Blame for README.md

---
library_name: transformers
pipeline_tag: text-generation
license: apache-2.0
base_model:
  - tsinghua-sigs-robot-lab/VeriLoop-E2
base_model_relation: quantized
language:
  - en
  - zh
tags:
  - veriloop
  - veriloop-coder
  - code
  - coding-agent
  - software-engineering
  - mathematical-reasoning
  - nvfp4
  - modelopt
  - tensorrt-llm
  - vllm
  - sglang
  - code-optimized
  - quantization
  - open-source
  - apache-2.0
  - qwen3_5
  - self-harness
  - harness-engineering
  - surface-host-adapter
  - evidence-binding
  - rollback
  - uncertainty-calibration
  - long-context
  - vertical-code-model
  - recursive-improvement
---

<div align="center">
  <h1>VeriLoop Coder E2 · NVFP4</h1>
  <p><strong>Coding-Optimized Quantized Model (NVIDIA NVFP4)</strong></p>
  <p>
    <a href="https://huggingface.co/tsinghua-sigs-robot-lab/VeriLoop-E2">Original Model ↗</a>
    ·
    <a href="https://github.com/rodrigoramosrs">GitHub</a>
    ·
    Apache-2.0
  </p>
</div>

---

## Overview

This repository contains an **NVFP4 quantization** of [VeriLoop E2](https://huggingface.co/tsinghua-sigs-robot-lab/VeriLoop-E2), an open 27B post-trained model built on **Qwen3.8-27B** for code, mathematics, and physics. Its core reasoning discipline is **VeriLoop-Governed Recurrence (VGR)**: candidate states are recursively proposed, externally checked, and retained only when the protected evidence state improves without regression.

Quantized by [Rodrigo Ramos](https://github.com/rodrigoramosrs).

## Quantization Approach

Produced with [NVIDIA Model Optimizer](https://github.com/NVIDIA/Model-Optimizer) using the canonical **`NVFP4_W4A4_WEIGHT_LOCAL_HESSIAN_CFG`** recipe: static per-block 4-bit weights (group size 16) + dynamic 4-bit activations, FP8 attention, local-Hessian calibration with MSE and fp8 scale sweep.

The `linear_attn` (Gated Delta Net) blocks and the self-attention projections ship in BF16 on purpose — matching validated official NVFP4 releases for this architecture family (which exclude `linear_attn*` per layer) and the widely-deployed MLP-only NVFP4 pattern. Only the MLP blocks (gate/up/down) are NVFP4. Embeddings, `lm_head` and small projections stay BF16 per the recipe.

Calibration ran on **512 samples × 512 tokens** (262144 tokens) of competitive-programming code from `nvidia/Nemotron-Competitive-Programming-v1`, keeping quantization faithful on the distributions that matter most for coding tasks.

The result retains the original model's strong software-engineering capabilities at roughly half of the BF16 footprint, ready for NVFP4-capable inference stacks on Blackwell GPUs.

## Files

| File | Size | Notes |
|---|---|---|
| `model-00001-of-00008.safetensors` … `model-00008-of-00008.safetensors` | ~29.2 GB total | NVFP4 MLP weights (packed U8 + FP8 block scales) + BF16 attention/GDN/embeddings/head |
| `model.safetensors.index.json` | — | Shard index (1427 tensors) |
| `hf_quant_config.json` | — | `quant_algo: NVFP4`, group size 16 |
| `config.json`, `tokenizer.*`, `merges.txt`, `vocab.json` | — | Config + tokenizer from the base model |
| `scripts/quantize_veriloop.py` | — | Full reproduction script (multi-GPU PTQ) |
| `scripts/validate_nvfp4.py` | — | Structural shard/index validator |

## Usage

This checkpoint targets **NVFP4-capable runtimes** (Blackwell GPUs). Plain `transformers` cannot load modelopt NVFP4 checkpoints directly.

> **Compatibility note:** per the NVIDIA recipe, the `linear_attn` blocks, the self-attention projections, a small set of projections (`linear_attn.in_proj_a/b`, `conv1d`), plus the embeddings and `lm_head`, intentionally ship in BF16. The `exclude_modules` list in `hf_quant_config.json` (mirrored in `config.json`, including fused names such as `in_proj_ba`/`in_proj_qkvz` that some loaders synthesize) declares exactly which modules those are — inference loaders must honor it and serve those modules unquantized. If a loader errors with *"weight is torch.bfloat16 but the quant config declares NVFP4"*, it is ignoring the exclusion list, not a checkpoint defect.

### vLLM

```bash
pip install 'vllm>=0.17.0'

vllm serve rodrigoramosrs/veriloop-coder-e2-nvfp4 \
    --tensor-parallel-size 1 \
    --max-model-len 131072 \
    --kv-cache-dtype fp8_e4m3 \
    --gpu-memory-utilization 0.92
```

### SGLang

```bash
pip install 'sglang[all]'

python -m sglang.launch_server \
    --model-path rodrigoramosrs/veriloop-coder-e2-nvfp4 \
    --trust-remote-code \
    --mem-fraction-static 0.88
```

### TensorRT-LLM

Build the engine from this checkpoint with `trtllm-build` using an NVFP4 + FP8-KV configuration, then serve with `trtllm-serve`. See the [TensorRT-LLM documentation](https://github.com/NVIDIA/TensorRT-LLM) for the exact flags for your GPU.

### Reproducing

```bash
pip install "nvidia-modelopt[all]" torch transformers accelerate datasets safetensors

# needs ~55 GB VRAM across CUDA GPUs for the 27B BF16 source model
python scripts/quantize_veriloop.py \
    --model ./model-bf16 \
    --output ./model-nvfp4 \
    --calib-size 512 \
    --calib-seq-len 512

python scripts/validate_nvfp4.py ./model-nvfp4
```

On Windows the Triton JIT needs a C compiler: run from a Visual Studio Native Tools prompt (`vcvars64`) with `CC` pointing at `cl.exe`. See the header of `scripts/quantize_veriloop.py` for multi-GPU layout tuning (`--layers-split`, `--gpu-order`).

## Runtime notes

- **Verified working:** FreeToken (NVFP4, Blackwell) and SGLang ≥ 0.5.20 with the loader patch below — end-to-end generation confirmed on both.
- **SGLang ≤ 0.5.20, stock:** serving this checkpoint yields degenerate output (constant-token repetition). Root cause, bisected: `qwen3_5_text.Qwen3_5ForCausalLM.load_weights` strips the leading `model.` before delegating to a body whose params are prefix-less, so essentially every weight misses (`... not found in params_dict` warnings) and the model runs uninitialized. Fix (`scripts/sglang-qwen35-loader.patch.py`, applied in `scripts/Dockerfile.sglang-qwen35fix`): strip `model.` plus an optional `language_model.` segment, yielding exactly the body-relative names for both `model.layers.*` and `model.language_model.*` checkpoints. Affects any text-only Qwen3.5 checkpoint on stock 0.5.20, quantized or not.

## Acknowledgements

- **Libo Wang** and the **Intelligent Robotics Laboratory, Tsinghua SIGS** for developing the original VeriLoop E2 model.
- **NVIDIA** for Model Optimizer and the NVFP4 quantization tooling.
- The original model repository: [tsinghua-sigs-robot-lab/VeriLoop-E2](https://huggingface.co/tsinghua-sigs-robot-lab/VeriLoop-E2)

## License

Apache-2.0. The weights are quantized from the original Apache-2.0 licensed model. See the [original repository](https://huggingface.co/tsinghua-sigs-robot-lab/VeriLoop-E2) for full licensing details and third-party notices.