build(base): bump ROCm baseline to 7.13

[MioYuuIH]

Summary

• Bump the ROCm GPU base image baseline to ROCm 7.13.0 and the matching ROCm-enabled PyTorch wheel set.
• Add the gfx1152 build target and keep generic gfx110x / gfx120x image targets supported.
• Split ROCm SDK target selection from PyTorch wheel bucket selection so apt packages and wheel indexes can use their own target names.
• Refresh ROCm host setup docs and ansible installer references for the ROCm 7.13 flow.

Implementation Notes

• ROCm is installed from AMD's Ubuntu 24.04 apt repository using target-specific amdrocm-core-sdk7.13-<target> packages.
• Generic CI targets use gfx110x / gfx120x for the ROCm SDK packages and gfx110X-all / gfx120X-all for PyTorch wheel buckets.
• The /opt/rocm symlink fallback is idempotent so it does not fail when the 7.13 packages already create those links through update-alternatives.
• Local Makefile builds default ROCM_SDK_TARGET to GPU_TARGET while CI passes explicit values from .github/build-config.json.

Validation

• jq empty .github/build-config.json
• YAML parse for .github/workflows/docker-build.yml
• python3 -m unittest tests.installer.test_gpu tests.installer.test_overlay
• Local Docker build for gfx110x with ROCM_SDK_TARGET=gfx110x and PYTORCH_WHL_TARGET=gfx110X-all
• Local image smoke test: /opt/rocm/bin/hipcc exists and torch.__version__ is 2.9.1+rocm7.13.0
• GitHub Actions Build Docker Images workflow dispatch run 26497377590 succeeded for base-gpu / all

[KerwinTsaiii]

Hi @MioYuuIH should we also update Pytorch version to the one with ROCm 7.13.0

python -m pip install --index-url https://repo.amd.com/rocm/whl/gfx1151/ \
    "torch==2.11.0+rocm7.13.0" \
    "torchvision==0.26.0+rocm7.13.0" \
    "torchaudio==2.11.0+rocm7.13.0"

[MioYuuIH]

Hi @MioYuuIH should we also update Pytorch version to the one with ROCm 7.13.0

python -m pip install --index-url https://repo.amd.com/rocm/whl/gfx1151/ \
    "torch==2.11.0+rocm7.13.0" \
    "torchvision==0.26.0+rocm7.13.0" \
    "torchaudio==2.11.0+rocm7.13.0"

aup-learning-cloud/dockerfiles/Base/Dockerfile.rocm

Lines 125 to 149 in 3013096

	# CI passes PYTORCH_WHL_TARGET explicitly via .github/build-config.json. For
	# local builds driven by dockerfiles/Makefile or auplc-installer — which only
	# set GPU_TARGET — we derive the wheel target here so ad-hoc
	# `docker build --build-arg GPU_TARGET=gfx120x` and `make base-rocm
	# GPU_TARGET=gfx120x` both Just Work.
	RUN WHL_TARGET="${PYTORCH_WHL_TARGET}" && \
	if [ -z "${WHL_TARGET}" ]; then \
	case "${GPU_TARGET}" in \
	gfx110x) WHL_TARGET="gfx110X-all" ;; \
	gfx120x) WHL_TARGET="gfx120X-all" ;; \
	*) WHL_TARGET="${GPU_TARGET}" ;; \
	esac; \
	fi && \
	INDEX_URL="${PYTORCH_INDEX_URL:-https://repo.amd.com/rocm/whl/${WHL_TARGET}/}" && \
	TORCH_SUFFIX="+rocm${ROCM_VERSION}" && \
	echo "Installing ROCm PyTorch from ${INDEX_URL} (GPU_TARGET=${GPU_TARGET}, WHL_TARGET=${WHL_TARGET})" && \
	python3 -m pip install --no-cache-dir \
	--index-url "${INDEX_URL}" \
	"torch==${PYTORCH_VERSION}${TORCH_SUFFIX}" \
	"torchvision==${TORCHVISION_VERSION}${TORCH_SUFFIX}" \
	"torchaudio==${TORCHAUDIO_VERSION}${TORCH_SUFFIX}" && \
	SDK_INC=$(python3 -c "import _rocm_sdk_core, os; print(os.path.join(os.path.dirname(_rocm_sdk_core.__file__), 'include'))" 2>/dev/null) && \
	if [ -n "$SDK_INC" ] && [ ! -e /opt/rocm/include/hip ]; then \
	ln -s "${SDK_INC}/hip" /opt/rocm/include/hip; \
	fi

We are not hardcoding the PyTorch wheel URL or ROCm target here.
The Dockerfile derives the wheel index from PYTORCH_WHL_TARGET / GPU_TARGET, and CI passes the target-specific values from .github/build-config.json. The package versions have also been updated to use the ROCm 7.13.0 wheel suffix:

• torch==2.9.1+rocm7.13.0
• torchvision==0.24.0+rocm7.13.0
• torchaudio==2.9.0+rocm7.13.0
  So the install path is target-aware, while the framework versions stay aligned with the current course baseline.

[KerwinTsaiii]

@MioYuuIH Please resolve conflicts