fix(BACKEND-GATE-ROCM-SGLANG): isolate offload inspection from installed kernels
Nobody has claimed this yet.
- Dominant language
- C++
- Stars
- 423
- Forks
- 53
- Avg merge
- 20h 26m
- Merged PRs (30d)
- 310
Description
Row: BACKEND-GATE-ROCM-SGLANG
Owner: current Strix benchmark campaign operator. Found during #3053 preparation retry fb77c7b0-4fa8-4e69-80d9-115845f6922e. The aab36a6a4 HIP identity fix passes, kernel compilation and initial sgl_kernel import pass, but upstream tests fail collection. The extension-targets.log proves llvm-objdump --offloading extracts host/HIP bundles beside the installed common_ops extension. Pinned SGLang load_utils.py glob common_ops.* accepts names containing .so and attempts only the first candidate, which is an extracted non-importable bundle.
Fix the preparation caller to inspect an exact verified copy in an isolated scratch directory, preserve the installed extension/package bytes and candidate inventory, retain gfx1151-only validation, and add a CLI regression with an inspector that creates adjacent artifacts. No upstream loader patch or pin change is needed. Use committed design before a fresh implementation and independent mutation review. Evidence directory: /mnt/nas_share/rc/strix-four-engine-3053.X94a3J/prepare-aab36a6a-rebuilt-03/logs. #3053 owns downstream kernel tests, AITER/model startup and benchmarks.
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start with the preparation caller and the pinned SGLang load_utils.py behavior described in the issue. Inspect extension-targets.log and the evidence directory to understand the adjacent artifacts, then use the committed design before implementing the isolated inspection. Done means installed package bytes and candidate inventory remain intact, gfx1151-only validation is retained, and the CLI regression covers an inspector that creates adjacent artifacts.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- cpp, python
- Domain
- backend, build-system, testing-qa
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Active
- Clarity
- Mostly clear
- Newbie friendliness
- 48/100