DeepSeek R1 Distill Qwen 32B
Model Overview¶
DeepSeek-R1 and its distilled models represent a significant advancement in reasoning capabilities for LLMs by combining RL, SFT, and distillation. The DeepSeek-R1-Distill-Qwen-32B is a distilled version of the Qwen-32B large language model (LLM), optimized for efficient performance while retaining high-quality generative capabilities and is particularly suited for scenarios where computational efficiency is critical.
- Model Architecture: DeepSeek-R1-Distill-Qwen-32B is based on a transformer architecture, distilled from the larger Qwen-32B model to reduce computational requirements while maintaining competitive performance. The distillation process ensures that the model retains the core capabilities of the original model, making it suitable for a wide range of text generation tasks.
- Repository: DeepSeek-V3
- Model Source: deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
- License: MIT License.
QPC Configurations¶
| Precision | SoCs / Tensor slicing | NSP-Cores (per SoC) | Full Batch Size | Chunking Prompt Length | Context Length (CL) | Generated URL | Download |
|---|---|---|---|---|---|---|---|
| MXFP6 | 4 | 8 | 8 | 128 | 8192 | https://dc00tk1pxen80.cloudfront.net/SDK1.20.2/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B/qpc_8cores_128pl_8192cl_8bs_4devices_mxfp6_mxint8.tar.gz | Download |
Run This Model¶
# Download QPC
mkdir -p deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
cd deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
wget <Download URL>
tar xzvf <downloaded filename.tar.gz>
# Run QPC
python3 -m QEfficient.cloud.execute --model_name deepseek-ai/DeepSeek-R1-Distill-Qwen-32B --qpc_path <path/to/qpc> --prompt "# shortest path algorithm\n" --generation_len 128
API Endpoint¶
# Start REST endpoint with vLLM
VLLM_QAIC_MAX_CPU_THREADS=8 VLLM_QAIC_QPC_PATH=/path/to/qpc python3 -m vllm.entrypoints.openai.api_server \
--host 0.0.0.0 \
--port 8000 \
--model deepseek-ai/DeepSeek-R1-Distill-Qwen-32B \
--max-model-len <Context Length> \
--max-num-seq <Full Batch Size> \
--max-seq_len-to-capture <Chunking Prompt Length> \
--device qaic \
--block-size 32