README.md

---
library_name: vllm
license: mit
tags:
  - foundation
  - amd
  - rocm
  - text-generation
pipeline_tag: text-generation
---

![](https://huggingface.co/AMD-PAVS-AI/deepseek_r1/resolve/main/Deepseek-R1.png)

# DeepSeek-R1: Optimized for AMD ROCm

DeepSeek-R1-Distill-Qwen-7B is a distilled reasoning language model that generates chain-of-thought answers for math and logic problems. This repository packages evaluation/inference for text reasoning / math problem solving using vLLM, exported and validated for **AMD ROCm** so it runs efficiently on AMD GPUs and CPUs.

This is based on the implementation of DeepSeek-R1 found [here](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B).
This repository contains configurations and scripts optimized for **AMD® ROCm™** platforms. You can use the [deepseek_r1 AMD scripts](https://github.com/AMD-PAVS/physical_ai_sdk/blob/main/models/deepseek_r1) to reproduce results or export with custom configurations. More details on model performance can be found [here](#accuracy-pipeline).

---

## Task Overview

**Task:** Text reasoning / math problem solving

**Dataset:** MATH-500 (500 competition math problems); sample prompts for `infer-text`

**Output metrics:** MATH-500 accuracy

> **vLLM note:** MATH-500 evaluation uses symbolic answer verification through `math_verify`, which parses `\boxed{}` expressions and checks equivalence to the reference solution.

---

## AMD ROCm Optimization

This model export has been adapted and validated for **AMD Instinct™ / Radeon™ GPUs** running **ROCm**, as well as AMD CPUs. Key points:

- Exported/tested with ROCm `7.2` and vLLM ROCm build `0.19.1` (built from source).
- Validated backends: **vLLM** (ROCm-enabled build).
- No code changes required versus the upstream DeepSeek-R1 implementation — only environment/runtime configuration differs.

| Runtime | Precision | Backend | Hardware | Notes |
|---|---|---|---|---|
| GPU | FP32 / FP16 / BF16 | vLLM | AMD RYZEN AI MAX+ 395 w/ Radeon 8060S | `VLLM_ROCM_USE_SKINNY_GEMM=0` set to avoid bf16/fp16 GEMM segfaults on gfx1151 |

---

## Getting Started

For setup instructions, evaluation scripts, and custom configuration options, see the [deepseek_r1 on GitHub](https://github.com/AMD-PAVS/physical_ai_sdk/blob/main/models/deepseek_r1).

---

## Model Details

**Model Type:** Distilled reasoning language model (text generation)

**Base Model:** `Qwen/Qwen2.5-7B` (Qwen2.5-7B)

**Model Stats:**
- Model variant: DeepSeek-R1-Distill-Qwen-7B
- Number of parameters: `7B`
- Precision tested: FP32, FP16, BF16

---

## Accuracy Pipeline

Higher MATH-500 accuracy means the model produces mathematically equivalent answers to ground truth more often — 100% would be perfect, ~0% is chance-level. Strong distilled reasoning models typically score ~85–95% on this benchmark.

### Metrics Explained

| Metric | Description |
|--------|-------------|
| MATH-500 Accuracy | Primary metric — fraction of problems where the model's final boxed answer is symbolically equivalent to the reference solution. |

### Accuracy Results

**Full Dataset Evaluation (MATH-500)** — filled from `deepseek-ai/DeepSeek-R1-Distill-Qwen-7B`; run `make metrics` to refresh:

<!-- accuracy-table-start -->
| Device | Backend | Precision | Variant | Accuracy (%) |
|--------|---------|-----------|---------|--------------|
| GPU | vLLM | FP32 | DeepSeek-R1-Distill-Qwen-7B | 90.00 |
| GPU | vLLM | FP16 | DeepSeek-R1-Distill-Qwen-7B | 90.00 |
| GPU | vLLM | BF16 | DeepSeek-R1-Distill-Qwen-7B | 90.00 |
<!-- accuracy-table-end -->

---

## Dig Deeper

Want to explore the full evaluation scripts, config options, and other AMD-optimized model examples?

📂 **[View the full project on GitHub](https://github.com/AMD-PAVS/physical_ai_sdk/blob/main/models/deepseek_r1)**

The GitHub repository includes:
- Setup and prerequisites for ROCm environments
- Scripts for the supported runners
- Additional model variants and datasets
- Benchmarking and reproduction instructions