DGX Spark Recipes

This repository contains small deployment notes and Spark vLLM recipe files for running optimized LLM serving setups.

Contents

Path	Description
spark-vllm-docker/rdtand-Qwen3.6-27B-PrismaQuant-5.5bit-vllm	Qwen3.6 27B PrismaQuant vLLM recipe, FlashQLA setup notes, and chat-template fix.
spark-vllm-docker/rdtand-Qwen3.6-35B-A3B-PrismaQuant-4.75bit-vllm	Qwen3.6 35B A3B PrismaQuant vLLM recipe, chat-template files, and initial benchmark report.
spark-vllm-docker/nvidia-Qwen3.6-35B-A3B-NVFP4	Qwen3.6 35B A3B NVIDIA NVFP4 vLLM recipe, chat-template files, and initial benchmark report.

Current Recipes

The active recipe areas are:

• rdtand/Qwen3.6-27B-PrismaQuant-5.5bit-vllm
• rdtand/Qwen3.6-35B-A3B-PrismaQuant-4.75bit-vllm
• nvidia/Qwen3.6-35B-A3B-NVFP4

For example, start with:

cd spark-vllm-docker/rdtand-Qwen3.6-27B-PrismaQuant-5.5bit-vllm

See each recipe README for build steps, runtime defaults, mod setup, and chat-template benchmark notes:

Qwen3.6 27B PrismaQuant vLLM Notes

Qwen3.6 35B A3B PrismaQuant vLLM Notes

Qwen3.6 35B A3B NVIDIA NVFP4 vLLM Notes

License

See LICENSE.txt.