| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | # Copyright 2025 Google Inc. HuggingFace Inc. team. All rights reserved. |
| 4 | |
| 5 | """Gemma4 thinking/reasoning output parsing utilities for offline inference. |
| 6 | |
| 7 | Standalone functions that parse decoded model text to extract structured |
| 8 | thinking content from Gemma4 models. These are pure-Python utilities with |
| 9 | zero heavy dependencies — they work on raw decoded strings from any |
| 10 | inference backend (vLLM, HuggingFace, TGI, etc.). |
| 11 | |
| 12 | For the OpenAI-compatible API reasoning parser (streaming + |
| 13 | non-streaming), see ``vllm.reasoning.gemma4_reasoning_parser``. |
| 14 | For tool call parsing, see ``vllm.tool_parsers.gemma4_utils``. |
| 15 | |
| 16 | Usage with vLLM offline inference:: |
| 17 | |
| 18 | from vllm import LLM, SamplingParams |
| 19 | from vllm.reasoning.gemma4_utils import parse_thinking_output |
| 20 | |
| 21 | llm = LLM(model="google/gemma-4-it") |
| 22 | outputs = llm.generate(prompt, SamplingParams(...)) |
| 23 | text = tokenizer.decode(outputs[0].outputs[0].token_ids, skip_special_tokens=False) |
| 24 | |
| 25 | # Extract thinking / answer (works with or without enable_thinking) |
| 26 | result = parse_thinking_output(text) |
| 27 | print(result["thinking"]) # chain-of-thought or None |
| 28 | print(result["answer"]) # final answer |
| 29 | |
| 30 | Ported from ``transformers.models.gemma4.utils_gemma4`` so that vLLM users |
| 31 | do not need a transformers dependency for output parsing. |
| 32 | """ |
| 33 | |
| 34 | # ---- Thinking Mode Utility ---- |
| 35 | |
| 36 | # Thinking delimiter tokens as they appear in decoded text. |
| 37 | # Gemma4 uses <|channel> (start) and <channel|> (end) as thinking delimiters. |
| 38 | _THINKING_START_TAG = "<|channel>" |
| 39 | _THINKING_END_TAG = "<channel|>" |
| 40 | |
| 41 | # Sentinel tokens that may appear in decoded output. |
| 42 | _TURN_END_TAG = "<turn|>" |
| 43 | |
| 44 | |
| 45 | def parse_thinking_output(text: str) -> dict[str, str | None]: |
| 46 | """Parse decoded Gemma4 model output. |
| 47 | |
| 48 | Use this on **all** Gemma4 output regardless of whether thinking mode |
| 49 | was enabled. It handles three cases: |
| 50 | |
| 51 | 1. **Thinking enabled, tags present** — splits on ``<|channel>``/ |
| 52 | ``<channel|>`` to separate chain-of-thought from the answer and |
| 53 | strips the ``thought\\n`` role label. |
| 54 | 2. **Thinking disabled, spurious label** — strips the bare |
| 55 | ``thought\\n`` prefix that some Gemma4 models emit even |
| 56 | without thinking mode. |
| 57 | 3. **Clean output** — returns the text unchanged. |
| 58 | |
| 59 | The answer text is always cleaned of trailing sentinel tokens |
| 60 | (``<turn|>``, ``<eos>``, etc.). |
| 61 | |
| 62 | Args: |
| 63 | text: Decoded model output text (from ``tokenizer.decode(...)``). |
| 64 | |
| 65 | Returns: |
| 66 | A dict with keys: |
| 67 | - ``"thinking"``: The chain-of-thought text, or ``None`` if no |
| 68 | thinking delimiters were found. |
| 69 | - ``"answer"``: The final answer text. |
| 70 | |
| 71 | Example:: |
| 72 | |
| 73 | >>> from vllm.reasoning.gemma4_utils import parse_thinking_output |
| 74 | >>> output_text = tokenizer.decode(outputs[0], skip_special_tokens=False) |
| 75 | >>> result = parse_thinking_output(output_text) |
| 76 | >>> print(result["thinking"]) # chain-of-thought reasoning or None |
| 77 | >>> print(result["answer"]) # final answer |
| 78 | |
| 79 | """ |
| 80 | if _THINKING_END_TAG in text: |
| 81 | parts = text.split(_THINKING_END_TAG, 1) |
| 82 | thinking_block = parts[0] |
| 83 | answer = _clean_answer(parts[1]) |
| 84 | |
| 85 | # Extract thinking content: strip the start tag if present |
| 86 | if _THINKING_START_TAG in thinking_block: |
| 87 | thinking = thinking_block.split(_THINKING_START_TAG, 1)[1] |
| 88 | else: |
| 89 | thinking = thinking_block |
| 90 | |
| 91 | # Strip the "thought\n" channel role label the model emits inside |
| 92 | # <|channel>thought\n...<channel|> (analogous to "user\n" in |
| 93 | # <|turn>user\n...<turn|>). |
| 94 | thinking = _strip_thought_label(thinking.strip()) |
| 95 | thinking = thinking.strip() |
| 96 | |
| 97 | return {"thinking": thinking, "answer": answer} |
| 98 | |
| 99 | # No thinking delimiters found. |
| 100 | # Strip spurious "thought\n" role label that some Gemma4 models sometimes |
| 101 | # emit even without thinking mode enabled, then clean trailing tokens. |
| 102 | answer = _strip_thought_label(text) |
| 103 | answer = _clean_answer(answer) |
| 104 | return {"thinking": None, "answer": answer} |
| 105 | |
| 106 | |
| 107 | def _strip_thought_label(text: str) -> str: |
| 108 | """Strip the spurious ``thought\\n`` label from the start of text. |
| 109 | |
| 110 | Only strips when ``thought`` appears as the very first word followed by |
| 111 | a newline — preserving the word ``thought`` in any other context. |
| 112 | """ |
| 113 | if text.startswith("thought\n"): |
| 114 | return text[len("thought\n") :] |
| 115 | return text |
| 116 | |
| 117 | |
| 118 | def _clean_answer(text: str) -> str: |
| 119 | """Clean trailing sentinel tokens from the answer text. |
| 120 | |
| 121 | Strips ``<turn|>``, ``<eos>``, and surrounding whitespace that the |
| 122 | model appends at the end of its response. |
| 123 | """ |
| 124 | text = text.strip() |
| 125 | # Strip trailing <turn|> (Gemma4 turn-end marker) |
| 126 | if text.endswith(_TURN_END_TAG): |
| 127 | text = text[: -len(_TURN_END_TAG)].rstrip() |
| 128 | # Strip trailing <eos> if present |
| 129 | if text.endswith("<eos>"): |
| 130 | text = text[:-5].rstrip() |
| 131 | return text |