[Fix] minimax-m3: Type nested tool-call arguments and count self-opened reasoning by ispobock · Pull Request #46133 · sgl-project/sglang

ispobock · 2026-10-23T17:38:51Z

Summary

Fixes in the MiniMax-M3 tool-call parser and chat serving: type tool-call arguments nested in schema-less containers, and count reasoning tokens when the model opens the reasoning block itself.

Changes to existing code

These apply to every model:

Function.strict defaults to None and is left out of the rendered tool schema when unset; it was False and always injected.
Tool schemas keep the key order the client sent.
An image detail of "default" is read as "auto".
The root role is accepted alongside system and developer.
A thinking request field is accepted; only chat templates that define a thinking mode read it.
Usage reports reasoning tokens in completion_tokens_details.reasoning_tokens, set only when there are some.
ModelConfig.self_opened_think_start_ids defaults to None; the scheduler fills it only when a reasoning parser is configured.

MiniMax-M3 only: the request's own thinking, reasoning_effort and chat_template_kwargs settle thinking_mode and the effort before --default-chat-template-kwargs fill what the request left unset, so a server default never re-enables thinking the request turned off.

Verification

test_minimax_m3_detector, test_reasoning_parser, test_serving_chat, test_protocol, test_batch_result_processor_*, test_model_config (reasoning-token fields exist without a reasoning parser)

Cut byte-for-byte from #45925 (MiniMax-M3.1-Flash support) at 10e655a.

CI States

Latest PR Test (Base): 🚫 Run #40090965301
Latest PR Test (Extra): ❌ Run #40090965282
Latest PR Test (AMD ROCm 10): 🚫 Run #40090965289

Co-authored-by: Yuhao Yang <47235274+yhyang201@users.noreply.github.com> Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com> Co-authored-by: Yan Huang <yan.huang@radixark.ai>

ispobock · 2026-10-23T17:46:11Z

/rerun-test --changed

github-actions · 2026-10-23T17:46:38Z

🚀 ubuntu-latest (7 tests): ✅ View workflow run

cd test/ && python3 registered/unit/entrypoints/openai/test_protocol.py
cd test/ && python3 registered/unit/entrypoints/openai/test_serving_chat.py
cd test/ && python3 registered/unit/function_call/test_minimax_m3_detector.py
cd test/ && python3 registered/unit/managers/test_batch_result_processor_hidden_states.py
cd test/ && python3 registered/unit/managers/test_batch_result_processor_mamba_boundary.py
cd test/ && python3 registered/unit/managers/test_batch_result_processor_spec_grammar.py
cd test/ && python3 registered/unit/parser/test_reasoning_parser.py

ispobock · 2026-10-24T11:36:14Z

/rerun-test --changed

github-actions · 2026-10-24T11:36:36Z

🚀 ubuntu-latest (8 tests): ✅ View workflow run

cd test/ && python3 registered/unit/configs/test_model_config.py
cd test/ && python3 registered/unit/entrypoints/openai/test_protocol.py
cd test/ && python3 registered/unit/entrypoints/openai/test_serving_chat.py
cd test/ && python3 registered/unit/function_call/test_minimax_m3_detector.py
cd test/ && python3 registered/unit/managers/test_batch_result_processor_hidden_states.py
cd test/ && python3 registered/unit/managers/test_batch_result_processor_mamba_boundary.py
cd test/ && python3 registered/unit/managers/test_batch_result_processor_spec_grammar.py
cd test/ && python3 registered/unit/parser/test_reasoning_parser.py

ispobock · 2026-10-24T11:42:17Z

/rerun-test --changed

github-actions · 2026-10-24T11:42:39Z

🚀 ubuntu-latest (8 tests): ✅ View workflow run

cd test/ && python3 registered/unit/configs/test_model_config.py
cd test/ && python3 registered/unit/entrypoints/openai/test_protocol.py
cd test/ && python3 registered/unit/entrypoints/openai/test_serving_chat.py
cd test/ && python3 registered/unit/function_call/test_minimax_m3_detector.py
cd test/ && python3 registered/unit/managers/test_batch_result_processor_hidden_states.py
cd test/ && python3 registered/unit/managers/test_batch_result_processor_mamba_boundary.py
cd test/ && python3 registered/unit/managers/test_batch_result_processor_spec_grammar.py
cd test/ && python3 registered/unit/parser/test_reasoning_parser.py

…alizer now does Since #46133, Python's `Function` keeps the client's key order and drops an unset `strict` or `defer_loading`, so a Jinja template no longer sees `"strict": false` added to every tool.

…lizer now does Since #46133, Python's `Function` keeps the client's key order and leaves out an unset `strict` or `defer_loading` (a null `strict` included), so GLM-5.3 and Qwen3.8 templates no longer see `"strict": false` added to every tool. One `function_dump` now serves their renderers and the template fallback; Kimi K3's dump accepts a null `strict` too. Regenerated the GLM-5.3 and Qwen3.8 parity fixtures from current Python; their prompts change only in tool cases (other changes are error wording). The fixture generator's stub server gains `is_minimax_m3`.

…okens_details Since #46133, Python's UsageProcessor also reports the reasoning count where OpenAI puts it, `usage.completion_tokens_details.reasoning_tokens` (null without reasoning), for chat and completions, unary and streamed. Regenerated the GLM-5.3 and Qwen3.8 OpenAI fixtures from current Python, keeping their recorded engine outputs; their tool cases' /generate ids also take the new tool dump. The generator's stub template owner gains the model config `load_chat_template` now reads.