[Bugfix][Gemma4] Fix offline parser truncation, adjust_request token leak, and chat template sync (#45553)
Signed-off-by: Luciano Martins <lucianommartins@users.noreply.github.com> Co-authored-by: Luciano Martins <lucianommartins@users.noreply.github.com>
This commit is contained in:
co-authored by
Luciano Martins
parent
b8bd773fe4
commit
6607a80dab
@@ -116,7 +116,9 @@
|
||||
}
|
||||
{%- endmacro -%}
|
||||
{%- macro format_argument(argument, escape_keys=True) -%}
|
||||
{%- if argument is string -%}
|
||||
{%- if argument is none -%}
|
||||
{{- 'null' -}}
|
||||
{%- elif argument is string -%}
|
||||
{{- '<|"|>' + argument + '<|"|>' -}}
|
||||
{%- elif argument is boolean -%}
|
||||
{{- 'true' if argument else 'false' -}}
|
||||
@@ -172,18 +174,21 @@
|
||||
{{- '<tool_response|>' -}}
|
||||
{%- endmacro -%}
|
||||
|
||||
{%- set ns = namespace(prev_message_type=None) -%}
|
||||
{#- ===== SETUP ===== -#}
|
||||
{%- set ns = namespace(prev_message_type=None, prev_non_tool_role=None) -%}
|
||||
{%- set loop_messages = messages -%}
|
||||
{%- set enable_thinking = enable_thinking | default(false) -%}
|
||||
{%- set preserve_thinking = preserve_thinking | default(false) -%}
|
||||
{{- bos_token -}}
|
||||
{#- Handle System/Tool Definitions Block -#}
|
||||
{%- if (enable_thinking is defined and enable_thinking) or tools or messages[0]['role'] in ['system', 'developer'] -%}
|
||||
{%- if enable_thinking or tools or (messages and messages[0]['role'] in ['system', 'developer']) -%}
|
||||
{{- '<|turn>system\n' -}}
|
||||
{#- Inject Thinking token at the very top of the FIRST system turn -#}
|
||||
{%- if enable_thinking is defined and enable_thinking -%}
|
||||
{%- if enable_thinking -%}
|
||||
{{- '<|think|>\n' -}}
|
||||
{%- set ns.prev_message_type = 'think' -%}
|
||||
{%- endif -%}
|
||||
{%- if messages[0]['role'] in ['system', 'developer'] -%}
|
||||
{%- if messages and messages[0]['role'] in ['system', 'developer'] -%}
|
||||
{%- if messages[0]['content'] is string -%}
|
||||
{{- messages[0]['content'] | trim -}}
|
||||
{%- elif messages[0]['content'] is sequence -%}
|
||||
@@ -217,31 +222,24 @@
|
||||
{%- if message['role'] != 'tool' -%}
|
||||
{%- set ns.prev_message_type = None -%}
|
||||
{%- set role = 'model' if message['role'] == 'assistant' else message['role'] -%}
|
||||
{#- Detect continuation: suppress duplicate <|turn>model when previous non-tool message was also assistant -#}
|
||||
{%- set prev_nt = namespace(role=None, found=false) -%}
|
||||
{%- if loop.index0 > 0 -%}
|
||||
{%- for j in range(loop.index0 - 1, -1, -1) -%}
|
||||
{%- if not prev_nt.found -%}
|
||||
{%- if loop_messages[j]['role'] != 'tool' -%}
|
||||
{%- set prev_nt.role = loop_messages[j]['role'] -%}
|
||||
{%- set prev_nt.found = true -%}
|
||||
{%- endif -%}
|
||||
{%- endif -%}
|
||||
{%- endfor -%}
|
||||
{%- endif -%}
|
||||
{%- set continue_same_model_turn = (role == 'model' and prev_nt.role == 'assistant') -%}
|
||||
{#- Detect continuation using tracked state — O(1) instead of O(n) backward scan -#}
|
||||
{%- set continue_same_model_turn = (role == 'model' and ns.prev_non_tool_role == 'assistant') -%}
|
||||
{%- if not continue_same_model_turn -%}
|
||||
{{- '<|turn>' + role + '\n' }}
|
||||
{%- if role == 'model' and not enable_thinking and not (message.get('reasoning') or message.get('reasoning_content')) -%}
|
||||
{{- '<|channel>thought\n<channel|>' -}}
|
||||
{%- endif -%}
|
||||
{%- endif -%}
|
||||
|
||||
{#- Render reasoning/reasoning_content as thinking channel -#}
|
||||
{#- Render reasoning/reasoning_content as thinking channel (tool-call turns only) -#}
|
||||
{%- set thinking_text = message.get('reasoning') or message.get('reasoning_content') -%}
|
||||
{%- if thinking_text and loop.index0 > ns_turn.last_user_idx and message.get('tool_calls') -%}
|
||||
{%- set thinking_gate = (loop.index0 > ns_turn.last_user_idx) or preserve_thinking -%}
|
||||
{%- if thinking_text and thinking_gate and message.get('tool_calls') -%}
|
||||
{{- '<|channel>thought\n' + thinking_text + '\n<channel|>' -}}
|
||||
{%- endif -%}
|
||||
|
||||
{%- if message['tool_calls'] -%}
|
||||
{%- for tool_call in message['tool_calls'] -%}
|
||||
{%- if message.get('tool_calls') -%}
|
||||
{%- for tool_call in message.get('tool_calls') -%}
|
||||
{%- set function = tool_call['function'] -%}
|
||||
{{- '<|tool_call>call:' + function['name'] + '{' -}}
|
||||
{%- if function['arguments'] is mapping -%}
|
||||
@@ -251,8 +249,13 @@
|
||||
{%- set ns_args.found_first = true -%}
|
||||
{{- key -}}:{{- format_argument(value, escape_keys=False) -}}
|
||||
{%- endfor -%}
|
||||
{%- elif function['arguments'] is string -%}
|
||||
{{- function['arguments'] -}}
|
||||
{%- elif function['arguments'] is none -%}
|
||||
{%- else -%}
|
||||
{{- raise_exception(
|
||||
"chat_template: tool_calls[].function.arguments must be a "
|
||||
"JSON object (mapping), not a string. Deserialize arguments "
|
||||
"before passing to the template."
|
||||
) -}}
|
||||
{%- endif -%}
|
||||
{{- '}<tool_call|>' -}}
|
||||
{%- endfor -%}
|
||||
@@ -262,7 +265,7 @@
|
||||
{%- set ns_tr_out = namespace(flag=false) -%}
|
||||
{%- if message.get('tool_responses') -%}
|
||||
{#- Legacy: tool_responses embedded on the assistant message (Google/Gemma native) -#}
|
||||
{%- for tool_response in message['tool_responses'] -%}
|
||||
{%- for tool_response in message.get('tool_responses') -%}
|
||||
{{- format_tool_response_block(tool_response['name'] | default('unknown', true), tool_response['response']) -}}
|
||||
{%- set ns_tr_out.flag = true -%}
|
||||
{%- set ns.prev_message_type = 'tool_response' -%}
|
||||
@@ -277,8 +280,8 @@
|
||||
{%- else -%}
|
||||
{%- set follow = loop_messages[k] -%}
|
||||
{#- Resolve tool_call_id to function name -#}
|
||||
{%- set ns_tname = namespace(name=follow.get('name') | default('unknown', true)) -%}
|
||||
{%- for tc in message['tool_calls'] -%}
|
||||
{%- set ns_tname = namespace(name=follow.get('name') or 'unknown') -%}
|
||||
{%- for tc in message.get('tool_calls') -%}
|
||||
{%- if tc.get('id') == follow.get('tool_call_id') -%}
|
||||
{%- set ns_tname.name = tc['function']['name'] -%}
|
||||
{%- endif -%}
|
||||
@@ -296,9 +299,9 @@
|
||||
{%- endfor -%}
|
||||
{{- format_tool_response_block(ns_tname.name, ns_txt.s) -}}
|
||||
{%- for part in tool_body -%}
|
||||
{%- if part.get('type') == 'image' -%}
|
||||
{%- if part.get('type') in ['image', 'image_url'] -%}
|
||||
{{- '<|image|>' -}}
|
||||
{%- elif part.get('type') == 'audio' -%}
|
||||
{%- elif part.get('type') in ['audio', 'input_audio'] -%}
|
||||
{{- '<|audio|>' -}}
|
||||
{%- elif part.get('type') == 'video' -%}
|
||||
{{- '<|video|>' -}}
|
||||
@@ -314,29 +317,26 @@
|
||||
{%- endif -%}
|
||||
|
||||
{%- set captured_content -%}
|
||||
{%- if message['content'] is string -%}
|
||||
{%- if message.get('content') is string -%}
|
||||
{%- if role == 'model' -%}
|
||||
{{- strip_thinking(message['content']) -}}
|
||||
{%- else -%}
|
||||
{{- message['content'] | trim -}}
|
||||
{%- endif -%}
|
||||
{%- elif message['content'] is sequence -%}
|
||||
{%- elif message.get('content') is sequence -%}
|
||||
{%- for item in message['content'] -%}
|
||||
{%- if item['type'] == 'text' -%}
|
||||
{%- if item.get('type') == 'text' -%}
|
||||
{%- if role == 'model' -%}
|
||||
{{- strip_thinking(item['text']) -}}
|
||||
{%- else -%}
|
||||
{{- item['text'] | trim -}}
|
||||
{%- endif -%}
|
||||
{%- elif item['type'] == 'image' -%}
|
||||
{%- elif item.get('type') in ['image', 'image_url'] -%}
|
||||
{{- '<|image|>' -}}
|
||||
{%- set ns.prev_message_type = 'image' -%}
|
||||
{%- elif item['type'] == 'audio' -%}
|
||||
{%- elif item.get('type') in ['audio', 'input_audio'] -%}
|
||||
{{- '<|audio|>' -}}
|
||||
{%- set ns.prev_message_type = 'audio' -%}
|
||||
{%- elif item['type'] == 'video' -%}
|
||||
{%- elif item.get('type') == 'video' -%}
|
||||
{{- '<|video|>' -}}
|
||||
{%- set ns.prev_message_type = 'video' -%}
|
||||
{%- endif -%}
|
||||
{%- endfor -%}
|
||||
{%- endif -%}
|
||||
@@ -345,19 +345,43 @@
|
||||
{{- captured_content -}}
|
||||
{%- set has_content = captured_content | trim | length > 0 -%}
|
||||
|
||||
{#- Forward-scan: find next non-tool message role for continuation detection -#}
|
||||
{%- set next_nt = namespace(role=None, found=false) -%}
|
||||
{%- for j in range(loop.index0 + 1, loop_messages | length) -%}
|
||||
{%- if not next_nt.found -%}
|
||||
{%- if loop_messages[j]['role'] != 'tool' -%}
|
||||
{%- set next_nt.role = loop_messages[j]['role'] -%}
|
||||
{%- set next_nt.found = true -%}
|
||||
{%- endif -%}
|
||||
{%- endif -%}
|
||||
{%- endfor -%}
|
||||
|
||||
{%- set continues_into_next = (
|
||||
role == 'model'
|
||||
and next_nt.role == 'assistant'
|
||||
and (not message.get('tool_calls') or ns_tr_out.flag)
|
||||
) -%}
|
||||
|
||||
{%- if ns.prev_message_type == 'tool_call' and not ns_tr_out.flag -%}
|
||||
{{- '<|tool_response>' -}}
|
||||
{%- elif continues_into_next -%}
|
||||
{{- '\n' -}}
|
||||
{%- elif not (ns_tr_out.flag and not has_content) -%}
|
||||
{{- '<turn|>\n' -}}
|
||||
{%- endif -%}
|
||||
|
||||
{#- Track previous non-tool role for next iteration (avoids O(n) backward scan) -#}
|
||||
{%- set ns.prev_non_tool_role = message['role'] -%}
|
||||
{%- endif -%}
|
||||
{%- endfor -%}
|
||||
|
||||
{%- if add_generation_prompt -%}
|
||||
{%- if ns.prev_message_type != 'tool_response' and ns.prev_message_type != 'tool_call' -%}
|
||||
{{- '<|turn>model\n' -}}
|
||||
{%- if not enable_thinking | default(false) -%}
|
||||
{%- if not enable_thinking -%}
|
||||
{{- '<|channel>thought\n<channel|>' -}}
|
||||
{%- endif -%}
|
||||
{%- elif ns.prev_message_type == 'tool_response' and enable_thinking -%}
|
||||
{{- '<|channel>thought\n' -}}
|
||||
{%- endif -%}
|
||||
{%- endif -%}
|
||||
{%- endif -%}
|
||||
|
||||
@@ -54,7 +54,7 @@ NO_REASONING = {
|
||||
"output": "This is content",
|
||||
"reasoning": None,
|
||||
"content": "This is content",
|
||||
"is_reasoning_end": False,
|
||||
"is_reasoning_end": True,
|
||||
}
|
||||
REASONING_WITH_CHANNEL = {
|
||||
"output": "<|channel>This is a reasoning section<channel|>This is the rest",
|
||||
|
||||
@@ -358,7 +358,7 @@ class TestGemma4ChatTemplate:
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "download_image",
|
||||
"arguments": '{"url": "https://example.com/x.png"}',
|
||||
"arguments": {"url": "https://example.com/x.png"},
|
||||
},
|
||||
},
|
||||
],
|
||||
@@ -392,7 +392,7 @@ class TestGemma4ChatTemplate:
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "process",
|
||||
"arguments": "{}",
|
||||
"arguments": {},
|
||||
},
|
||||
},
|
||||
],
|
||||
|
||||
+20
-3
@@ -423,6 +423,8 @@ class Gemma4Parser(ParserEngine):
|
||||
tools: list[Tool] | None = None,
|
||||
**kwargs,
|
||||
) -> None:
|
||||
chat_kwargs = kwargs.get("chat_template_kwargs", {}) or {}
|
||||
self._thinking_enabled = chat_kwargs.get("enable_thinking", True)
|
||||
super().__init__(
|
||||
tokenizer,
|
||||
tools,
|
||||
@@ -437,6 +439,21 @@ class Gemma4Parser(ParserEngine):
|
||||
self._prefix_stripped: bool = False
|
||||
self._is_first_feed: bool = True
|
||||
|
||||
def adjust_request(
|
||||
self,
|
||||
request: ChatCompletionRequest | ResponsesRequest,
|
||||
) -> ChatCompletionRequest | ResponsesRequest:
|
||||
"""Skip ``skip_special_tokens=False`` when thinking is disabled.
|
||||
|
||||
When there are no reasoning channel tokens to preserve,
|
||||
keeping the default prevents tool-call delimiter tokens
|
||||
from leaking into content (e.g. with ``tool_choice="none"``).
|
||||
"""
|
||||
chat_template_kwargs = getattr(request, "chat_template_kwargs", None) or {}
|
||||
if not chat_template_kwargs.get("enable_thinking", True):
|
||||
return request
|
||||
return super().adjust_request(request)
|
||||
|
||||
def _reset(self, initial_state=None) -> None:
|
||||
super()._reset(initial_state=initial_state)
|
||||
self._reasoning_text = ""
|
||||
@@ -494,12 +511,12 @@ class Gemma4Parser(ParserEngine):
|
||||
if tool_call_id is not None and tid == tool_call_id:
|
||||
return True
|
||||
if new_turn_id is not None and tid == new_turn_id:
|
||||
return False
|
||||
return not self._thinking_enabled
|
||||
if tool_response_id is not None and tid == tool_response_id:
|
||||
return False
|
||||
return not self._thinking_enabled
|
||||
if end_id is not None and tid == end_id:
|
||||
return True
|
||||
return self._reasoning_ended
|
||||
return True
|
||||
|
||||
def _events_to_delta(
|
||||
self,
|
||||
|
||||
@@ -35,8 +35,6 @@ Ported from ``transformers.models.gemma4.utils_gemma4`` so that vLLM users
|
||||
do not need a transformers dependency for output parsing.
|
||||
"""
|
||||
|
||||
import json
|
||||
|
||||
import regex as re
|
||||
|
||||
# Tool call delimiter tokens as they appear in decoded text.
|
||||
@@ -52,42 +50,23 @@ _ESCAPE_TOKEN = '<|"|>'
|
||||
def _parse_tool_arguments(args_str: str) -> dict[str, str]:
|
||||
"""Parse tool call arguments from the Gemma4 compact format.
|
||||
|
||||
Handles the ``key:<|"|>value<|"|>`` format used by Gemma4, with fallback
|
||||
to heuristic key-value extraction. Also tolerates the slightly different
|
||||
``key: "value"`` format (space + plain quotes) that some chat templates
|
||||
produce.
|
||||
Delegates to the native ``<|"|>``-aware parser from
|
||||
``vllm.parser.gemma4``, which handles internal quotes, nested
|
||||
objects, arrays, and all Gemma4 value types correctly.
|
||||
|
||||
Args:
|
||||
args_str: Raw argument string from inside ``call:name{...}``.
|
||||
|
||||
Returns:
|
||||
Dictionary of argument name → value.
|
||||
Dictionary of argument name → string value.
|
||||
"""
|
||||
if not args_str or not args_str.strip():
|
||||
return {}
|
||||
|
||||
# Replace Gemma4 escape tokens with standard quotes.
|
||||
cleaned = args_str.replace(_ESCAPE_TOKEN, '"')
|
||||
from vllm.parser.gemma4 import _parse_gemma4_args
|
||||
|
||||
# Try JSON parsing first (handles nested values, arrays, etc.).
|
||||
try:
|
||||
parsed = json.loads("{" + cleaned + "}")
|
||||
# Ensure all values are strings for consistency.
|
||||
return {k: str(v) if not isinstance(v, str) else v for k, v in parsed.items()}
|
||||
except (json.JSONDecodeError, ValueError):
|
||||
pass
|
||||
|
||||
# Fallback: extract key:"value" pairs (allow optional space after colon).
|
||||
arguments = {}
|
||||
for key, value in re.findall(r'(\w+):\s*"([^"]*)"', cleaned):
|
||||
arguments[key] = value
|
||||
|
||||
if not arguments:
|
||||
# Last resort: extract key:value pairs (unquoted).
|
||||
for key, value in re.findall(r"(\w+):\s*([^,}]+)", args_str):
|
||||
arguments[key] = value.strip().strip('"').replace(_ESCAPE_TOKEN, "")
|
||||
|
||||
return arguments
|
||||
parsed = _parse_gemma4_args(args_str)
|
||||
return {k: str(v) if not isinstance(v, str) else v for k, v in parsed.items()}
|
||||
|
||||
|
||||
def parse_tool_calls(text: str, *, strict: bool = False) -> list[dict]:
|
||||
|
||||
Reference in New Issue
Block a user