tenstorrent / tt-forge-fe

The TT-Forge FE is a graph compiler designed to optimize and transform computational graphs for deep learning models, enhancing their performance and efficiency.
https://docs.tenstorrent.com/tt-forge-fe/
Apache License 2.0
19 stars 3 forks source link

Llama 3.2 1B - Add specific op tests #533

Open nvukobratTT opened 1 week ago

nvukobratTT commented 1 week ago

Summary

To confirm that each op individually works for the Llama model, let's add a set of op tests that are specific to this variation of the Llama model.

To do so, here are a few hints:

  1. Print unique ops appearing in Llama.
    • To do so, use this feature
  2. For each unique op (and related input shape, attribute, etc.) create an e2e op test
    • The recommendation is to keep these model-related op tests as part of the model directory. E.g. forge/test/mlir/llama/ops/*
  3. Make sure to run e2e tests as further as possible
    • To make this possible, you can use compile depth (e.g. compiler_cfg.compile_depth = forge.CompileDepth.SPLIT_GRAPH)
mstojkovicTT commented 16 hours ago

Additional info for each of the ops that doesnt work: