Model outputs only "!!!!!!!!"
I fire up vLLM, the model loads beautifully across my dual RTX 3090s. Everything looks perfect. The weights are loaded, the KV cache is allocated, the stars are aligned. I send my first prompt: "Hello! How are you today?" The model responds: !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! Okay, maybe a bad prompt. Let me try something more elaborate. "Write a Python function to calculate the Fibonacci sequence." !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! "Explain quantum physics to a 5-year-old." !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! I tried every prompt I could think of — creative writing, coding, math, philosophy. The model had one answer for everything, and that answer was exclamation marks. Not even good exclamation marks. Just the same character, repeated infinitely, like a stuck typewriter having an existential crisis. ## Technical details - Model: cyankiwi/Qwen3.5-27B-AWQ-BF16-INT8 - Backend: vLLM 0.19.1rc1 (also tested on dev292) - Hardware: 2x RTX 3090 (48GB total VRAM) - Symptom: Model loads successfully, but every single prompt produces only !!! characters - Quantization: Attempted with --quantization compressed-tensors --enforce-eager — same result -