Arlo Phoenix
Arlo Phoenix
If you use the official docker images/have hipBlasLt installed, you could try my fork [here](https://github.com/arlo-phoenix/bitsandbytes-rocm-5.6). Also probably not the best place to ask this, there are several user ROCm ports,...
> @arlo-phoenix Could you make a PR? I'd love to have this working/maintained in this repo. @TimDettmers is there a reason not to accept this? I'd love that as well....
> If it can't be easily overcome, maybe just disable 4Bit support and throw a exception when it is used on ROCm. Then someone can just maintain a fork that...
> What do you think about this approach? @rickardp, @abhilash1910, and @arlo-phoenix: this is also relevant for your integration. Please give us feedback on the `extern C` approach so we...
The Adam8bit optimizer issue probably wasn't from ROCm itself, but the general issue that's been fixed in the latest version 0.41.1. I made my own [fork](https://github.com/arlo-phoenix/bitsandbytes-rocm-5.6) that just uses defines...
> Hi @arlo-phoenix , nice fork, what would i comment out of the Makefile to test 6900 XT (gfx1030)? > > I have already "export HSA_OVERRIDE_GFX_VERSION=10.3.0" You shouldn't have to...
All header should be found under $ROCM_HOME/include. To control where it searches for the libraries and headers you could try changing these two ```bash HIP_INCLUDE := -I $(ROCM_HOME)/include -I $(ROOT_DIR)/csrc...
> I am able to build arlo-phoenix's fork of bitandbytes using the docker image rocm/pytorch. The nightly version of the image gives errors in make hip though. Ok that's weird......
Should work now. Atleast the tests from bitsandbytes for nf4 succeeded. See the [README](https://github.com/arlo-phoenix/bitsandbytes-rocm-5.6/blob/main/README.md) in the fork for updated install instructions (I improved the Makefile to take the ROCM_TARGET as...
> If there is a BLAS or even a matrix multiplication API available, it should be easy to add. There's [Vitis BLAS](https://www.xilinx.com/products/design-tools/vitis/vitis-libraries/vitis-blas.html), but no idea if it actually works. It's...