feat: Gemma 4 tool call + reasoning parser

[Thump604]

Summary

• Gemma 4 tool call parser: parses <|tool_call>call:name{params}<tool_call|> with <|"|> string escaping
• Gemma 4 reasoning parser: extracts <|channel>thought...<channel|> blocks into reasoning_content
• Registered as --tool-call-parser gemma4 and --reasoning-parser gemma4
• CLI choices updated

Addresses the server-side equivalent of ml-explore/mlx-lm#1096 (Gemma 4 native tool calls not parsed).

Details

Tool parser (tool_parsers/gemma4_tool_parser.py):

• Regex extracts function name and parameter block from <|tool_call>call:name{...}<tool_call|>
• Handles <|"|> escaped string values, numeric/boolean coercion
• Arguments serialized to JSON string for OpenAI API compatibility
• Streaming support with start-marker detection

Reasoning parser (reasoning/gemma4_parser.py):

• Splits <|channel>thought\n...reasoning...<channel|> from final answer
• Non-streaming: returns (reasoning, content) tuple
• Streaming: tracks channel boundaries, emits reasoning deltas then content deltas

Test plan

• Non-streaming: tool_calls field populated with correct function name + JSON arguments
• Non-streaming: reasoning_content separated from content, both present
• Streaming: SSE chunks flow correctly
• Concurrent: 3 simultaneous requests with BatchedEngine
• Thinking + tools: model produces empty think channel then tool call (correct behavior)
• Parser standalone unit tests pass

[Thump604]

Quick merge signal on this one: this is the Gemma 4 parser pair we are already relying on locally on the current Gemma lane (reasoning_parser=gemma4, tool_call_parser=gemma4). No known blockers on our side, branch is mergeable, and CI is green.

If you're doing a fast sweep, this one looks ready.

[janhilgard]

+1 for merge — we've been running this in production on Gemma 4 26B-A4B (mixed 6/8-bit) for several days now with --tool-call-parser gemma4 --reasoning-parser gemma4 and it's been solid. Tool calls parse correctly, reasoning content separates cleanly, and streaming works as expected under concurrent load.

Clean implementation, no issues encountered. Would be great to get this into main. 🙏

[waybarrios]

Thanks for this one. The reasoning parser landed via #268 and the tool call parser is covered by #269 which we're about to merge. Your review and testing on the Gemma 4 work across multiple PRs made a real difference in getting this shipped. Closing in favor of #269.