acoustic solar
Hey guys hope you are well, we're an AI company testing your cloud provisions for the first time and have run into some odd behaviour. Pods are great, however with serverless we seem to be running into strange behaviour.
We've deployed 4x v2.11.3 vLLM workers (via your templates) for Qwen3/GPT-OSS-120b with a variety of configs. All models output pure jibberish:
Here's one example:
Input:
{
"input": {
"prompt": "Hello How are you?",
"temperature": 1
}
}
Response:
" \n\n! This is a day? \nThis is a test? \n=\n(Defect ID: ) ""<Email>"" has been removed first."\n\nIt looks like you want to extract the content of an email, split by the separators ________________________`, and process the extracted content. This can be broken down into the following steps:\n\n1) Find and extract the messages based on the separators.\n2) Strip newlines if required.\n3) Process the extracted content (assigning to"
Given that these exact models, with the exact configs work perfectly on pods - could you please help?
Thanks
ionic prism