fix(L4T13 backends): switch vllm/sglang/vllm-omni to PyPI aarch64+cu130 wheels

[localai-bot]

Summary

Restores the L4T13 (JetPack 7 / NVIDIA Thor) backends by retiring the now-broken pypi.jetson-ai-lab.io/sbsa/cu130 mirror in favor of PyPI's official aarch64 + cu130 wheels - per the PyTorch team's April 2026 announcement.

Three logical changes, split across three commits:

#	Commit	What
1	fix(vllm): switch L4T13 backend to PyPI aarch64+cu130 wheels	The original ABI-mismatch fix (vllm backend)
2	refactor(vllm): retire l4t13 pyproject.toml in favor of requirements-*.txt	Mechanical follow-up - with [tool.uv.sources] gone, pyproject.toml is dead weight; replace with the standard requirements-${profile}.txt pattern
3	fix(sglang,vllm-omni): switch L4T13 backends to PyPI aarch64+cu130 wheels	Same root cause + fix applied to the other two L4T13 backends

The bug

vllm crashed at import on cuda13-nvidia-l4t-arm64-vllm images:

ImportError: /backends/cuda13-nvidia-l4t-arm64-vllm/venv/lib/python3.12/site-packages/vllm/_C.abi3.so:
undefined symbol: _ZN3c1013MessageLoggerC1EPKciib

That demangles to c10::MessageLogger::MessageLogger(char const*, int, int, bool) - the torch 2.10 form. Verified by nm -D on the actual wheels from pypi.jetson-ai-lab.io/sbsa/cu130:

• vllm-0.20.0+cu130-cp312-cp312-linux_aarch64.whl: _C.abi3.so references the 4-arg (char const*, int, int, bool) signature (built against torch 2.10).
• torch-2.10.0-cp312-cp312-linux_aarch64.whl: libc10.so exports that symbol.
• torch-2.11.0-cp312-cp312-linux_aarch64.whl: replaced it with MessageLogger(c10::SourceLocation, int, bool) - old form gone.

All three L4T13 backends pinned their torch / vllm / sglang to that mirror with no version pins and uv picked the latest torch (2.11.0) next to the older-ABI vllm/sglang wheel → undefined symbol at import. vllm-omni hadn't been reported broken yet but carries the exact same anti-pattern.

The fix

PyPI now publishes aarch64 + cu130 manylinux wheels directly:

package	PyPI wheel	notes
torch	torch-2.11.0-cp312-cp312-manylinux_2_28_aarch64.whl	__version__ = '2.11.0+cu130', cuda: '13.0'
torchvision	torchvision-0.26.0-cp312-cp312-manylinux_2_28_aarch64.whl
torchaudio	torchaudio-2.11.0-cp312-cp312-manylinux_2_28_aarch64.whl
vllm	vllm-0.20.0-cp38-abi3-manylinux_2_35_aarch64.whl	Requires-Dist: torch==2.11.0, torchvision==0.26.0, torchaudio==2.11.0
sglang	sglang-0.5.12-cp312-cp312-manylinux_2_34_aarch64.whl	Requires-Dist: torch==2.11.0, torchaudio==2.11.0

ABI verified end-to-end on the PyPI vllm wheel: _C.abi3.so references c10::MessageLogger::MessageLogger(c10::SourceLocation, int, bool) and libc10.so exports that exact symbol.

flash-attn is dropped everywhere: PyPI has no aarch64 wheel, but vLLM 0.20+ bundles its own vllm_flash_attn (fa2 + fa3) inside the main wheel.

For sglang, the new requirements-l4t13-after.txt matches the cublas13 split (sglang[all]>=0.5.11). The [all] extra was historically dropped on aarch64 because the transitive decord dep had no aarch64 wheel; sglang has since switched to decord2 on aarch64 (which does), so [all] is now safe and unblocks Gemma 4 / MTP recipes on Jetson Thor that were previously blocked by the 0.5.1.post2 cap on the L4T mirror.

Cleanup

pyproject.toml for vllm and sglang existed only to host [tool.uv.sources]. With that block gone, those files did nothing the standard requirements-${profile}.txt flow couldn't. Replaced with the same two-file pattern every other build profile uses; the special l4t13 elif branches in both install.shs vanish entirely. libbackend.sh's installRequirements already handles the requirements-install.txt build-deps pass, the C_INCLUDE_PATH export for PORTABLE_PYTHON, and runProtogen. vllm-omni's l4t13 vllm-install branch collapses into the cublas13 branch (both now: pip install vllm --torch-backend=auto).

Test plan

• CI image build for -nvidia-l4t-cuda-13-arm64-vllm completes (installRequirements resolves torch 2.11.0 + vllm 0.20.0 from PyPI without an extra index).
• CI image build for -nvidia-l4t-cuda-13-arm64-sglang completes (torch 2.11.0 + sglang 0.5.12, [all] extra resolves with decord2 on aarch64).
• CI image build for -nvidia-l4t-cuda-13-arm64-vllm-omni completes (PyPI vllm + clone-from-source vllm-omni).
• Confirm vllm import no longer raises undefined symbol: _ZN3c1013MessageLoggerC1EPKciib on Jetson Thor.
• Confirm the backend starts via NATS install on nvidia-thor without timing out at startup.
• Smoke-test a small chat completion against vllm + sglang backends on Thor.