| 123456789101112131415161718192021222324252627282930313233343536373839 |
- #!/usr/bin/env python3
- """Always split reasoning from content, whatever the thinking kwargs say.
- GLM-5.3-Flash was trained to think on every turn; its official template has no
- non-thinking mode. Our template therefore maps `thinking: false` /
- `enable_thinking: false` to `Reasoning Effort: Low` and still opens `<think>`
- (the empty `<think></think>` we used before is out of distribution and breaks
- long structured output). The model now always emits a short reasoning block.
- Stock glm47_moe drops its <think> terminals and starts in CONTENT when either
- kwarg is false, so that reasoning would land in `content`. This makes the
- parser track <think> on every request: thinking off returns a short `reasoning`
- and a clean `content`.
- Same contract as the other patchers: one anchor, exactly once, or the build
- fails.
- """
- from pathlib import Path
- PATH = Path("/usr/local/lib/python3.12/dist-packages/vllm/parser/glm47_moe.py")
- OLD = """ self.thinking_enabled = (
- True
- if thinking is None and enable_thinking is None
- else bool(thinking) or bool(enable_thinking)
- )
- """
- NEW = """ # GLM53-REASONING-ALWAYS: the template always opens <think> (thinking
- # off means low effort), so the parser must always track it.
- del thinking, enable_thinking
- self.thinking_enabled = True
- """
- text = PATH.read_text()
- if text.count(NEW) == 1 and text.count(OLD) == 0:
- print("[reasoning-always] already patched")
- else:
- assert text.count(OLD) == 1, f"[reasoning-always] anchor matched {text.count(OLD)} times; stock tree changed"
- PATH.write_text(text.replace(OLD, NEW))
- print("[reasoning-always] glm47_moe always parses <think>")
|
