amd-ci-test-bisect

name: amd-ci-test-bisect description: > Add or update AMD/ROCm SGLang regression tests, choose the right MI325/MI35x CI suite, reproduce AMD CI locally with the upstream amd_ci container scripts, and bisect AMD CI regressions on main/nightly. Use when an agent-generated fix needs register_amd_ci coverage, when selecting MI325 vs MI35x runners, when debugging pr-test-amd or nightly AMD failures, or when documenting the register_amd_ci flow for ROCm-specific fixes.

AMD CI Test / Bisect

This skill is the AMD/ROCm adaptation of SGLang's write-sglang-test, ci-workflow-guide, and sglang-bisect-ci-regression skills.

Use it for four things:

1. add a regression test for an AMD-only or ROCm-sensitive fix
2. register that test with register_amd_ci(...)
3. reproduce the AMD CI job locally through the upstream scripts/ci/amd/* flow
4. bisect an AMD CI regression across MI325 / MI35x runners, ROCm versions, and image tags

Read references/suite-matrix.md when you need exact suite/runner mappings or local reproduction commands.

Read references/bisect-playbook.md when you are debugging a failing AMD CI run on main or nightly.

Core Rules

1. register_amd_ci is AST-parsed. Keep est_time, suite, nightly, and disabled as module-level literal constants. Do not hide them behind helper functions, variables, or computed expressions.

2. Only add AMD registration when AMD coverage is the point. Use register_amd_ci(...) for ROCm-only kernels, HIP/aiter paths, MI35x/gfx950 behavior, RCCL/distributed AMD paths, or an AMD regression fix. Do not duplicate backend-independent tests onto AMD just because AMD exists.

3. Choose the lightest AMD suite that proves the fix. Prefer 1-GPU MI325 first. Escalate to MI35x only when the failure depends on gfx950 / MI355X / ROCm 7.2 image behavior. Escalate to 2/4/8 GPU only when the bug requires distributed state, DP/TP, or model scale.

4. MI325 and MI35x are different validation targets. In upstream AMD CI, MI325-class runners are generally normalized onto mi30x images. MI35x runners use MI35x images and amd_ci_exec.sh injects GPU_ARCHS=gfx950. Treat them as separate contracts, not interchangeable hardware.

5. Reproduce with the AMD CI scripts, not ad-hoc shell state. Use ensure_vram_clear.sh, amd_ci_start_container.sh, amd_ci_install_dependency.sh, and amd_ci_exec.sh. That is the closest match to what GitHub Actions actually runs.

6. For regressions, separate code regressions from runner/image drift. Always record runner label, GPU family, ROCm version, and container image tag before blaming a commit.

Workflow

Phase 1: Classify the fix

Ask:

• Is the bug AMD-only, or just observed first on AMD?
• Does it depend on is_in_amd_ci(), HIP kernels, aiter, RCCL, or ROCm image contents?
• Is it MI35x/gfx950-specific, or should MI325 coverage be enough?
• Does it need 1 GPU, 2 GPU, 4 GPU, or 8 GPU to reproduce?

If the answer is "common logic, no AMD-specific path", keep the test on CPU/CUDA and do not add AMD registration just for symmetry.

Phase 2: Add the test

Default SGLang authoring rules still apply:

• use CustomTestCase
• make tearDownClass defensive
• prefer mocks/unit tests when a server is unnecessary
• place CI-discovered tests under test/registered/**

For AMD-specific behavior, the common pattern is:

from sglang.test.ci.ci_register import register_amd_ci

register_amd_ci(est_time=120, suite="stage-b-test-1-gpu-small-amd")

When the same file should run on both CUDA and AMD, register both explicitly and gate behavior inside the test body with is_in_amd_ci().

Use disabled="reason" instead of deleting coverage when a suite is temporarily too expensive or unstable.

Phase 3: Pick the suite

Use references/suite-matrix.md.

Practical default:

• 1-GPU ROCm correctness/kernel issue: stage-a-test-1-gpu-small-amd or stage-b-test-1-gpu-small-amd
• 1-GPU heavy model / memory issue on AMD: stage-b-test-1-gpu-large-amd
• MI355X/gfx950-specific issue: stage-b-test-1-gpu-small-amd-mi35x or stage-c-test-large-8-gpu-amd-mi35x
• distributed AMD issue: stage-b-test-2-gpu-large-amd or stage-c-test-4-gpu-amd
• long model/eval coverage: nightly AMD suites

Phase 4: Reproduce locally

From a SGLang checkout:

bash scripts/ci/amd/ensure_vram_clear.sh
bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
bash scripts/ci/amd/amd_ci_install_dependency.sh
bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" \
  python3 run_suite.py --hw amd --suite stage-b-test-1-gpu-small-amd

Notes:

• amd_ci_start_container.sh auto-detects MI325/MI35x from the runner hostname.
• MI325/MI300 runners collapse to mi30x images; MI35x keeps mi35x.
• amd_ci_exec.sh auto-adds SGLANG_IS_IN_CI_AMD=1, SGLANG_USE_AITER=1, and on MI35x also GPU_ARCHS=gfx950.
• For disaggregated tests, use amd_ci_start_container_disagg.sh.

Phase 5: Bisect regressions

Use references/bisect-playbook.md.

The AMD twist is important:

• pr-test-amd.yml does push / PR / rerun-stage coverage, not scheduled cron
• nightly-test-amd-rocm720.yml is the scheduled source of truth for long AMD coverage

So:

• use push runs on main for per-commit AMD regressions
• use scheduled nightly AMD runs for large-model / MI35x / long-running regressions

Output Requirements

When you finish an AMD test / CI / bisect task, report:

1. whether the fix is AMD-only or cross-backend
2. the chosen register_amd_ci suite and why
3. whether validation target is MI325/mi30x or MI35x/gfx950
4. the exact local reproduction command
5. if bisecting, whether the root cause is:
   • code regression
   • runner/hardware-specific
   • image / ROCm version drift
   • flaky / nondeterministic

Minimal Checklist

• test added or updated in the right folder
• register_amd_ci(...) literals are valid for AST parsing
• suite matches the real hardware requirement
• local AMD CI reproduction command was documented
• any MI35x-only assumption is stated explicitly
• reviewer can tell whether this should also keep CUDA coverage