Address streaming responses incongruences #7180
Description
LocalAI version:
3.7.0
Environment, CPU architecture, OS, and Version:
Describe the bug
It seems that the way streaming is implemented in LocalAI breaks some clients, such as sst/opencode#3994 , and see also: #1194
To Reproduce
N/A yet
Expected behavior
Both opencode and other software using streaming token should work
Logs
Additional context
- A list of things that doesn't seem to work are noted in task update but no model output with LocalAI / Qwen3 8B sst/opencode#3994 (comment)
- Streaming implementation is in https://github.com/mudler/LocalAI/blob/master/core/http/endpoints/openai/chat.go for OpenAI