MiMo V2.6 Pro MOPD SGLang Intervention Runtime

Canonical runtime artifacts are published on Hugging Face.

This repository packages the Blackfrost-AI SGLang compatibility layer and the reversible pass-2 runtime intervention used to serve MiMo V2.6 Pro MOPD on NVIDIA Blackwell. It is a runtime package, not a model checkpoint. Model weights, tokenizer files, prompt text, captured activations, evaluation traces, compiled kernels, and deployment credentials are not included.

The included release contains:

• six qualified SGLang 0.5.19 overrides for the MiMo MXFP4 path;
• an isolated-overlay installer that never edits the shared Python environment;
• two 6,144-dimensional BF16 direction artifacts and a portable manifest;
• a TP8 launch profile matching the qualified 8-GPU deployment;
• exact source revisions, hashes, compatibility limits, and third-party notices.

Method credit

The refusal-direction/abliteration workflow, boundary capture-and-freeze methodology, and lambda-strength intervention framing are derived from public work by Keys (drowzeys). Blackfrost-AI adapted and refit the method for MiMo-V2.6-Pro-MOPD, implemented the SGLang runtime loader, and ran the iterative pass evaluation. Keys/drowzeys did not author these exact direction tensors and is not implied to endorse this release. See CREDITS.md.

Tested configuration

Pass 2 was selected from five experimental passes using automated marker-based triage plus separate control prompts. That triage is not semantic review, a broad quality benchmark, or a safety guarantee. Revalidate the intervention on your own workload before deployment.

Install

Use an environment that already has the GPU-specific SGLang stack installed. The exact versions observed in the qualified environment are recorded in environment-tested.txt. CUDA, PyTorch, FlashInfer, and FlashAttention wheels are platform-specific, so this repository does not replace them.

hf download Blackfrost-AI/MiMo-V2.6-Pro-MOPD-Derisking-Intervention-Runtime \
  --local-dir mimo-pro-runtime

python3 mimo-pro-runtime/scripts/verify_package.py
python3 mimo-pro-runtime/scripts/materialize_overlay.py \
  --output mimo-pro-runtime/runtime-overlay

materialize_overlay.py requires a pristine sglang==0.5.19. It verifies the six upstream base hashes, copies the installed package into a private overlay, applies the qualified overrides, and records the base environment. An extracted official package can instead be supplied with --base-package.

Download the pinned checkpoint separately:

hf download XiaomiMiMo/MiMo-V2.6-Pro-MOPD \
  --revision adea8e2c5373181e5a973fa1ecb343cb31af214b \
  --local-dir /models/MiMo-V2.6-Pro-MOPD

Serve

The defaults reproduce the qualified TP8 profile:

MODEL_PATH=/models/MiMo-V2.6-Pro-MOPD \
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
HOST=127.0.0.1 \
PORT=30000 \
  bash mimo-pro-runtime/scripts/launch_pro_mopd.sh

The launcher defaults to the cumulative two-direction pass-2 manifest with alpha 2.0 per direction, source layer 38, and target layers 46-49. The API model ID defaults to mimo-2.6-pro. Override INTERVENTION_MANIFEST to use another compatible manifest. To measure a loader-only baseline, launch SGLang directly through the materialized overlay without setting BLACKFROST_MIMO_MOE_INTERVENTION.

Additional SGLang arguments can be appended to the launcher. It binds to loopback by default; place authentication or a private network in front of it before allowing remote clients.

Verify the live API

python3 mimo-pro-runtime/scripts/smoke_test.py \
  --base-url http://127.0.0.1:30000/v1

The smoke test checks /models and submits a deterministic chat completion.

What the intervention does

At each configured layer, the runtime applies both rank-one projections after the down-projected MLP/MoE output:

y <- y - alpha * d * (d^T y)

Vectors are normalized and re-orthogonalized in manifest order at load time. The implementation checks the schema, provenance metadata, SHA-256 hashes, orientation, target layers, alpha values, hidden width, finite values, and path confinement before placing vectors on GPU. Tensor payloads use PyTorch's restricted weights_only loader. Router decisions and model weights are not modified.

Set BLACKFROST_MIMO_MOE_INTERVENTION before importing or launching SGLang; configuration is cached per process.

Scope

• Model weights remain governed by their upstream repository terms.
• The runtime overrides are modifications of Apache-2.0 SGLang 0.5.19.
• Blackfrost-AI releases its runtime changes and included direction artifacts under Apache-2.0.
• The package is research software and should be requalified after changing the checkpoint, tokenizer/system prompt, runtime, GPU architecture, alpha, or direction set.

See COMPATIBILITY.md, MODIFICATIONS.md, PROVENANCE.md, SECURITY.md, and release-manifest.json before adapting the runtime.