Xiaoze Fan
97 lines
4.1 KiB
Python
97 lines
4.1 KiB
Python
"""`--tool-call-parser auto` / `--reasoning-parser auto`: architecture -> parser name.
|
|
|
|
Three tables are maintained independently -- the model registry, the substring cascade in
|
|
``server/args.py``, and the two parser factories -- and nothing links them. Add a model family and
|
|
forget the cascade, and the model serves perfectly: it just stops emitting tool calls and stops
|
|
separating out its thinking, because it silently fell through to the generic default. Every other
|
|
parser test still passes, since they all start from a parser name that is already correct.
|
|
|
|
So this drives the *live* registry rather than a list: a newly registered architecture is covered
|
|
the moment it is added, and has to be dispositioned here to stay green.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
from unittest.mock import patch
|
|
|
|
import pytest
|
|
|
|
from freetoken.models.register import _MODEL_REGISTRY
|
|
from freetoken.server.args import parse_args
|
|
from freetoken.server.function_call_parser import FunctionCallParser
|
|
from freetoken.server.reasoning_parser import ReasoningParser
|
|
|
|
ARCHITECTURES = sorted(_MODEL_REGISTRY)
|
|
|
|
# The cascade also reads the model path, so the path here is deliberately anonymous: it must
|
|
# resolve off the checkpoint's own architecture, not off a directory somebody happened to name.
|
|
ANON_PATH = "/models/anon"
|
|
|
|
# Families with no thinking format of their own. Everything else must resolve to a real reasoning
|
|
# parser, and a new architecture landing here is the bug this file exists to catch.
|
|
NO_REASONING_FORMAT = {
|
|
"LlamaForCausalLM",
|
|
"MistralForCausalLM",
|
|
"Mistral3ForConditionalGeneration",
|
|
"Qwen2ForCausalLM",
|
|
}
|
|
|
|
# `llama3` is the end of the cascade -- the answer when nothing matched.
|
|
GENERIC_TOOL_CALL_FALLBACK = "llama3"
|
|
NO_DEDICATED_TOOL_FORMAT = {"LlamaForCausalLM"}
|
|
|
|
|
|
class _Config:
|
|
def __init__(self, data: dict) -> None:
|
|
self._data = data
|
|
|
|
def to_dict(self) -> dict:
|
|
return self._data
|
|
|
|
|
|
def _inferred(architecture: str) -> tuple[str, str | None]:
|
|
"""(tool_call_parser, reasoning_parser) that `auto` picks for this architecture."""
|
|
config = _Config({"architectures": [architecture], "torch_dtype": "bfloat16"})
|
|
with patch("freetoken.utils.cached_load_hf_config", lambda _path: config):
|
|
args, _run_shell = parse_args(["--model", ANON_PATH])
|
|
return args.tool_call_parser, args.reasoning_parser
|
|
|
|
|
|
@pytest.mark.parametrize("architecture", ARCHITECTURES)
|
|
def test_inferred_parser_names_are_names_the_factories_know(architecture):
|
|
"""A name the cascade invents but no factory can build fails at request time, not at boot."""
|
|
tool_call, reasoning = _inferred(architecture)
|
|
|
|
assert tool_call in FunctionCallParser.ToolCallParserEnum, tool_call
|
|
if reasoning is not None:
|
|
assert reasoning in ReasoningParser.ReasoningParserEnum, reasoning
|
|
|
|
|
|
def test_only_the_families_without_a_thinking_format_get_no_reasoning_parser():
|
|
fell_through = {a for a in ARCHITECTURES if _inferred(a)[1] is None}
|
|
assert fell_through == NO_REASONING_FORMAT
|
|
|
|
|
|
def test_only_the_families_without_a_tool_format_get_the_generic_fallback():
|
|
fell_through = {
|
|
a for a in ARCHITECTURES if _inferred(a)[0] == GENERIC_TOOL_CALL_FALLBACK
|
|
}
|
|
assert fell_through == NO_DEDICATED_TOOL_FORMAT
|
|
|
|
|
|
def test_qwen3_5_is_not_shadowed_by_the_generic_qwen_branch():
|
|
"""The cascade matches substrings in order, so the specific arm has to come first: a bare
|
|
``"qwen" -> qwen25`` reached earlier would swallow every later Qwen and lose its tool format."""
|
|
assert _inferred("Qwen3_5MoeForConditionalGeneration")[0] == "qwen3_coder"
|
|
assert _inferred("Qwen3_5ForConditionalGeneration")[0] == "qwen3_coder"
|
|
assert _inferred("Qwen3MoeForCausalLM")[0] == "qwen25"
|
|
|
|
|
|
def test_an_explicit_choice_beats_inference():
|
|
config = _Config({"architectures": ["DeepseekV4ForCausalLM"], "torch_dtype": "bfloat16"})
|
|
with patch("freetoken.utils.cached_load_hf_config", lambda _path: config):
|
|
off, _ = parse_args(["--model", ANON_PATH, "--reasoning-parser", "off"])
|
|
pinned, _ = parse_args(["--model", ANON_PATH, "--reasoning-parser", "qwen3"])
|
|
assert off.reasoning_parser is None
|
|
assert pinned.reasoning_parser == "qwen3"
|