【Feature】: add GLM-47 tool parser and support thinking/non-thinking mode toggle`

[astrophel0]

PR Description

1. Background

Currently, vLLM does not support Function Calling for the GLM-47 (GLM4) series models. This PR introduces a dedicated tool parser for GLM-47 and adds a control mechanism for its hybrid "thinking" (reasoning) mode.

2. Changes

• New Tool Parser: Added the glm47 tool parser. Users can enable it by passing the command-line argument:

  --tool-call-parser glm47

• Thinking Mode Toggle: Added support for switching between thinking and non-thinking modes via chat_template_kwargs.
  • Control Parameter: "chat_template_kwargs": {"enable_thinking": true/false}.
  • Default Behavior: Thinking mode is enabled by default.
• Bug Fix (Multi-turn Conversation):
  • Issue: The previous logic relied on string matching (checking if </think> existed in the prompt) to determine the reasoning state, which was unreliable in multi-turn dialogues.
  • Fix: Replaced the string-matching logic with explicit control via the enable_thinking parameter, ensuring consistent behavior across complex conversation histories.

3. Test Plan

• Verified Function Calling performance with GLM-47 models using the --tool-call-parser glm47 flag.
• Tested multi-turn chat completions to ensure enable_thinking correctly toggles the model's reasoning output without state confusion.

---

Notes: This PR improves the robustness of GLM4 integration by moving away from prompt-parsing heuristics to explicit parameter control.

[liwei109]

@kurkol Please review this PR.

[kurkol]

Thanks for your work on adapting the GLM-4.7 model. However, during the adaptation, please avoid overwriting entire vLLM source files. Instead, use localized replacements or override specific class methods to maximize the vllm-kunlun plugin’s compatibility with different vLLM versions. You can refer to other adaptation PRs for concrete examples.

[kurkol]

In particular, you can follow the approach used in this PR, which replaces the relevant class methods: #75

[astrophel0]

In particular, you can follow the approach used in this PR, which replaces the relevant class methods: #75

To keep the PR focused, I have only modified two methods within this class to support toggling between thinking and non-thinking modes. This ensures minimal impact on the existing logic.

[kurkol]

In particular, you can follow the approach used in this PR, which replaces the relevant class methods: #75

To keep the PR focused, I have only modified two methods within this class to support toggling between thinking and non-thinking modes. This ensures minimal impact on the existing logic.

OK, then just fix the file naming — it should be glm, not gim.

[astrophel0]

In particular, you can follow the approach used in this PR, which replaces the relevant class methods: #75

To keep the PR focused, I have only modified two methods within this class to support toggling between thinking and non-thinking modes. This ensures minimal impact on the existing logic.

OK, then just fix the file naming — it should be glm, not gim.

done