---
base_model: Qwen/Qwen3.8-27B
library_name: peft
license: apache-2.0
tags: [verilog, rtl, hardware-design, lora, verilog-eval]
---

# RecurRTL-MATE — `localizer` adapter

LoRA adapter for **Qwen/Qwen3.8-27B**, one of four role adapters in the RecurRTL-MATE
recursive RTL-generation pipeline. This one identifies which part of a failing design is responsible.

## Usage

```python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.8-27B", device_map="auto")
model = PeftModel.from_pretrained(base, "REPO_ID", subfolder="localizer")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.8-27B")
```

The base model is ~27B parameters; in 4-bit it needs roughly 17 GB of VRAM.

## Training

QLoRA (4-bit NF4 base, bf16 compute), rank 16, alpha 32, on an NVIDIA H100 NVL.
Adapters were trained on recursive-repair trajectories rather than on benchmark
reference solutions.

## Base-model baseline

The **unmodified** base model scores **77.56% pass@1** (121/156) on VerilogEval v2
spec-to-rtl — n=1, temperature 1.0, thinking enabled, official `sv-generate --rules`
prompt, scored with the official iverilog harness. Compile rate 88.5%; among designs
that compiled and simulated, 87.68% passed.

That figure is the **baseline these adapters are meant to improve on**, measured on the
4-bit GGUF build of the same base weights. Treat it as the reference point, not as a
result for this adapter.

## Benchmark caveat

`Prob099_m2014_q6c` is broken upstream in NVlabs/verilog-eval: its testbench binds ports
`Y2`/`Y4` that its own reference module does not declare, so the reference cannot compile
against its own testbench and no model can pass it. It is counted as a failure above
because published numbers count it too.