[rromenskyi]

A request body can carry recurrent_loops (1-4) for a looped model (Dory). No value keeps the default: the model's config, or --recurrent-loops.

• The loop count is the last element of ModelProvider.model_key for looped models. Prompt cache entries, warm-next jobs and handoffs are all keyed on it, so a cache made with 2 loops is never reused for 3. Their layouts differ, and so does the output-layer KV.
• ModelDescription.loops: a request with another count doesn't join the running batch. The batch drains, then load() calls model.set_recurrent_loops(n). Weights aren't reloaded.
• dory.Model.set_recurrent_loops resets the cache plan.
• Validation: an integer 1-4, else 400.

Tests (tests/test_looped_model_config.py):

• the provider switches loops and back without reloading;
• --recurrent-loops stays the default;
• a switched model gives the same logits as one loaded with that count;
• over HTTP, real ModelProvider, tiny Dory with a Qwen tokenizer: requests with 3, none and 1 loops, and bad values get 400.

test_server, test_warm_next_turn and test_server_handoff pass.

🤖 Generated with Claude Code