anonymize authors

README.md

@@ -260,77 +260,4 @@ cd /root/prima.cpp
 (prima-v4) ./llama-cli -m download/qwq-32b-q4_k_m.gguf -c 1024 --world 4 --rank 3 --prefetch --gpu-mem 8
 ``` 
 
-> If your host machine does not have a GPU, ignore the `--gpu-mem` option.
-
-## ❓ FAQ
-
-**1. How can I manually set the workload for each device?**
-
-By default, prima.cpp automatically profiles devices and assigns workloads. However, if you want to manually control the layer distribution, you can use the `-lw` (or `--layer-window`, `--n-layer-window`) and `-ngl` options:
-
-```shell
-# on head device without a GPU, rank 0, use the option "-lw":
-./llama-cli -m download/qwq-32b-q4_k_m.gguf -c 1024 -n 256 -p "what is edge AI?" --world 4 --rank 0 --master 192.168.1.2 --next 192.168.1.3 --prefetch -lw "16,16,16,16"
-
-# on worker device with 8 GiB VRAM, rank 1, use the option "-ngl":
-./llama-cli -m download/qwq-32b-q4_k_m.gguf -c 1024 --world 4 --rank 1 --master 192.168.1.2 --next 192.168.1.4 --prefetch -ngl 16
-
-# on worker device with 11 GiB VRAM, rank 2, use the option "-ngl":
-./llama-cli -m download/qwq-32b-q4_k_m.gguf -c 1024 --world 4 --rank 2 --master 192.168.1.2 --next 192.168.1.5 --prefetch -ngl 16
-
-# on worker device without a GPU, rank 3:
-./llama-cli -m download/qwq-32b-q4_k_m.gguf -c 1024 --world 4 --rank 3 --master 192.168.1.2 --next 192.168.1.2 --prefetch
-```
-
-- `-lw` sets the total model layers each device should handle. The format is a comma-separated list, one value per device, in rank order. You can also set `"8,8,8,8"`, `"4,4,4,4"`, `"16,16,24,8"`.
-- `-ngl` sets how many of those model layers should run on the GPU. 
-
-> Example: if `-lw "16,16,16,16"` is passed to the head device, then each of the 4 devices will handle 16 model layers. A worker with `-ngl 8` (if a GPU is available) will run 8/16 layers on the GPU.
-
-**2. How to run in chat mode like in llama.cpp?**
-
-To enable chat (conversation) mode, simply add the `-cnv` flag on the head device:
-
-```shell
-# on head device, rank 0, use the option "-cnv":
-./llama-cli ... --rank 0 -p "You are an AI assistant" -cnv
-```
-
-To quit the chat mode, input `quit` or `exit`.
-
-**3. How to force prefetching after computing?**
-
-By default, prima.cpp only advises the OS to prefetch upcoming layer weights. The actual prefetching is then scheduled and handled by the OS, which may introduce some uncertainty. To explicitly trigger prefetching right after computing, you can use the `--force` flag on each device:
-
-```shell
-# on each device, use the option "--force":
-./llama-cli ... --prefetch --force
-```
-
-This enables more aggressive overlap but also introduce extra memory access latency. Use `--force` only after testing, as its effect depends on your hardware and OS behavior.
-
-**4. Does it support Windows?**
-
-Not yet—but it's on the roadmap. Currently, prima.cpp can run on Linux, macOS, Android and HarmonyOS (via Termux). You can mix heterogeneous devices in the cluster.
-
-**5. Does it support Vulkan or AMD GPUs?**
-
-Not yet. Now prima.cpp supports only CUDA-based GPUs. Vulkan is in our roadmap, and AMD GPUs will be supported once we have that device.
-
-## ❤️ Acknowledgment
-This project builds upon the incredible work from the open-source community, especially [ggml, gguf](https://github.com/ggml-org/ggml), and [llama.cpp](https://github.com/ggml-org/llama.cpp). We gratefully acknowledge their contributions.
-
-## 📚 Cite Us
-If you find this work helpful, please do not hesitate to cite us and send a star! 🤩
-
-```bibtex
-@misc{li2025primacpp,
-    title={PRIMA.CPP: Speeding Up 70B-Scale LLM Inference on Low-Resource Everyday Home Clusters}, 
-    author={Zonghang Li and Tao Li and Wenjiao Feng and Mohsen Guizani and Hongfang Yu},
-    year={2025},
-    eprint={2504.08791},
-    archivePrefix={arXiv},
-    primaryClass={cs.DC},
-    url={https://arxiv.org/abs/2504.08791}, 
-}
-```
+> If your host machine does not have a GPU, ignore the `--gpu-mem` option.