Open parmesant opened 3 months ago
Try flash-attn 2.5.1 on nvcr 23.12 or 24.01.
The symbol "_ZN2at4_ops15sum_IntList_out4callERKNS_6TensorEN3c1016OptionalArrayRefIlEEbSt8optionalINS510ScalarTypeEERS2" is a mangled CPP function name, to demangle it use this Demangler tool
the function is
at::_ops::sum_IntList_out::call(at::Tensor const&, c10::OptionalArrayRef<long>, bool, std::optional<c10::ScalarType>, at::Tensor&)
So what happens is that flash-attn built with a pytorch version that does not align, don't know which version should..
@tsvisab, not sure if you've resolved this but for anyone who might come across this but installing from source solved it for me.
GPU: 2x RTX 4090 Memory: 128GB CPU: 64 cores CUDA: 12.3.52 NVIDIA Driver: 545.23.08 PyTorch Container: 23.11 nvcc --version (on host machine):
Cuda compilation tools, release 12.3, V12.3.107 Build cuda_12.3.r12.3/compiler.33567101_0
I am trying to fine-tune mistral 7b instruct v0.2 and am running into these errors-
First, I run into an error due to
protobuf
(I have 4.24.4 and the error message suggests downgrading it to 3.20.x or lower) (Upon downgrading, pip gives this errorcudf 23.10.0 requires protobuf<5,>=4.21, but you have protobuf 3.20.3 which is incompatible.
but things continue to work)After downgrading to 3.20.3, I run into an issue with
flash-attn
After upgrading
flash-attn
to 2.3.6 (latest at the time of the pytorch container release), I get this pip warningERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts. transformer-engine 1.0.0+66d91d5 requires flash-attn<=2.0.4,>=1.0.6, but you have flash-attn 2.3.6 which is incompatible.
and this error-Even if I install the latest version 2.5.2 (latest at the time of writing this), I get a similar error-
Some other things I've tried-
transformer-engine