Skip to main content

Documentation Index

Fetch the complete documentation index at: https://www.radixark.com/llms.txt

Use this file to discover all available pages before exploring further.

In agentic / multi-turn workflows, Miles uses SGLang’s pretokenized prefix mechanism so the conversation history is not re-tokenized every turn. That requires the chat template to satisfy an append-only invariant: rendering messages [1..N] must produce a string that is an exact prefix of rendering [1..N+1]. Some community templates violate this. They use loop.last or other context-dependent Jinja logic that flips bits across turns, and the result is silent tokenization drift, divergent log-probabilities, and gradient blow-up after a few iterations of multi-turn RL. Miles ships a verifier and an autofix.

​
Quick start

​
Verify a HuggingFace template

python scripts/tools/verify_chat_template.py --model Qwen/Qwen3-0.6B
Failing output (illustrative): 
Template source: HuggingFace: Qwen/Qwen3-0.6B
Thinking cases:  disabled

  [FAIL] single_tool-N3                Prefix mismatch
  [PASS] single_tool-N3-no_tools
  [FAIL] multi_turn-N4                 Prefix mismatch
  ...
Verdict: FAIL - template is NOT append-only after last user message

​
Apply Miles’s autofix

If a fixed template ships for that model, --autofix swaps it in and re-runs the suite: 
python scripts/tools/verify_chat_template.py --model Qwen/Qwen3-0.6B --autofix

Template source: fixed template: .../templates/qwen3_fixed.jinja
Verdict: PASS - template IS append-only after last user message

​
Verify a local Jinja file

python scripts/tools/verify_chat_template.py --template path/to/my_template.jinja

​
Include thinking-specific cases

For Qwen3.5, GLM-5, and other models that toggle enable_thinking, add --thinking to also run thinking-specific trajectories. 
python scripts/tools/verify_chat_template.py --model Qwen/Qwen3.5-0.8B --autofix --thinking

​
CLI

usage: verify_chat_template.py (--template PATH | --model MODEL_ID)
                               [--autofix] [--thinking]
FlagWhat
--template PATHLocal .jinja template.
--model MODEL_IDHF model ID.
--autofixApply Miles’s fixed template if available.
--thinkingAlso run thinking-specific cases.
Exit code is 0 on pass, 1 on fail.

​
How it works

For each test case (a list of messages), the verifier renders progressive prefixes and checks the invariant character by character: 
for n in range(1, len(messages)):
    full   = render(messages[: n + 1])
    prefix = render(messages[: n])
    assert full.startswith(prefix), f"break between turn {n} and {n+1}"
The trajectory specs and cases live in miles/utils/test_utils/chat_template_verify.py. Standard cases cover single-tool, multi-turn, parallel-tool, and long-chain trajectories; the thinking suite adds variants that toggle enable_thinking. A break almost always comes from loop.last, conditional whitespace, or a closing token that’s only emitted on the final turn.

​
Using the fixed template at training time

Once you have the right template, point Miles at it: 
ROLLOUT_ARGS+=(
   --chat-template-path /opt/miles/utils/chat_template_utils/templates/qwen3_fixed.jinja
)
Built-in fixed templates that ship with Miles live under miles/utils/chat_template_utils/templates/ (e.g. qwen3_fixed.jinja, qwen3.5_fixed.jinja, qwen3_thinking_2507_and_next_fixed.jinja).

​
What “append-only” buys you

Without itWith it
Re-tokenize everything each turnTokenize only the new turn
O(N²) tokenization costO(N) tokenization cost
Subtle drift between turnsBit-stable tokens
Multi-turn RL collapses after ~50 stepsStable across thousands of steps
Running the verifier as part of every model’s pre-flight is recommended.