REBUILD.md

Rebuild and serve (d3-mini-agentcode, pre-release)

Use the merged weights (this repo)

The repo root is a complete d3 package: the base layout, runtime code, tokenizer and config, with the fine-tuned backbone weights and answer readout.

from transformers import AutoModel
model = AutoModel.from_pretrained("stonefish-labs/d3-mini-agentcode", trust_remote_code=True)

MODEL_MANIFEST.json lists every file's SHA-256 and size. On every load the runtime checks the files it reads (verify="fast": it hashes files up to 64 MB and checks the size of larger ones). verify="full" hashes them all, and verify="release" also checks the card and other files the runtime doesn't read. SHA256SUMS covers every file except itself and the manifest.

Rebuild the merged weights from the adapter

adapter/ holds the LoRA adapter (adapter_model.safetensors, adapter_config.json), the trained answer readout (readout.safetensors) and the training settings (training_args.json).

1. Download the base package at the pinned revision: huggingface-cli download vllm-sr/d3-mini --revision df99bf4ca4418bc6908ee68e9b7479af1f3b8800 --local-dir d3-mini
2. Merge (requires torch and safetensors; runs on CPU or GPU): python adapter/merge.py --base d3-mini --adapter adapter --out merged --name d3-mini-agentcode
   • Each adapted weight becomes W ← bf16(W + (α/r)·B·A), with α/r = 2.
   • The trained readout replaces the base readout.
   • Everything else is copied from the base, and a new MODEL_MANIFEST.json is written.
3. Check: the rebuilt model-0000{1,2}-of-00002.safetensors and readout.safetensors match the hashes in this repo's SHA256SUMS. They did on our rebuild.

Serve

• transformers: as above. model.system_one(state=..., questions=...); see examples/usage.py.
• vLLM Semantic Router: the package keeps d3-mini's format (d3-code-readout-v1), so runtimes that serve d3-mini should serve it too (see the d3-mini card). We only ran it with the bundled runtime above.
• Hardware: the bf16 weights take about 9 GB. On an RTX 4090, batches of 8 at up to 8,192 tokens peaked at 12.7 GiB allocated.