Issue in Code, at 114:24: def _fwd_kernel(..)
#1
by
Hosna
- opened
If we run the same example in the ReadMe, the code gives some errors:
assertion error: q.cuda--> so we change the model and input tocuda- If error 1 is solved, then it gives another error of not identifying an attribute
trans_b is an invalid argument - If we even remove this argument, we end up in an error of
incompatible dot product multiplicationwhere q & k for example are of the same shapes while we need transpose. :AssertionError: assert lhs.shape[1].value == rhs.shape[0].value
If anyone knows how to solve this,
Hosna
changed discussion title from
Issue in Code
to Issue in Code, at 114:24:
def _fwd_kernel(
Hosna
changed discussion title from
Issue in Code, at 114:24:
def _fwd_kernel(
to Issue in Code, at 114:24:
def _fwd_kernel(..)
Hey,
Can you please try to install triton from source?
git clone https://github.com/openai/triton.git;
cd triton/python;
pip install cmake; # build-time dependency
pip install -e .
Hi there!
I am facing the same issue and I am not able to install triton from source on my server. Platform and torch restrictions.
Can we find another way?
After installing triton from source i get a different error when loading model.
AttributeError: module 'triton' has no attribute 'autotune'
Any updates? I am facing the same issue as @amberzqt .