tensorflow-upstream VS rocm-build

I tried to train another model using InceptionResNetV2 and the same issues happens. Also, this happens even using the model.predict() method if using the GPU. Probably this is an issue related to the AMD Radeon RX 6700 XT or some mine misconfiguration. System Inormation: ArchLinux 6.1.32-1-lts - AMD Radeon RX 6700 XT - gfx1031 Opened issues: - https://github.com/RadeonOpenCompute/ROCm/issues/2250 - https://github.com/ROCmSoftwarePlatform/tensorflow-upstream/issues/2125

AMD has had their ROCM Tensorflow port for quite a while now: https://github.com/ROCmSoftwarePlatform/tensorflow-upstream and it works pretty well on my RX 6800; unfortunately it's only compatible with Linux as the ROCm stack is built on top of the Linux kernel I believe

I mean, tensorflow has a fork with ROCm support which is maintained by AMD https://github.com/ROCmSoftwarePlatform/tensorflow-upstream although I'm not entirely sure what you're AI workloads are specifically, I'm just throwing out tensorflow because it's popular. On the enterprise side they also have radeon instinct MI, although I assume you're probably not using enterprise HW but I wanted to throw it out there anyway.

It might work, but will likely require building pytorch/tensorflow from source. I compiled both with ROCm 5.0 today, Pytorch is painless but TF required using this repo maintained by AMD instead

I saw this https://github.com/ROCmSoftwarePlatform/tensorflow-upstream and I thought it could be a way. Do you think it is not stable and ready to be used ?

The repo exists: https://github.com/ROCmSoftwarePlatform/tensorflow-upstream

https://github.com/ROCmSoftwarePlatform/tensorflow-upstream#tensorflow-rocm-port

I tried some case ROCm & Tensorflow-rocm as is on this page( tensorflow-upstream/tensorflow-rocm-release.md at develop-upstream · ROCmSoftwarePlatform/tensorflow-upstream · GitHub ), but I failed to run a simple CNN model with fashion-mnist datasets.

Scripts to help with building rocm and hip. It will also help work out dependencies. You will need to modify the scripts for them to work and not all are required. https://github.com/xuhuisheng/rocm-build

Short answer no. Long answer "in theory" yes. I tried this [1] but gave up as building rocm + deps takes up to 6h :/ Official statement [2]

[1] https://github.com/xuhuisheng/rocm-build

I am in the same boat with a gfx03 card. What patch did you use? The ones here? https://github.com/xuhuisheng/rocm-build

I also tried to compile pytorch with its Vulkan backend, but ended throwing the towel as LDFLAGS are a mess to get right (I successfully compiled it, but that was only part of the build chain, and decided I had better things to spend time on). I wonder how that would perform; ncnn works pretty decently.

Unofficial black magic is available: https://github.com/xuhuisheng/rocm-build/tree/master/navi10 (pytorch 1.12.0 is outdated but can run SD)

also have an RX570, currently running latest Tensorflow and ROCm 4.1. had to recompile some parts of ROCm 4.1 libraries to get tensorflow to work. mostly followed this guide: https://github.com/xuhuisheng/rocm-build/tree/master/gfx803