Baidu’s LLM suite trains DeepSeek on GPUs, NPUs, XPUs, and DCUs

What it does PaddleNLP is Baidu’s large-language-model toolkit built atop the PaddlePaddle framework. It covers the full life cycle—training, fine-tuning, compression, and inference—for models like DeepSeek-R1, Qwen3, and Llama, with a heavy emphasis on multi-hardware support. The suite wraps 4D distributed training, speculative decoding, and quantization into a single set of APIs that claim to let you swap between NVIDIA GPUs, Kunlun XPUs, Ascend NPUs, and other accelerators without touching low-level code.

The interesting bit The project is aggressively hardware-agnostic at a time when most LLM tooling assumes CUDA. Its Unified Checkpoint system lets you resume training across different machine topologies, and the README touts single-machine DeepSeek-R1 inference at over 1000 tokens/s with FP8 quantization—numbers that, if accurate, suggest the backend optimizations are more than skin-deep.

Key highlights

• Supports training and inference across NVIDIA GPU, Kunlun XPU, Ascend NPU, Suiyuan GCU, and Haiguang DCU from one interface.
• Claims 4D distributed training (data, sharded data, tensor, pipeline parallelism) configurable via a Trainer API.
• Unified Checkpoint compresses storage by 78.5% and supports async saves; training can resume on differently sized clusters.
• Ships with FP8/INT8/4-bit quantization and MTP speculative decoding, claiming >2100 tokens/s for 4-bit DeepSeek-R1 on a single node.
• Includes niche tools like PP-UIE for 8K-context information extraction and MergeKit for model fusion.

Caveats

• The README is Chinese-first; English docs exist but the latest feature announcements and community activity are predominantly in Mandarin.
• It is tightly bound to the PaddlePaddle ecosystem, which is distinct from the PyTorch-based stacks it benchmarks against (e.g., transformers and LLama-Factory).
• Several headline features are still in v3.0 beta, and some performance claims link to external benchmarks rather than showing methodology inline.

Verdict Worth evaluating if you need a single toolkit that spans training to quantized inference across non-NVIDIA hardware. Give it a pass if you are unwilling to adopt the PaddlePaddle framework, since the suite is firmly inside Baidu’s ecosystem.