[yhl-amd]

Summary

Adds .github/workflows/lmcache-rocm-wheel.yaml, a manually dispatched workflow that builds an LMCache wheel for ATOM's image ABI (ROCm 7.2.4 / torch 2.10.0, gfx942 + gfx950) from any LMCache commit and can publish it as a pinned GitHub Release on this repository.

Why

ATOM's image ships AMD's torch 2.10.0+rocm7.2.4.git3d3aa833. LMCache publishes a wheel for that ABI only with tagged releases (v*-rocm-torch210). Its rolling nightly-rocm wheel is built against torch 2.11 for the upstream vLLM image. I installed today's nightly (0.5.6.dev98+rocm7.2, built from dev@05fc77a) in an ATOM image, and its native extension fails to load:

ImportError: .../lmcache/cuda_ops.cpython-312-x86_64-linux-gnu.so: undefined symbol: _ZN3c104cuda20CUDACachingAllocator9allocatorE

So LMCache changes that ATOM depends on, such as LMCache/LMCache#5132 for #2250, cannot be used until LMCache's next tag. The nightly release is also deleted and republished every day, so its URL cannot be pinned.

How it works

• Input: lmcache_commit must be a full 40-character SHA; branch names are rejected. publish defaults to off, which builds and validates only. runner is build-only-atom by default; linux-atom-mi35x-1 additionally runs the GPU kernel check.
• Build: runs LMCache's own .github/scripts/build_rocm_torch210_wheel.sh from the requested commit, inside the same digest-pinned rocm/pytorch:rocm7.2.4_..._pytorch_release_2.10.0 image that LMCache's build_rocm_torch210_artifacts.yml uses. The build steps are not duplicated in ATOM.
• Version: follows LMCache's nightly numbering (setuptools-scm guess-next-dev from the last stable tag), with the commit and ABI in the local segment, e.g. 0.5.6.dev98+g05fc77a0.rocm7.2.4.torch2.10.git3d3aa833.cxx11abi1.
• Validation in a clean container of the build image:
  • LMCache's validate_rocm_torch210_wheel.py checks the exact torch/HIP/ROCm/CXX11 ABI and the wheel tag. On a GPU runner it also runs the HIP D2H/H2D kernels.
  • New .github/scripts/validate_lmcache_wheel_for_atom.py requires the ATOM MP adapters, execute_object_group_transfer, and the native .so modules. It also parses --null-block-id -1 --separate-object-groups through the MP server's own parser.
• Publish runs in a separate job on ubuntu-latest, so the write token never reaches the self-hosted runner. It creates tag lmcache-v<version>-g<sha8>-rocm-torch210 as a pre-release that is not marked latest, so ATOM's own release keeps its place. It refuses to overwrite an existing tag, and the release notes include the ARG LMCACHE_WHEEL_* lines to paste into the Dockerfiles. LMCache's Apache-2.0 LICENSE is included in the wheel's dist-info.

Validation

Ran the build job's run: steps locally, read straight from the YAML, with LMCache at 05fc77a0a7ababd9a7f2e343a771bc4bbc5b65cb:

• The build succeeded: lmcache-0.5.6.dev98+g05fc77a0.rocm7.2.4.torch2.10.git3d3aa833.cxx11abi1-cp312-cp312-manylinux_2_39_x86_64.whl
• LMCache validator passed, including GPU kernels on gfx950: OK: ROCm torch 2.10 wheel ABI; ATOM adapters available ... hip 7.2.53211
• ATOM validator passed: OK: lmcache 0.5.6.dev98+g05fc77a0... has the ATOM MP adapters and server flags
• actionlint 1.7.7 passes on all workflows; black and ruff pass on the new script.

The build job also passed on GitHub, on build-only-atom, triggered by a temporary push trigger that has since been reverted: https://github.com/ROCm/ATOM/actions/runs/35982738885. That run produced lmcache-0.5.6.dev98+g05fc77a0...whl (sha256 323ed34f...7be6) and both validators passed. The publish job was skipped, and it has not run yet; it first runs on the post-merge workflow_dispatch with publish=true.

Follow-up

Once this is merged, dispatch it with 05fc77a0a7ababd9a7f2e343a771bc4bbc5b65cb and publish=true, then pin the resulting release in docker/Dockerfile and docker/atom_release.dockerfile for #2250.

🤖 Generated with Claude Code

[github-actions]

🏷️ CI Guide

Runs automatically on every eligible PR before approval:

• ✅ Pre Checkin: Black, Ruff, catalog schema validation, non-GPU unit tests

Heavy model tests:

• ✅ Run after the PR is approved and Pre Checkin passes
• ✅ Run immediately when an approval review is submitted
• ✅ Can be requested before approval with labels

Label	Tests
ci:full	Run all heavy PR model tests: native ATOM, vLLM, and SGLang
ci:atom	Run native ATOM model accuracy tests
ci:vllm	Run ATOM vLLM OOT model accuracy tests
ci:sglang	Run ATOM SGLang model accuracy tests

Heavy jobs are skipped when the PR is not approved and no matching ci:* label is present.
Add labels via the sidebar or gh pr edit 2388 --add-label <label>