[NVolcz]

normalize_messages merges consecutive user messages, and consecutive system messages, into one message, and puts nothing between their texts, so the words of two messages run together:

<|turn>user
Hi.What is 2+2?<turn|>

Clients do send such messages. For example, Pi drops a reply the user aborted, which leaves two user messages in a row.

The messages are now passed on as they are, and the chat template decides how to render them. This is what llama.cpp, vLLM, MLX-LM, TGI and transformers do; Ollama merges them only for models on its older template path, and renders them as separate turns for Gemma 4. Gemma 4's template renders separate turns:

<|turn>user
Hi.<turn|>
<|turn>user
What is 2+2?<turn|>

A template that requires the roles to alternate, such as Gemma 3's, now rejects such a request, as it does on llama.cpp and vLLM.

How to reproduce

Send two user messages in a row, then ask about the second one:

curl -s http://localhost:52625/v1/chat/completions -H 'Content-Type: application/json' -d '{
  "model": "gemma4-it:e4b", "temperature": 0, "think": false, "max_tokens": 20,
  "messages": [
    {"role": "user", "content": "Apples are red."},
    {"role": "user", "content": "Bananas are yellow."},
    {"role": "user", "content": "What is the first word of my second message? Answer with that word only."}
  ]}' | jq -r '.choices[0].message.content'

On main the answer is What: the model sees one message, "Apples are red.Bananas are yellow.What is the first word…". On this branch it is Bananas.

I found this while reading the prompts flm renders, with a small debug branch that logs every prompt to a file (log-prompt, FLM_LOG_PROMPT=<file>).