fix broken cuda and rocm images

[baptistecolle]

AutoAWQ change its setup.py, this broke the building of our docker image

[IlyasMoutawwakil]

can you fix the same TORCH_VERSION issue in rocm image as well (we build the same packages there)

[baptistecolle]

done, rocm is fixed

[IlyasMoutawwakil]

I'm not sure about that (see failing PyTorch tests), I think https://github.com/casper-hansen/AutoAWQ/blob/main/setup.py#L10 is triggering the installation of ipex https://github.com/huggingface/optimum-benchmark/actions/runs/10908257743/job/30273811027?pr=263#step:5:129

[baptistecolle]

Now, everything concerning the broken Docker image due to an update with AutoAWQ should have been fixed.

There are still some failing tests in our pipeline at of today:
(because of #265 - issue with the codecarbon lock)

• API CPU Tests / run_api_cpu_tests (pull_request)
• API CUDA Tests / run_api_cuda_tests (pull_request)
  (issue with the use of the deprecated method - is_torch_tpu_available Fix is_torch_tpu_available in ORT Trainer optimum#2028, so we need to wait for next release of optimum for the fix)
• CLI CUDA Torch-ORT Multi-GPU Tests / run_cli_cuda_torch_ort_multi_gpu_tests
• CLI CUDA Torch-ORT Single-GPU Tests / run_cli_cuda_torch_ort_single_gpu_tests