💥 One RTX 5090. Qwen3.8 NVFP4. vLLM OOMs on start. Not weights. CUDA graph cap… · @QwenAIFans · TwKit

@QwenAIFans

💥 One RTX 5090. Qwen3.8 NVFP4. vLLM OOMs on start. Not weights. CUDA graph capture wants 784 MiB outside utilization. Raise or lower it; ~47 MiB stays free. Add --enforce-eager and it boots. At 32K: about 91k KV tokens. via vLLM Recipes 👇 https://recipes.vllm.ai/Qwen/Qwen3.8-27B

曝光 158 · 评论 1 · 点赞 4 · 书签 0