codemaivanngu/fmrg-b200-portable

Preparation update — 2026-11-01

Read the new GenEval/VLM/reconstructed-data handoff. GenEval Modal B200 technical canary passed; company qualification remains pending. Full launch remains disabled. Older immutable core artifacts are preserved.

Full-paper preparation update — 2026-10-29

Read FULL_PAPER_HANDOFF.md for the audited contracts, new evaluator assets, offline setup and remaining blockers. No full benchmark is authorized. The older core runtime instructions below remain a historical immutable setup reference.

FMRG public portable bundle

Handoff with pinned download URLs and setup command.

Artifact commit: 73353f15eab0024925c942d197a7ed39a71aca0d. All files are public.

FMRG portable runtime and pilot assets

This bundle supports offline setup on an allocated NVIDIA B200 Linux x86_64 node. It contains a dependency runtime at /opt/fmrg, separately pinned model/input assets, pilot source, and pilot outputs. No company node has been qualified yet.

Use the immutable publication manifest and setup helper linked in HANDOFF.md. Download through the operator's approved HF proxy, verify the helper and manifest SHA-256, then call setup_public.sh MANIFEST_URL MANIFEST_SHA DOWNLOAD_DIR NEW_DATA_DIR NEW_RUN_DIR. These must be approved absolute paths; the download directory and parent directories must exist. Setup requires write access to /opt, refuses an existing /opt/fmrg or data/output destination, and runs the imports/kernel/tiny/full technical gates. Keep scheduler-provided GPU visibility. Do not run outside the allocation or install dependencies from PyPI on the node. Allow roughly 100 GiB free for downloaded archives and extraction; full datasets and experiment outputs need additional storage. Verify actual capacity first.

The public runtime repacks the qualified dependency archive, removing cloud launchers and local infrastructure paths in its tools directory. Dependency, interpreter and upstream source file bytes are preserved. Its archive SHA differs from the historical qualification archive; see manifests/archive-audit.json. The source upstream is jrrhuang/fmrg at be485ba40f1d1c163b95ad61112aca6baa13ed22. The runtime uses Python 3.11.14, Torch 2.7.1+cu128, CUDA 12.8 and SDPA. Ubuntu22/Modal B200 was qualified; Ubuntu24 and the company node still require their own receipts. Host prerequisites include curl, tar, sha256sum, Python 3 for validation, an NVIDIA driver supporting CUDA 12.8, and Ubuntu libgl1/libglib2.0-0 support libraries (the prior clean-restore container supplied these). Missing host libraries must be resolved through the operator's approved host/container setup; the helper does not run apt or change the system driver.

Assets retain their original immutable bytes. Their embedded manifest calls the original export private; publication was subsequently explicitly authorized on 2026-10-28. This label is historical and is not the current HF visibility. The bundle includes FLUX.1-dev, flow-map LoRAs, CLIP ViT-H, three input PNGs and six frozen measurements. It excludes full AFHQ/FFHQ datasets and the full paper reward/evaluation ensemble. Public availability does not replace upstream terms: the original FLUX license is included in the assets and under licenses/. The flow-map upstream has no separate license metadata. Dependencies retain their individual license files inside the runtime. Model authors retain rights. Rights to use FLUX.1-dev and its derivatives are granted directly by Black Forest Labs to recipients under the included license.

The FLUX.1 [dev] Model is licensed by Black Forest Labs Inc. under the FLUX.1 [dev] Non-Commercial License. Copyright Black Forest Labs Inc. IN NO EVENT SHALL BLACK FOREST LABS INC. BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH USE OF THIS MODEL.

Pilot results: G has 288 configs and eight oracle checks; I/J/R have 72/96/24 outputs. R is CLIP-only. These are small screening runs, not full reproduction of the paper tables. All generated output images are AI-generated. See results/PILOT_RESULTS.md and the raw results archive for scope and exact metrics. J is not certified bitwise reproducible; the prior clean-restore smoke differed by up to 7/255 per PNG channel. The reward counter autograd_calls hooks torch.autograd.grad only; zero does not mean no Tensor.backward() execution. Resolve loss, schedule, data split/tuning and evaluator contracts before a full scientific run. The helper launches technical gates only, never scientific jobs.