[Bug] Ollama endpoint /api/chat fails with empty chunks when stream: true (OpenAI endpoint works) #436

Environment

• OS: Windows 11
• Hardware: AMD Ryzen AI 300 / Strix Point (Radeon 860M NPU)
• FastFlowLM Version: v0.9.36
• Model: qwen3.5:4b (also tested with qwen3:0.6b)
• Command: flm serve qwen3.5:4b --port 8080 --ctx-len 32768 --pmode turbo -r 1 --q-len 10

Description

When using the Ollama-compatible endpoint (/api/chat) with "stream": true, the server instantly returns empty SSE chunks containing only metadata (no content) and immediately closes the stream with "done": true.

This completely breaks tools like Claude Code (via proxies like claudish or LiteLLM) that rely on the Ollama streaming API.

Steps to Reproduce:
Run a simple streaming request against the Ollama endpoint:

curl -s -N -X POST http://localhost:8080/api/chat -H "Content-Type: application/json" -d "{\"model\": \"qwen3.5:4b\", \"messages\": [{\"role\": \"user\", \"content\": \"Hello, reply with 'hi'.\"}], \"stream\": true}"

Observed Behavior (Empty Chunks):

{"model":"qwen3.5:4b","message":{"role":"assistant","content":""},"done":true,"done_reason":"stop","prompt_eval_count":24,"eval_count":0,"total_duration":1966916608,"load_duration":4300,"prompt_eval_duration":964379584,"eval_duration":0}

Expected Behavior:
The stream should return the generated text chunk by chunk.
Note: The OpenAI-compatible endpoint (/v1/chat/completions) works perfectly fine with stream: true on the exact same model. Also, setting "stream": false on the Ollama endpoint works correctly.