server: add /v1/responses support

[RodriMora]

Summary

• add /v1/responses endpoint by converting Responses payloads to chat-completions and emitting Responses-style SSE events
• track response IDs/state in server slots and serialize Responses output/final payloads
• document /v1/responses usage and add responses compatibility scenarios

Testing

• cmake --build build --target llama-server (works too with CUDA enabled on my system)
• behave -i server.feature -n "OAI Responses Compatibility"

Notes

• i tried to run the full server test suite but it does not run, embeddings scenario fails on main due to /embedding response shape. This seems to be failing too on the main branch, I can look into those in a separate pr

Curently as a draft as I do more test. Currently works via regular curl

curl http://localhost:5000/v1/responses \
                       -H "Content-Type: application/json" \
                       -H "Authorization: Bearer no-key" \
                       -d '{
                     "model": "MiniMax-M2.1",
                     "input": "Say hello in one short sentence."
                   }'

Based off ggml-org/llama.cpp#18486

Edit: Used an AI agent to help create the draft, as this seems like a good use case of an implementation for it as it doesn't involve a new model or architecture which seems to yield bad results. I didn't see any contributing rules against it. If that's the case feel free to close it.

[firecoperana]

I don't think these tests are well maintained. Just need to test in llama-server

[ikawrakow]

This is still draft?

[RodriMora]

Tool calling doesn't work perfectly but I checked mainline llama.cpp and it has the same problems, so it has feature parity in terms of regular text completion for streaming and non-streaming test:

Launch command:

./build/bin/llama-server \
                                                                  --model /mnt/llms/models/bartowski/MiniMaxAI_MiniMax-M2.1-GGUF/MiniMaxAI_MiniMax-M2.1-Q6_K/MiniMaxAI_MiniMax-M2.1-Q6_K-00001-of-00005.gguf \
                                                                  --alias "MiniMax-M2.1" \
                                                                  --ctx-size 128000 \
                                                                  -ger \
                                                                  -ngl 99 \
                                                                  --host 0.0.0.0 \
                                                                  --port 5000 \
                                                                  --jinja -np 4

Test:

curl -s http://192.168.10.115:5000/v1/responses \
                         -H "Content-Type: application/json" \
                         -H "Authorization: Bearer no-key" \
                         -d '{
                       "model": "MiniMax-M2.1",
                       "instructions": "You are a helpful assistant.",
                       "input": "Write a limerick about exceptions",
                       "max_output_tokens": 32
                     }'
{"completed_at":1770028662,"created_at":1770028662,"id":"resp_L1bEyCME6xhrIk2ePfniibiVoa6s6zNR","model":"MiniMax-M2.1","object":"response","output":[{"id":"rs_F0JvQ8MRkVid9hpBBHxxUzzj51zXrK1L","summary":[],"type":"reasoning","content":[{"text":"The user wants a limerick about exceptions. A limerick is a humorous five-line poem with an AABBA rhyme scheme, where lines 1","type":"reasoning_text"}],"encrypted_content":"","status":"completed"}],"status":"completed","usage":{"input_tokens":29,"output_tokens":32,"total_tokens":61}}

[firecoperana]

The model name is defaulted to "gpt-3.5-turbo-0613" when I leave the model empty in the request message. Mainline returns the correct model name. Otherwise, they look good.

[RodriMora]

The model name is defaulted to "gpt-3.5-turbo-0613" when I leave the model empty in the request message. Mainline returns the correct model name. Otherwise, they look good.

Nice catch, thanks. If model is empty/missing now, it uses the loaded model name instead of defaulting to "gpt-3.5-turbo-0613".

[firecoperana]

Yeah, it's a bug not related to this PR. I will fix later.

[firecoperana]

It is fine to merge now.