MoS DFlash Routers
This repository contains three experimental request-level router snapshots used by the MoS/DFlash studies. A router classifies one prompt and selects one of five draft experts:
0 code
1 math
2 factual_qa
3 creative_writing
4 general
These artifacts are not token-level routers, standalone language models, or complete serving systems. They require hidden states from the exact target model and feature contract recorded beside each weight file.
Release layout
request-level/
offline-v2/qwen3-8b/
intraining-sidecar/qwen3-8b/
c3/qwen3-4b/
registry/
schema/
verification/
manifests/
| Release | Feature contract | Head |
|---|---|---|
offline-v2/qwen3-8b |
layers 1/9/17/25/33; mean, max, and last pooling; stored z-score statistics | LayerNorm(61440) → Linear(512) → GELU → Linear(5) |
intraining-sidecar/qwen3-8b |
layers 1/9/17/25/33; mean pooling | LayerNorm(20480) → Linear(512) → GELU → Linear(5) |
c3/qwen3-4b |
layers 1/9/17/25/33; mean pooling | LayerNorm(12800) → Linear(512) → GELU → Linear(5) |
The offline-v2 source was a PyTorch checkpoint. This release converts only its tensor state into safetensors and does not publish the pickle container.
Loading
Install the two runtime dependencies:
pip install -r requirements.txt
Then load a release with the included helper:
from pathlib import Path
from router_loader import load_offline_v2, load_sidecar
root = Path("request-level/offline-v2/qwen3-8b")
head, feature_mean, feature_std, config = load_offline_v2(
root / "router_head.safetensors",
root / "router_config.json",
)
head.eval()
load_sidecar handles either in-training sidecar release. Feature extraction
is intentionally not hidden inside the loader: callers must implement the
recorded tokenizer, prompt template, target-layer, and pooling contract
exactly.
The offline-v2 deployment contract additionally zeros z-scored dimensions
whose stored standard deviation is at most 2e-6, then clamps the remaining
values to [-10, 10]. Use prepare_offline_features from
router_loader.py.
Verification boundary
The release manifest binds every public file to SHA-256 and byte size. The included validation report covers:
- exact tensor equality for the offline-v2 weights-only conversion;
- safetensors header, key, shape, dtype, and finite-forward checks;
- removal of machine-local paths and private per-sample material.
It does not claim task-level accuracy reproduction on a new inference backend. Exact tokenizer, target revision, prompt formatting, hidden-state indexing, and numeric dtype still need end-to-end parity testing.
Raw prompts, hidden-state feature tensors, per-sample routing records, training logs, and raw pickle checkpoints are not included.