Having a proper support for triangular upper when the inputs are a list of CallNodes will allow us to be a step closer to successfully implementing Qwen 1.5 (0.5B) (See https://github.com/tenstorrent/tt-buda-demos/issues/20).
Is this a genuine issue??? Or can it be ignored for now?
2. NaN tensor values for Grayskull e75
When tested on @marty1885's e75, he ran into an error where his tensor values were NaN.
But weirdly @JonathanALevine's e150 was able to successfully compile and run it until running into some errors later.
*This is a draft for now since this is just a workaround and not a proper fix yet.
This PR is to address https://github.com/tenstorrent/tt-tvm/issues/3
Having a proper support for triangular upper when the inputs are a list of
CallNode
s will allow us to be a step closer to successfully implementing Qwen 1.5 (0.5B) (See https://github.com/tenstorrent/tt-buda-demos/issues/20).Explanation
When compiling Qwen 1.5 (0.5B) (https://github.com/tenstorrent/tt-buda-demos/pull/37), one of its OP codes is
aten::triu
with its inputs containing nested functions of OP calls.self.trilu
seems to be able to successfully handle these inputs whenmode="upper"
to do triangular upper operation.Issues
1
Op(trilu)
instead ofOp(triu)
After doing
self.trilu(inputs, input_types, mode="triu")
, the resulting output would be:Is this a genuine issue??? Or can it be ignored for now?
2. NaN tensor values for Grayskull e75
When tested on @marty1885's e75, he ran into an error where his tensor values were NaN. But weirdly @JonathanALevine's e150 was able to successfully compile and run it until running into some errors later.
*This is a draft for now since this is just a workaround and not a proper fix yet.