docs(vllm): record tsingmicro tsm260610 0.20.2 F/T E2E result (#872)

docs(vllm): record tsingmicro tsm260610 0.20.2 F/T E2E result

The TX8110 backend passes the 0.20.2 dual-compiler end-to-end run: both the FlagTree and the vendor Triton path serve Qwen3-4B and return byte-identical outputs for the anchor prompts, with the same operator routing on both paths.

Records the three TX8110-specific pitfalls fixed along the way (paged KV cache writes need base indexing, 0-dim .item() faults, strided device reads are ~80x slower than contiguous), plus the dispatch configuration that routes attention to the vendor backend and rms_norm / rotary_embedding to reference.

Only the runtime side was exercised, so the per-backend facility booleans (deps_app / launch_docs / image_tag) stay unset.

This PR was written in part with the assistance of generative AI.

docs(vllm): link the tsingmicro 0.20.2 enablement PR

The verified wheel was built from the PR head, so the backend record and the matrix track it.

This PR was written in part with the assistance of generative AI.

Co-authored-by: flagos-ci noreply@flagos.net

build-infra

FlagOS container image build infrastructure. This repo defines and builds the base images — a vendor’s SDK and toolchain on an OS base — that FlagOS runtime images are built on.

📖 Documentation: https://flagos-ai.github.io/release-info/

The docs (supported backends, per-image environment, dependencies) are generated from configs.yaml + base/, so they can’t drift from the source.

Layout

Path	What
`configs.yaml`	Source of truth: per-backend deps, versions, image env
`base/<vendor>-<backend>`	Base image containerfiles
`scripts/build_base.py`	Build a base image (reads OCI labels + registry)
`scripts/generate_matrix.py`	Emit the CI build matrix from `configs.yaml`
`scripts/build_runtime.py`	Build a runtime image
`.github/build-config.yml`	Global build config: registry + per-backend runners
`runtime/`	Runtime Containerfile + per-image readmes
`packaging/flagtree/`	Low-glibc FlagTree wheel builders
`docs/`	Hugo docs site + data-generation scripts

Quick start

Build a base image locally:

python scripts/build_base.py nvidia-cuda12.8 --dry-run   # preview
python scripts/build_base.py nvidia-cuda12.8 --push      # build + push

In CI, base images are built on demand via the Base Image Build (manual) workflow (workflow_dispatch) — pick a backend (or all) and whether to push.

To add a vendor/backend, see the onboarding guide.