Add vLLM backend support

[sozercan]

Summary

• Adds vLLM as a third backend option (alongside llama-cpp and diffusers) for high-throughput GPU inference with HuggingFace safetensors models
• vLLM is CUDA-only, amd64-only — validated end-to-end on NVIDIA A100 80GB with Qwen2.5-0.5B-Instruct
• Includes build-time workarounds for the pre-built LocalAI vLLM backend image (flash_attn ABI fix, AsyncLLM API patch)

Changes

• pkg/utils/const.go — BackendVLLM constant
• pkg/aikit2llb/inference/backend.go — Backend tag/name/alias resolution, dependency installation, compatibility patches for vLLM
• pkg/aikit2llb/inference/vllm.go — Python base deps + gcc/libc6-dev for Triton JIT
• pkg/aikit2llb/inference/convert.go — CUDA apt packages for vLLM
• pkg/aikit2llb/inference/download.go — HuggingFace repo-level downloads (huggingface://namespace/model)
• pkg/build/build.go — Validation (vLLM requires CUDA, amd64)
• test/aikitfile-vllm.yaml — Test aikitfile (runtime model download)
• .github/workflows/test-docker-gpu.yaml — vLLM in GPU test matrix
• website/docs/ — Specs and GPU docs updated

Test plan

• make test — 148 tests pass (including new vLLM cases for backend tag/name/alias, build validation, ARM64 rejection)
• make lint — 0 issues
• Build AIKit image (docker buildx build .)
• Build vLLM model image (docker buildx build -f test/aikitfile-vllm.yaml)
• Run inference on GPU VM: curl /v1/chat/completions with Qwen2.5-0.5B-Instruct — success

[Copilot]

Pull request overview

Adds vLLM as an additional LocalAI backend option to support high-throughput CUDA GPU inference with HuggingFace safetensors models, including build-time validation and HuggingFace repo-level model downloads.

Changes:

• Introduces vllm backend constant plus backend tag/name/alias resolution and install flow (including compatibility patching).
• Adds HuggingFace repo-level downloads (huggingface://namespace/model) via an hf-cli helper image.
• Extends build validation, tests, docs, and GPU CI workflow matrix to cover vLLM (CUDA-only, amd64-only).

Reviewed changes

Copilot reviewed 12 out of 12 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
website/docs/specs-inference.md	Documents vllm as an allowed backend option.
website/docs/gpu.md	Adds vLLM usage guidance and example config for GPU inference.
test/aikitfile-vllm.yaml	Adds a vLLM test aikitfile that downloads the model at runtime.
pkg/utils/const.go	Adds BackendVLLM constant.
pkg/build/build.go	Enforces vLLM runtime requirement (runtime: cuda) and supports vLLM in backend allowlist.
pkg/build/build_test.go	Adds test cases for vLLM validation and arm64 rejection.
pkg/aikit2llb/inference/backend.go	Adds vLLM tag/name/alias logic, dependency install hook, and compatibility patching for the backend image.
pkg/aikit2llb/inference/backend_test.go	Adds backend tag/name/alias tests for vLLM on CUDA/amd64.
pkg/aikit2llb/inference/vllm.go	Adds vLLM dependency installation (Python base deps + gcc/libc6-dev).
pkg/aikit2llb/inference/vllm_test.go	Adds a basic non-panicking test for vLLM dependency install wiring.
pkg/aikit2llb/inference/convert.go	Ensures CUDA libs are installed when vLLM backend is selected.
pkg/aikit2llb/inference/download.go	Adds HuggingFace repo-level download support using hf CLI image + optional HF token secret.
.github/workflows/test-docker-gpu.yaml	Adds vLLM to GPU test matrix and validates a chat completion response.

[Copilot]

installVLLMDependencies installs Python deps via installPythonBaseDependencies(s, merge) (which produces a diff from the current state), but then computes the gcc/libc6-dev diff from the original state s again. Merging two diffs that both modify dpkg/apt state (e.g., /var/lib/dpkg/status) can lead to lost updates or non-deterministic results. Consider running the gcc install step on top of the state returned by installPythonBaseDependencies (or performing both installs in a single Run and taking one Diff) so dpkg changes are applied sequentially instead of merged.

[Copilot]

This test only checks that .choices is a non-empty string. If the API returns { "choices": null } or { "choices": [] }, jq '.choices' outputs null/[] (both non-empty strings) and the step will still pass. Consider using jq -e with a predicate that asserts choices is an array with length > 0 (and fail otherwise) to avoid false positives in the vLLM GPU test.