Using without nvidia libraries or how to make it work with them using ZLUDA?

[test3211234]

I'm trying to use ZLUDA with paddlepaddle-gpu to make it run on an AMD GPU, and while PyTorch/Transformers/Diffusers all work, paddlepaddle-gpu doesn't work because it uses nvidia. These have different dlls, that are not included with ZLUDA, like having not only cublas but also nvblas which isn't in ZLUDA, so I get an error. Is there a workaround? I need the speed of GPU.

lshqqytiger/ZLUDA#55 (comment)

[GreatV]

To use ZLUDA with paddlepaddle-gpu on an AMD GPU, you face the challenge that PaddlePaddle's GPU implementation heavily relies on NVIDIA-specific libraries such as cublas and nvblas, which are not compatible with ZLUDA.

Here is a detailed breakdown of potential workarounds for your situation:

---

1. ZLUDA Limitations and Missing NVIDIA Libraries

• ZLUDA does not include a full implementation of NVIDIA's proprietary libraries like nvblas or cublas. These libraries are critical for PaddlePaddle's GPU operations.
• While PyTorch/Transformers/Diffusers may work with ZLUDA because they either have better fallback mechanisms or use custom operations, PaddlePaddle's reliance on NVIDIA libraries makes it inherently incompatible without further modifications.

---

2. Possible Workarounds

Below are some approaches you can try:

a. Use CPU Mode for PaddlePaddle

If GPU acceleration is unavailable due to hardware or library constraints, one option is to use the CPU version of PaddlePaddle (paddlepaddle instead of paddlepaddle-gpu). While this sacrifices speed, it ensures compatibility.

Steps:

1. Uninstall paddlepaddle-gpu:

   pip uninstall paddlepaddle-gpu

2. Install the CPU version:

   pip install paddlepaddle

b. Explore ROCm (AMD GPU Support)

AMD provides the ROCm (Radeon Open Compute) platform for GPU computation, which includes libraries similar to NVIDIA's CUDA and cuBLAS. However, PaddlePaddle does not natively support ROCm at the moment, so this solution would require significant development effort.

Steps (if you want to explore custom development):

1. Investigate whether the PaddlePaddle source code can be modified to use ROCm libraries (like rocBLAS) instead of cuBLAS.
2. This approach would require you to build PaddlePaddle from source and implement AMD GPU support manually.

c. Wait for ZLUDA Updates

If ZLUDA extends support for AMD GPUs in the future or implements replacements for missing NVIDIA libraries, it may eventually become possible to use PaddlePaddle with AMD GPUs via ZLUDA. Monitor the ZLUDA GitHub repository for updates.

d. Use Alternative Frameworks

If PaddlePaddle is not a strict requirement, consider using frameworks like PyTorch or TensorFlow, which have broader support for heterogeneous hardware. PyTorch, for example, has native ROCm support for AMD GPUs.

---

3. Error Reference

The linked GitHub Discussion mentions missing libcublasLt.so.12, which directly points to the absence of NVIDIA's cuBLAS library on your system. Since ZLUDA cannot provide a substitute for libcublasLt.so.12, this error is expected when trying to run NVIDIA-dependent code on non-NVIDIA hardware.

---

4. Conclusion

At this time, there is no straightforward way to make paddlepaddle-gpu work with ZLUDA on an AMD GPU due to the reliance on NVIDIA-specific libraries. Your best options are:

• Switch to the CPU version of PaddlePaddle.
• Explore ROCm and modify PaddlePaddle for AMD GPU support.
• Use alternative frameworks with better AMD GPU compatibility, such as PyTorch.

If you need GPU acceleration immediately and have access to an NVIDIA GPU, using that hardware with paddlepaddle-gpu would be the simplest solution.

---

Response generated by feifei-bot | chatgpt-4o-latest

[test3211234]

That's sad, I think you should consider supporting this, it could bring in many many more users and would be very useful. This is without a doubt the best OCR library and I really need to use it with GPU.