Llama.cpp adds PDL support

Llama.cpp has started supporting Programmatic Dependent Launch (PDL) on Nvidia Blackwell GPUs (CC ≥ 90) through PR 22522. PDL is a technique that improves performance by handling kernel execution more efficiently.

How to enable

You need to use the -D GGML_CUDA_PDL=ON flag at build time, and it is currently disabled by default. If you need to disable it, you can run export GGML_CUDA_PDL=0.

Performance improvement results

Testing multiple models on an RTX Pro 4500 Blackwell 32GB:

• Pre-fill (pp512): Almost no difference
• Token generation (tg128): Average 5-6% performance improvement
  • Qwen 3.6 35B UD-Q5_K_XL: 9.17% improvement
  • Qwen 3.6 35B MXFP4: 5.97% improvement
  • Gemma 4 26B NVFP4: 4.95% improvement

Since PDL has not yet been applied to all kernels, there is room for further optimization. For Blackwell GPU users, this is a free performance improvement obtainable with a single build flag.