph. You can also save the Goldens in a golden_map in the flatbuffer itself, as the support for that already exists (second way is preferred). There should be a way to correlate the dumped tensor to the ttir op it came from. I will then take the Goldens and the ttir graph and run it in our existing infra in MLIR. Ultimately, this will help in the future once we integrate golden into tt-forge itself.
Thanks @tapspatel! We're planing to fix up and enable our intermediate verifications in next period. We'll include this as additional set of requirements
llama_prefill.mlir.txt
I've attached the llama prefill model given from @dgolubovicTT. We have mechanisms in MLIR itself to run golden comparisons if the golden data exists. Part of the bringup is being able to generate the golden intermediates/output for each of the ttir ops in the llama model.
I'm requesting for there to be some support to dump the goldens as .pt tensors for each op in the ttir graph. You can also save the Goldens in a golden_map in the flatbuffer itself, as the support for that already exists (second way is preferred). There should be a way to correlate the dumped tensor to the ttir op it came from. I will then take the Goldens and the ttir graph and run it in our existing infra in MLIR. Ultimately, this will help in the future once we integrate golden into tt-forge itself.
Relevant issue: tenstorrent/tt-mlir#1768 (stage 4)
The text was updated successfully, but these errors were encountered: