vllm-ultimate-dgx-spark

AEON-7

AEON vLLM Ultimate — vLLM 0.24.0 built from source for DGX Spark / Blackwell (sm_121a/GB10). One image serves the whole AEON fleet (Gemma-4-26B-A4B, Qwen3.6-27B, Qwen3.6-35B-A3B) with DFlash speculative decoding, NVFP4 weights + Triton NVFP4/FP8 KV cache, high-concurrency DFlash fix, UMA cudagraph clamp, FlashInfer 0.6.12.