[Bug] `/v1/chat/completions` defaults to streaming when `stream` is omitted

[lipchev]

Summary

When calling the OpenAI-compatible /v1/chat/completions endpoint without explicitly specifying the stream parameter, the server appears to default to stream: true.

This behavior is not aligned with the OpenAI API convention, where omitting stream defaults to non-streaming (false).

As a result, OpenAI-compatible clients (e.g. .NET SDKs) that omit the stream field receive SSE (data:) responses instead of JSON, causing deserialization failures.

---

Steps to reproduce

1. Load a model via:

POST /api/inference/load

2. Call:

POST /v1/chat/completions

with body:

{
  "messages": [
    { "role": "user", "content": "Hello" }
  ]
}

(note: no stream field)

---

Expected behavior

• Response is standard JSON:

{
  "choices": [...]
}

• Equivalent to:

"stream": false

---

Actual behavior

• Response is streamed using SSE format:

{"id":"chatcmpl-bec20bc3c001","object":"chat.completion.chunk" ...}
{"id":"chatcmpl-bec20bc3c001","object":"chat.completion.chunk" ...}

• This causes failures in clients expecting JSON (e.g. System.Text.Json errors like):

'd' is an invalid start of a value

---

Additional context

• Explicitly setting:

"stream": false

produces correct JSON output.

• Explicitly setting:

"stream": true

produces expected streaming output.

• The issue only occurs when the stream field is omitted.

---

Impact

• Breaks compatibility with OpenAI SDKs and abstractions (e.g. Microsoft .NET clients), which omit stream when non-streaming is intended.
• Requires consumers to always explicitly set "stream": false, which is not expected behavior.

---

Suggested fix

Default stream to false when the field is not provided, to match OpenAI API behavior.

---

Environment

• Unsloth Studio (local)
• Endpoint: http://localhost:8888/v1/chat/completions
• Tested via Postman and .NET OpenAI client

---

Thanks for the great work on the OpenAI-compatible API — this small fix would significantly improve interoperability.

[Imagineer99]

Thanks for sharing, looks valid, I will investigate.

[rolandtannous]

I will push a separate PR for this