Gemma4 NPU2 conversions predate Google's July 2026 template fixes — porting notes for a refresh (2 parser shims) #634

Summary

Google shipped a canonical chat-template fix wave for Gemma 4 in July (template header: "Fixed tool-calling loops, turn closures, and thinking content-ordering. Published 2026-07-09"; plus a 2026-07-20 response_template addition to tokenizer_config.json). The FastFlowLM/Gemma4-E4B-IT-NPU2 conversion (last modified 2026-04-28) still carries the April template, which has the tool-call-swallowing x-regex (content captured before tool_calls) and the O(n) continuation scan.

We grafted Google's July files into the local model folder and verified they work through flm serve — but two small adjustments were needed for FLM's parsers, which you'd presumably want in the refreshed conversion. Sharing both so the update is a drop-in.

Porting note 1 — Jinja: adjacent string literals

Google's July chat_template.jinja uses Python-style adjacent string literals inside a raise_exception(...) call (one message split across three quoted lines). FLM's template engine rejects it:

Failed to load model: Expected closing parenthesis in call args at row 259, column 29

Fix: join the three literals into a single one-line string. Semantics unchanged; that's the only spot in the file.

Porting note 2 — tokenizer_config: bos/eos_token_id required

Google's July tokenizer_config.json drops bos_token_id and eos_token_id. FLM reads them and fails:

Failed to load model: [json.exception.type_error.302] type must be number, but is null

Fix: carry them over from the previous config (bos_token_id: 2, eos_token_id: [1, 106]).

Verification (with both shims applied)

• FLM v0.9.45, Windows 11, Ryzen AI 5 340 (XDNA2), driver 32.0.203.314
• Plain chat: normal.
• Tool calling: POST /v1/chat/completions with a tools array returns finish_reason: "tool_calls", a well-formed tool_calls entry with JSON arguments, and content: null — no tool-call text bleeding into content, which the April x-regex ordering was prone to.

Ask

A refreshed Gemma4-E4B-IT-NPU2 (and siblings) built against Google's July files, with the two shims above. Weights don't appear to have changed in the July wave, so this may be a metadata-only republish. Happy to test.