masstensor's picture
Release Llama-3.1-8B-Instruct-InWeight-OS full weights + modular skill cartridges + paper PDF
1949354 verified
Raw History Blame Contribute Delete
1.13 kB
"""
Hot-swappable Skill Cartridge Loader for In-Weight OS.
Allows installing or uninstalling any subset of {safety, tools, fast_router}
into the terminal MLP layer of Qwen3-8B or Llama-3.1-8B-Instruct in ~20ms.
"""
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
def install_cartridges(model, cartridge_paths):
with torch.no_grad():
for pt_path in cartridge_paths:
payload = torch.load(pt_path, map_location="cpu", weights_only=True)
layer_idx = int(payload["terminal_fsm_layer"])
mlp = model.model.layers[layer_idx].mlp
cols = payload["cols"].long()
dev = mlp.gate_proj.weight.device
dt = mlp.gate_proj.weight.dtype
mlp.gate_proj.weight[cols, :] = payload["gate"].to(device=dev, dtype=dt)
mlp.up_proj.weight[cols, :] = payload["up"].to(device=dev, dtype=dt)
mlp.down_proj.weight[:, cols] = payload["down"].to(device=dev, dtype=dt)
print(f"[In-Weight OS] Installed cartridge '{payload['skill_name']}' ({len(cols)} dormant neurons at Layer {layer_idx})")
return model