feat: add Llama-3 renderer for Llama-3.2-1B/3B-Instruct

[hallerite]

Summary

Hand-coded Llama3Renderer for Meta's Llama-3.x chat template, plus matching parse_llama_3 parser. Initial scope: Llama-3.2-1B-Instruct and Llama-3.2-3B-Instruct (auto-routed via MODEL_RENDERER_MAP). No version bump.

How tests work without a Meta-license HF token

MODEL_RENDERER_MAP registers the canonical meta-llama/... paths so production callers auto-route. Tests load the tokenizer via the unrestricted unsloth/Llama-3.2-{1B,3B}-Instruct mirror — the chat-template SHA matches Meta's bit-for-bit and the underlying tiktoken-BPE files are identical. CI doesn't need an HF_TOKEN with Meta license access.

Implementation notes

• No <think> / reasoning channel — Llama-3 doesn't ship one. preserve_*_thinking constructor flags raise NotImplementedError if set (matches DefaultRenderer's contract for the same case).
• <|begin_of_text|> (BOS) is emitted at the start of every render; system block is always emitted with the fixed Cutting Knowledge Date / Today Date preamble even when no system message is supplied.
• date_string is a constructor kwarg, defaulting to "26 Jul 2024" (the chat template's strftime fallback) so output stays deterministic. Override per-instance for production runs that want today's date.
• tools_in_user_message defaults to True (matches chat template). Tools + JSON signatures inject into the first user message; pass False to flip to system-block mode. Both modes parity-tested.
• Single tool call per assistant message (chat template raises otherwise). Tool calls render as a JSON blob {"name": "...", "parameters": ...} inside the assistant body. Tool responses render under role ipython regardless of source role; mirrors the chat template's content | tojson branch — including the Jinja quirk that strings are iterable, so plain-string tool content gets JSON-quoted.
• parse_llama_3 detects the JSON tool-call body shape with a strict starts-with-{ + parses-as-dict-with-name check; malformed JSON falls through to content rather than dropping silently.

Tests

47 dedicated tests in tests/test_llama_3.py:

• MODEL_RENDERER_MAP shape + factory routing
• Constructor contract (default date, preserve_*_thinking rejection, tools_in_user_message toggle)
• Byte parity vs apply_chat_template across 11 conversation shapes (system + user, user-only, multi-turn, gen prompt, whitespace trimming, custom date, tools-in-user, tools-in-system, tool call round-trip, dict tool response, multiple-tool-calls rejection)
• parse_response (plain, tool call, malformed JSON fallthrough)
• Bridge contract (extends prev verbatim, matches fresh render, rejects assistant in extension, synthesises close on truncation)

Test plan

• pytest tests/test_llama_3.py — 47 cases pass on both 1B and 3B mirrors
• Full suite (pytest tests/ --ignore=tests/test_client.py) — 947 pass, 48 skipped, 1 xfailed (no regressions)
• Pre-commit hooks (ruff check + format) clean
• Maintainer with Meta-license HF_TOKEN can verify meta-llama/Llama-3.2-1B-Instruct parity directly (the unsloth mirror has been bit-verified, but a once-off canonical run is good defense in depth)

🤖 Generated with Claude Code