README.md

---
license: mit
base_model:
- deepseek-ai/DeepSeek-R1-0528
---


# Model Overview

- **Model Architecture:** DeepSeek-R1-0528
  - **Input:** Text
  - **Output:** Text
- **Supported Hardware Microarchitecture:** AMD MI350/MI355
- **ROCm**: 7.0
- **PyTorch**: 2.8.0
- **Transformers**: 5.0.0
- **Operating System(s):** Linux
- **Inference Engine:** [SGLang](https://docs.sglang.ai/)/[vLLM](https://docs.vllm.ai/en/latest/)
- **Model Optimizer:** [AMD-Quark](https://quark.docs.amd.com/latest/index.html) (V0.11)
    - **Base model:**
      - **Weight quantization:** self_attn Perchannel, FP8E4M3, Static; MOE OCP MXFP4, Static
      - **Activation quantization:** self_attn Pertoken, FP8E4M3, Dynamic; MOE OCP MXFP4, Dynamic
    - **Mtp:**
      - **Weight quantization:** self_attn Perchannel, FP8E4M3, Static; MOE OCP MXFP4, Static
      - **Activation quantization:** self_attn Pertoken, FP8E4M3, Dynamic; MOE OCP MXFP4, Dynamic
- **Calibration Dataset:** [Pile](https://huggingface.co/datasets/mit-han-lab/pile-val-backup)

This model was built with deepseek-ai DeepSeek-R1-0528 model by applying [AMD-Quark](https://quark.docs.amd.com/latest/index.html) for quantization.

# Model Quantization

The model was quantized from [deepseek-ai/DeepSeek-R1-0528](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528) using [AMD-Quark](https://quark.docs.amd.com/latest/index.html). Both weights and activations were quantized. 

**Preprocessing requirement:**

Before executing the quantization script below, the original FP8 model must first be dequantized to BFloat16.
You can either perform the dequantization manually using this [conversion script](https://github.com/deepseek-ai/DeepSeek-V3/blob/main/inference/fp8_cast_bf16.py), or use the pre-converted BFloat16 model available at [amd/DeepSeek-R1-0528-BF16](https://huggingface.co/amd/DeepSeek-R1-0528-BF16).

**Quantization scripts:**
```
cd Quark/examples/torch/language_modeling/llm_ptq/
export exclude_layers="*mlp.gate.* *lm_head model.layers.61.eh_proj model.layers.61.shared_head.head model.layers.61.embed_tokens"
python3 quantize_quark.py --model_dir amd/DeepSeek-R1-0528-BF16 \
                          --quant_scheme mxfp4 \
                          --layer_quant_scheme '*self_attn*' ptpc_fp8 \
                          --exclude_layers $exclude_layers \
                          --skip_evaluation \
                          --model_export hf_format \
                          --output_dir amd/DeepSeek-R1-0528-MXFP4-MTP-MoEFP4 \
                          --multi_gpu
```

### Accuracy

<table>
  <tr>
   <td><strong>Benchmark</strong>
   </td>
   <td><strong>DeepSeek-R1-0528</strong>
   </td>
   <td><strong>DeepSeek-R1-0528-MXFP4-MTP-MoEFP4(this model)</strong>
   </td>
  </tr>
  <tr>
   <td>GSM8K
   </td>
   <td>94.24
   </td>
   <td>94.90
   </td>
  </tr>
</table>

### Reproduction

Docker image: rocm/vllm-dev:base_main_20260212

Step 1: start a vLLM server with the quantized DeepSeek-R1 checkpoint

```bash
vllm serve amd/DeepSeek-R1-0528-MXFP4-MTP-MoEFP4 \
  --tensor-parallel-size 8 \
  --dtype auto \
  --speculative-config '{"method":"mtp","num_speculative_tokens":1}' \
  --gpu-memory-utilization 0.9 \
  --block-size 1 \
  --trust-remote-code \
  --port 8000
```
Note: CLI parameters such as `--tensor-parallel-size`, `--gpu-memory-utilization`, and `--port` can be adjusted as needed to match the target runtime environment.

Step 2: in a second terminal, run the GSM8K evaluation client against the running server.

```bash
python3 tests/evals/gsm8k/gsm8k_eval.py
```

# License
Modifications Copyright(c) 2025 Advanced Micro Devices, Inc. All rights reserved.