feat(baremetal): bump vLLM default to 0.28.0 with glibc gate

[BaoYunkai]

Problem

1. Hyperloom still defaulted bare-metal vLLM to 0.27.1 while 0.28.0 is the current ROCm release.
2. vLLM 0.28.0+ wheels require glibc >= 2.39 (Ubuntu 24.04+), but docs still treated 22.04 and 24.04 equally and the installer did not fail early on incompatible hosts.

Fix

1. Bump the bare-metal vLLM default to 0.28.0 (rocm723) and sync the pinned version across compatibility docs, install docs, example skills, and Slurm model metadata.
2. Enforce the glibc requirement inside install_vllm_framework() before any venv/pip work (including --check-only / --dry-run), with actionable guidance for Docker or a pre-0.28 override.
3. Document Ubuntu 24.04 as the recommended bare-metal OS; mark 22.04 as legacy for vLLM 0.28.0+.
4. Trim hyperloom-setup to a short UX explanation of the installer gate; remove the unconditional VLLM_VERSION=0.27.1 from the normal vLLM command.
5. Revert the SGLang profiler-hotfix gating from this PR; it will land in a separate PR based on observed torch.version.hip.

Test plan

• pytest src/hyperloom/inference_optimizer/tests/test_setup_cli.py -k "hotfix or glibc or vllm_install"
• pytest src/hyperloom/inference_optimizer/tests/test_baremetal_doc_version_consistency.py
• ruff format --check .

[xiaofei-zheng]

Review

What this PR does: bumps the bare-metal VLLM_VERSION default from 0.27.1 to 0.28.0 (variant stays rocm723) and propagates the literal to every place that names it — docs/compatibility.rst, docs/install/install.md, the four example SKILL.md recipes, assets/slurm/models.tsv, IMAGE_HINT, and the mirrored comments in test_baremetal_doc_version_consistency.py.

Blocking

The 0.28.0 wheels raise the glibc floor to 2.39, so bare-metal vLLM install now hard-fails on Ubuntu 22.04 — install_baremetal.sh:80.

The index contents differ from 0.27.1 in more than the version:

0.27.1: vllm-0.27.1+rocm723-cp312-cp312-manylinux_2_34_x86_64.whl
        torch-2.11.0+gitd0c8b1f-cp312-cp312-manylinux_2_35_x86_64.whl
0.28.0: vllm-0.28.0+rocm723-cp312-cp312-manylinux_2_39_x86_64.whl
        torch-2.12.0+git6bbd260-cp312-cp312-manylinux_2_39_x86_64.whl

manylinux_2_39 requires glibc >= 2.39 (Ubuntu 24.04+). Ubuntu 22.04 ships glibc 2.35, which satisfied both 0.27.1 wheels and satisfies neither 0.28.0 wheel.

Failure path (isolated mode, the default for vLLM): install_vllm always creates a fresh venv and installs unconditionally — there is no "already present, skip" branch. It first scrapes ${VLLM_ROCM_INDEX}/torch/ and pins torch==2.12.0+git6bbd260 (install_baremetal.sh:900); pip finds no wheel matching the interpreter's supported platform tags and the script hits die "failed to install ROCm torch==... into ${VLLM_VENV_ROOT}" at line 901. Even if torch resolution were skipped, pip install vllm==0.28.0+rocm723 fails the same way at line 908.

This breaks two paths the repo documents as supported:

1. docs/compatibility.rst:50 still lists Ubuntu 22.04 as a supported OS, and docs/compatibility.rst:187 still tells users the bare-metal route installs vllm==0.28.0+rocm723. On a 22.04 ROCm 7.2.3 host provisioned per the doc, the documented command now fails with Could not find a version that satisfies the requirement.
2. IMAGE_HINT (install_baremetal.sh:190) recommends running the installer inside vllm/vllm-openai-rocm:v0.28.0, and docs/compatibility.rst:191 asserts the bare-metal stack "matches" that image. That image's base is Ubuntu 22.04 (org.opencontainers.image.version=22.04 on the config blob for the v0.28.0 manifest), so the installer cannot install its own matching wheel inside the very image it points at.

Note the docker path is unaffected — the image ships vLLM prebuilt. Only the bare-metal/isolated-venv path regresses.

Suggested resolution (either is fine, but the PR shouldn't land as-is):

• gate the vLLM bare-metal path on glibc >= 2.39 with an explicit, actionable error (and cover it in --check-only / --dry-run, which today print a clean preflight and exit 0 on a host where the real run will die), and update the support matrix + the "matches its container image" sentence to state that bare-metal vLLM 0.28.0 requires Ubuntu 24.04; or
• keep the 0.27.1 default until the 22.04 requirement is dropped repo-wide.

Verification performed

• Fetched both wheels.vllm.ai/rocm/{0.27.1,0.28.0}/rocm723/{vllm,torch}/ index listings directly; tags quoted above are the real filenames.
• Pulled the vllm/vllm-openai-rocm:v0.28.0 manifest + config blob from Docker Hub to confirm the 22.04 base.
• Read install_vllm end to end on the PR head to confirm the isolated path installs unconditionally and that both pip failures are fatal (die), not warnings.
• Confirmed test_baremetal_doc_version_consistency.py only asserts the installer default against docs/compatibility.rst; the version/image literals in install.md, the four SKILL.md files, models.tsv, and IMAGE_HINT are unguarded — not blocking for this PR, but it is why the bump was an 8-file hand sweep.

No other blocking issues: the doc/example string replacements are complete and consistent with the new default, and the version-consistency test still passes against them.

[ZhengGong-amd]

Request changes

Please keep this PR to the vLLM default upgrade and make the installation boundary authoritative. Do not add further prompt rules or version-inference fallbacks.

1. Fix CI first. ruff format --check . currently fails in test_setup_cli.py.

2. Enforce the vLLM 0.28 ABI requirement in install_vllm_framework(), before any venv or pip mutation. The 0.28.0+rocm723 vLLM and torch wheels are manylinux_2_39; Ubuntu 22.04's glibc 2.35 cannot install them. The new hyperloom-setup instruction is advisory and is bypassed by direct installer/backend/automation calls; --check-only and --dry-run currently look healthy too. Check the actual glibc capability (not merely the Ubuntu label), fail with actionable Docker / pre-0.28 override guidance, and cover: glibc 2.35 rejection, 2.39 acceptance, and a permitted pre-0.28 override. Keep the prompt as a short UX explanation of that installer result, not as a second policy implementation.

3. Remove the unconditional export VLLM_VERSION=0.27.1 from the normal vLLM command in hyperloom-setup/SKILL.md. A user on Ubuntu 24.04 who follows that command never receives the new 0.28.0 default. Put the downgrade in a separate, explicitly conditional snippet.

4. Split out the SGLang profiler-hotfix change. It is independent of the vLLM upgrade, and the new helper is not a valid runtime detector: SGLANG_ROCM_EXTRA defaults to rocm724 and wins before the real HIP version is read, so an actual rocm720 SGLang container with that env unset skips its required overlay. .env does not restore this variable before the gate and the default is later written back. The current tests force the env value, so they miss the real path. In a separate PR, use the SGLang interpreter's observed torch.version.hip as the single source of truth; do not infer runtime state from an install-index setting or /opt/rocm. Verify real rocm720/rocm724 images and test unset/stale config plus the exact supported HIP boundaries.

This produces one small, deterministic vLLM-default change now; the host-library mutation can then be reviewed independently against real stack evidence.