ahd-9b-uspto

LoRA adapter for Qwen/Qwen3.5-9B.

Contents

• lora/ahd-9b-uspto/adapter_model.safetensors — LoRA adapter (attention/MLP deltas, vLLM-ready layout)
• lora/ahd-9b-uspto/adapter_model.tinker-original.safetensors — original export, kept for provenance and as the source of the output-head delta
• fold_lm_head.py — applies the output-head (lm_head) delta into a fresh copy of the base model

Usage

Requires torch, safetensors, and vLLM.

# 1. Download the base model Qwen/Qwen3.5-9B, e.g. to ./Qwen3.5-9B

# 2. Apply the output-head delta into a fresh model copy
python fold_lm_head.py lora/ahd-9b-uspto ./Qwen3.5-9B ./ahd-9b-uspto-base

# 3. Serve the model copy together with the adapter
vllm serve ./ahd-9b-uspto-base \
  --enable-lora --max-lora-rank 128 \
  --lora-modules ahd-9b-uspto=./lora/ahd-9b-uspto \
  --enable-auto-tool-choice --tool-call-parser qwen3_xml \
  --max-model-len 65536

Then request the model as ahd-9b-uspto through the OpenAI-compatible API.

Notes

• The fold step also writes a generation_config.json (greedy decoding) into the model copy; vLLM applies it automatically.
• --max-lora-rank 128 is required: the adapter packs per-head q/k/v deltas into fused projection blocks.
• The model reasons in <think>...</think> blocks before answering; tool calls use the Qwen3 XML format.