YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
DeepSeek-R1-Distill-Qwen-1.5B - RK3576 W4A16
✅ 兼容性确认
| 属性 | 值 |
|---|---|
| 平台 | RK3576 ✅ |
| 量化 | W4A16 ✅ |
| 文件大小 | 1.3 GB |
| 内存需求 | ~2.5 GB (含Swap) |
| NPU核心 | 1-2个 |
📥 使用方法
1. 下载到RK3576设备
huggingface-cli download JiahaoLi/DeepSeek-R1-Distill-Qwen-RK3576 \
DeepSeek-R1-Distill-Qwen-1.5B_W4A16_RK3576.rkllm \
--local-dir ./models/DeepSeek-R1-1.5B
2. 运行推理
cd ./models/DeepSeek-R1-1.5B
rkllm DeepSeek-R1-Distill-Qwen-1.5B_W4A16_RK3576.rkllm
3. Python API
from rkllm.api import RKLLM
llm = RKLLM()
llm.load_rkllm("DeepSeek-R1-Distill-Qwen-1.5B_W4A16_RK3576.rkllm")
llm.build(target_platform='rk3576', num_npu_core=2)
# 测试推理
response = llm.inference(["你好"])
print(response)
📊 性能预期
| 指标 | 预期值 |
|---|---|
| 推理速度 | 15-20 tokens/s |
| 内存占用 | ~2.5 GB |
| 上下文长度 | 4096 tokens |
| 量化精度 | W4A16 (4-bit权重, 16-bit激活) |
⚠️ 系统要求
- 硬件: Rockchip RK3576 (6 TOPS NPU)
- 内存: 至少 4 GB RAM + 2 GB Swap
- 驱动: RKNPU v0.9.6+
- 运行时: RKLLM Runtime 1.2.3+
🔗 相关链接
- 原始仓库: https://huggingface.co/JiahaoLi/DeepSeek-R1-Distill-Qwen-RK3576
- DeepSeek-R1: https://github.com/deepseek-ai/DeepSeek-R1
- RKLLM Toolkit: https://github.com/airockchip/rknn-llm
📝 许可证
MIT License (继承自DeepSeek-R1)
标注时间: 2026-08-04 验证状态: ✅ 文件头验证通过 (daee b336)
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support