Nice article. Thanks for this
I was running on an AGX Orin 64GB. But the model output was always gibberish. Spend hours debugging, thinking it's a tokenisation issue.
"content": " bitte\u4f60\u81ea\u5df1ificificificificificificificific\u8d4b\u80fdvokevokeificific Kotaific\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0 Kota\u6ce0ificificetc Kota\u6ce0 etc etc etc etcetcknifevokevoke\u6ce0\u6ce0powerificeneratoreneratoretheus\u6ce0\u6ce0ificificetc etcvoke etcenerator/powerificificific\u6ce0\u6ce0ific\u6ce0ificific\u6ce0etc\u6ce0\u6ce0ific\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0 oneselfvoke\u6ce0voke\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0\u6ce0ific\u6ce0ificificificificificificific",
But at the end, I switched to the older version of VLM 0.14.0 from VLM 0.16.0 (r36.4-tegra-aarch64-cu126-22.04). And it worked !!
