Skip to main content

​
Agentic Chat Templates (TITO)

Multi-turn agentic rollout in Miles runs on TITO (Token-In-Token-Out): each turn’s token sequence is a bit-perfect prefix of the next, so the trainer sees exactly the tokens the engine produced — no re-tokenization, no drift. The why is in the blog (No Token Left Behind); this page is how. Your harness only ever sends and receives OpenAI chat messages, never tokens. Miles keeps the per-trajectory append-only token buffer (ids + logprobs + routed experts) internally and ships it straight to training.

​
Prerequisites

Your rollout loop must keep two invariants, or TITO is rejected at runtime:
  • Append-only messages. Each turn = previous messages + new ones on the tail; past turns are never edited. The only exception is retrying the latest turn — a single-step rollback to the last assistant checkpoint. Diverging earlier, or rolling back more than one turn, is rejected.
  • Declared roles match the flag. --tito-allowed-append-roles must list exactly the roles your harness appends after the first assistant turn (tool is always implied) — Miles resolves a prefix-stable template for that exact set, and appending any role outside it is rejected at runtime.

​
Pick your --tito-model

No auto-detection — pick the family matching your model. Miles then auto-resolves the fixed template from (--tito-model, --tito-allowed-append-roles); pass --chat-template-path only to override. The table below lists common models; the full set and verified role surfaces live in issue #712.
Your model--tito-modelMax --tito-allowed-append-roles
Qwen3qwen3tool user
Qwen3.5qwen35tool user
GLM-4.7 / GLM-5glm47tool user system
NVIDIA Nemotron 3 Super / Ultranemotron3tool user system
Kimi K2.5 / K2.6kimi25 / kimi26tool user
DeepSeek-V3.2 / V4deepseekv32 / deepseekv4tool
anything elsedefaulttool
More models: issue #712.

​
Turn it on

ROLLOUT_ARGS+=(
   --use-session-server          # entry point; required for the two flags below
   --hf-checkpoint Qwen/Qwen3-4B
   --tito-model qwen3
   --tito-allowed-append-roles tool user
)

​
Example

A full multi-turn agentic setup on the session-server TITO path lives in examples/experimental/swe-agent-v2: its launchers wire --use-session-server + --tito-model glm47 + --tito-allowed-append-roles user tool against a real SWE agent.

​
Add a new model

Models in the table are verified by Miles maintainers — just pick the family. To support a new model (or a new append-role surface), register a TITOTokenizer subclass plus its fixed Jinja template (or HF-native + kwargs) and SUPPORTED_TEMPLATES rows in tito_tokenizer.py, then verify with both scripts — either failing blocks it. Each prints Verdict: PASS/FAIL. 
# CPU / fast — rendered token sequence is append-only
python scripts/tools/verify_chat_template.py \
    --model <hf-id> --tito-model <family> --tito-allowed-append-roles tool user

# GPU / e2e — still holds under real model inference
python scripts/tools/verify_session_tito_tokenizer.py \
    --hf-checkpoint <hf-id> --tito-model <family> --tito-allowed-append-roles tool user \
    --sglang-reasoning-parser <rp> --sglang-tool-call-parser <tcp> --rollout-num-gpus-per-engine 1