Fix ROCm torch install: use find-links, not --index-url (devcontainer, README, hip_utils)

[demandal25]

Summary

Fix ROCm torch installation across the repo: point pip at the radeon wheel mirror with -f/--find-links instead of --index-url, which fails with No matching distribution found for torch==... (from versions: none). Started as a devcontainer build fix and extended to the user-facing install docs and error messages that carried the same broken command.

What changed

• .devcontainer/rocm/Dockerfile — install torch with -f https://repo.radeon.com/rocm/manylinux/rocm-rel-${ROCM_VERSION}/ instead of --index-url; add a post-install assertion that fails the build if the installed torch is not a ROCm/HIP build (torch.version.hip is None); document the rationale in column-1 comments above the RUN.
• README.md — switch the two radeon torch install commands to -f, and rewrite the NOTE that previously recommended --index-url so it reflects reality (why --index-url fails, why -f still selects the ROCm wheel, and how to verify).
• flashinfer/hip_utils.py — switch the radeon torch commands in the "PyTorch has no ROCm support" error and the ROCm-version-mismatch warning (pip and uv variants) to -f.

Unrelated --index-url https://pypi.amd.com/simple/ lines (a real PEP 503 index) are intentionally left unchanged.

Why

repo.radeon.com/rocm/manylinux/rocm-rel-X.Y/ is a flat directory listing of wheels, not a PEP 503 simple index — its per-package path (.../torch/) returns 404. --index-url makes pip request that path, so it finds zero candidates and fails. -f/--find-links reads the flat listing directly, matching what docker/Dockerfile.rocm_ci already does.

Switching to -f still resolves to AMD's internal ROCm build rather than PyPI's CUDA wheel. The radeon wheel carries a local version (torch-2.9.1+rocm7.2.0.lw.git7e1940d4-cp312-cp312-linux_x86_64.whl), and PEP 440 ranks 2.9.1+rocm... strictly above PyPI's plain 2.9.1, so pip selects it even though both satisfy ==2.9.1. Confirmed with pip install --dry-run (resolved to the +rocm7.2.0 wheel). The devcontainer additionally guards this at build time with the HIP assertion, and the README documents a one-line check for manual installs.

Test plan

• Built the image end to end: docker build --build-arg USER_UID=$(id -u) --build-arg USER_GID=$(id -g) --build-arg USERNAME=demandal -t dm-fi-rocm72-torch291-py312 -f .devcontainer/rocm/Dockerfile . — succeeds, and the HIP assertion printed torch 2.9.1+rocm7.2.0.git7e1940d4 (hip=7.2.26015-fc0010cf6a) before the build continued (proving the && chain and the check both run).
• Verified torch inside the image is the ROCm build (torch.version.hip set, torch.version.cuda None), not a PyPI CUDA wheel.
• README and hip_utils.py changes are text/guidance only; the repo's required pre-commit check runs on this PR in CI.

[Copilot]

Pull request overview

Updates the ROCm devcontainer image build to install the AMD ROCm PyTorch wheel from the Radeon wheel mirror correctly, by treating the mirror as a flat wheel directory rather than a PEP 503 “simple” index.

Changes:

• Switch torch installation in the ROCm devcontainer from --index-url to -f/--find-links pointing at the Radeon ROCm wheel directory.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated 1 comment.

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated no new comments.

Comments suppressed due to low confidence (2)

.devcontainer/rocm/Dockerfile:90

• Now that the devcontainer uses -f https://repo.radeon.com/.../rocm-rel-${ROCM_VERSION}/ (because the URL is not a PEP 503 simple index), repo documentation and user-facing guidance that still recommends --index-url will likely continue to fail the same way (e.g., README.md and the flashinfer/hip_utils.py error message both point users at --index-url). Please update those references (or point at a true /simple index if one exists) so users don’t hit the same torch install failure outside the devcontainer.

    /bin/micromamba run -n ${MAMBA_ENV_NAME} pip install --no-cache-dir torch==${TORCH_VERSION} -f https://repo.radeon.com/rocm/manylinux/rocm-rel-${ROCM_VERSION}/ && \
    /bin/micromamba run -n ${MAMBA_ENV_NAME} python -c "import torch, sys; hip = torch.version.hip; print(f'torch {torch.__version__} (hip={hip})'); sys.exit('ERROR: installed torch has no ROCm/HIP build (hip is None)' if hip is None else 0)" && \

.devcontainer/rocm/Dockerfile:86

• The Dockerfile switches to -f/--find-links, but the in-file comment only explains the HIP build assertion; it doesn't record why --index-url is intentionally avoided here. Given the repo has multiple places that recommend --index-url, capturing the PEP 503/flat-directory rationale inline will help prevent accidental reverts and future confusion.

This issue also appears on line 89 of the same file.

# Create a new micromamba env and install needed packages for flashinfer development.
# After installing torch, assert it is a ROCm/HIP build (torch.version.hip is not
# None) so the build fails fast if -f/--find-links ever resolves a PyPI CPU/CUDA
# wheel instead of the radeon ROCm wheel.

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated no new comments.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated no new comments.

Comments suppressed due to low confidence (3)

flashinfer/hip_utils.py:567

• Same version-format issue as above: the FLASHINFER_ROCM_VERSION export and the uv/pip URL are built from {system_rocm} (often X.Y.Z), but the radeon directory naming is rocm-rel-X.Y. This guidance should use the already-computed system_rocm_major_minor.

                f"  export FLASHINFER_ROCM_VERSION={system_rocm}\n"
                f"  uv pip install torch==2.7.1 -f "
                f"https://repo.radeon.com/rocm/manylinux/rocm-rel-{system_rocm}/\n"

README.md:151

• The PR description says the broken --index-url https://repo.radeon.com/rocm/manylinux/... usage was fixed “across the repo”, but there are still examples that recommend --index-url for the radeon wheel directory (e.g., CLAUDE.md:42 and pyproject.toml:19, plus amd-flashinfer-jit-cache/pyproject.toml:4,33). This leaves conflicting installation guidance within the same repo.

**NOTE:** The radeon repo is a flat wheel listing, not a PEP 503 index, so
`--index-url` fails with "No matching distribution found" — use `-f`
(`--find-links`). pip still prefers the ROCm wheel over a same-version PyPI
wheel because its `+rocm<X.Y>` local version ranks higher. Confirm the result
with `python -c "import torch; assert torch.version.hip, 'not a ROCm build'"`.

flashinfer/hip_utils.py:563

• system_rocm is typically a 3-segment version (e.g., 7.2.0 from /opt/rocm/.info/version), but the radeon wheel directories are major.minor (e.g., rocm-rel-7.2). The warning currently builds a URL with {system_rocm}, which can point users to a non-existent directory.

This issue also appears on line 565 of the same file.

                f"  pip install torch==2.7.1 -f "
                f"https://repo.radeon.com/rocm/manylinux/rocm-rel-{system_rocm}/\n\n"