[Frontend] Replace legacy Gemma4 parsers with engine-based implementation (#45588)

Signed-off-by: Ben Browning <bbrownin@redhat.com>
Co-authored-by: Flora Feng <4florafeng@gmail.com>
This commit is contained in:
Ben Browning
2026-06-15 21:34:07 +00:00
committed by GitHub
co-authored by Flora Feng
parent 25ee659db0
commit 76a373eff4
20 changed files with 2809 additions and 1333 deletions
+113 -2
View File
@@ -29,6 +29,7 @@ from vllm.entrypoints.openai.chat_completion.protocol import (
ChatCompletionToolsParam,
)
from vllm.parser.engine.registered_adapters import (
Gemma4Parser,
Qwen3Parser,
)
@@ -48,6 +49,7 @@ class Scenario:
reasoning: str | None = None
content: str | None = None
tool_calls: list[ToolCallSpec] | None = None
after_tool_response: bool = False
# ── Scenarios ────────────────────────────────────────────────────────
@@ -132,6 +134,12 @@ SCENARIOS: list[Scenario] = [
reasoning="",
content="The epoch timestamp is 1779111346.",
),
Scenario(
id="tool-after-tool-response",
description="Tool call immediately after tool response (agentic flow)",
tool_calls=[_READ_TOOL],
after_tool_response=True,
),
]
@@ -250,7 +258,13 @@ def _validate_sample(sample: Sample, parser_cls: type, **kwargs) -> None:
"""Replay sample through the real parser and assert correctness."""
tokenizer = MockTokenizer(vocab=dict(sample.vocab), tokens=sample.tokens)
parser = parser_cls(tokenizer, sample.tools, **kwargs)
deltas = replay_streaming(parser, sample.tokens, chunk_size=1, tools=sample.tools)
deltas = replay_streaming(
parser,
sample.tokens,
chunk_size=1,
tools=sample.tools,
prompt_token_ids=sample.prompt_token_ids,
)
output = collect_output(deltas)
assert_parse_output(output, sample)
@@ -273,6 +287,7 @@ def _make_sample(
expected_tool_calls: list[dict] | None,
tools: list[dict] | None,
chat_template_kwargs: dict | None = None,
prompt_token_ids: list[int] | None = None,
) -> Sample:
tokens = _tokenize(segments, vocab)
return Sample(
@@ -286,10 +301,11 @@ def _make_sample(
expected_tool_calls=expected_tool_calls,
tools=_validate_tools(tools),
chat_template_kwargs=chat_template_kwargs,
prompt_token_ids=prompt_token_ids,
)
# ── Qwen3 / NemotronV3 (XML tool format, starts in REASONING) ───────
# ── Qwen3 (XML tool format, starts in REASONING) ────────────────────
_QWEN3_VOCAB: dict[str, int] = {
"<think>": 50,
@@ -376,10 +392,105 @@ def _build_qwen3(
return sample
# ── Gemma4 (channel reasoning, custom arg format) ────────────────────
_GEMMA4_VOCAB: dict[str, int] = {
"<|channel>": 50,
"<channel|>": 51,
"<|tool_call>": 48,
"<tool_call|>": 49,
'<|"|>': 52,
"<|turn>": 53,
"<|tool_response>": 54,
}
_GEMMA4_THOUGHT_PREFIX = "thought\n"
_GEMMA4_QUOTE = '<|"|>'
def _gemma4_value_segments(value: Any) -> list[tuple[str, bool]]:
"""Render a value in Gemma4 arg format as segments."""
if isinstance(value, str):
return [(_GEMMA4_QUOTE, True), (value, False), (_GEMMA4_QUOTE, True)]
if isinstance(value, bool):
return [("true" if value else "false", False)]
if isinstance(value, (int, float)):
return [(str(value), False)]
if isinstance(value, dict):
segs: list[tuple[str, bool]] = [("{", False)]
for i, (k, v) in enumerate(value.items()):
if i > 0:
segs.append((",", False))
segs.append((f"{k}:", False))
segs.extend(_gemma4_value_segments(v))
segs.append(("}", False))
return segs
if isinstance(value, list):
segs = [("[", False)]
for i, item in enumerate(value):
if i > 0:
segs.append((",", False))
segs.extend(_gemma4_value_segments(item))
segs.append(("]", False))
return segs
return [(json.dumps(value, ensure_ascii=False), False)]
def _gemma4_tool_segments(tc: ToolCallSpec) -> list[tuple[str, bool]]:
segs: list[tuple[str, bool]] = [
("<|tool_call>", True),
(f"call:{tc.name}", False),
("{", False),
]
for i, (key, value) in enumerate(tc.arguments.items()):
if i > 0:
segs.append((",", False))
segs.append((f"{key}:", False))
segs.extend(_gemma4_value_segments(value))
segs.append(("}", False))
segs.append(("<tool_call|>", True))
return segs
def _gemma4_segments(scenario: Scenario) -> list[tuple[str, bool]]:
segs: list[tuple[str, bool]] = []
if scenario.reasoning is not None:
segs.append(("<|channel>", True))
segs.append((_GEMMA4_THOUGHT_PREFIX, False))
segs.append((scenario.reasoning, False))
segs.append(("<channel|>", True))
if scenario.content is not None:
segs.append((scenario.content, False))
if scenario.tool_calls:
for tc in scenario.tool_calls:
segs.extend(_gemma4_tool_segments(tc))
return segs
def _build_gemma4(scenario: Scenario, validate: bool = True) -> Sample:
prompt_token_ids = None
if scenario.after_tool_response:
prompt_token_ids = [_GEMMA4_VOCAB["<|tool_response>"]]
sample = _make_sample(
sample_id=f"gemma4-{scenario.id}",
description=scenario.description,
vocab=_GEMMA4_VOCAB,
segments=_gemma4_segments(scenario),
expected_reasoning=scenario.reasoning,
expected_content=_qwen3_expected_content(scenario),
expected_tool_calls=_expected_tc(scenario),
tools=_expected_tools(scenario),
prompt_token_ids=prompt_token_ids,
)
if validate:
_validate_sample(sample, Gemma4Parser)
return sample
# ── Registry and public API ──────────────────────────────────────────
_BUILDERS: dict[str, Any] = {
"qwen3": _build_qwen3,
"gemma4": _build_gemma4,
}