Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
"""datasets pads plain messages with `tool_calls: None`; templates must test the value."""
import pytest
from tokenizers import Tokenizer, models
from transformers import PreTrainedTokenizerFast
from unsloth.chat_templates import get_chat_template
TEMPLATES = ["llama-3.1", "llama-3.2", "llama-3.3", "gpt-oss"]
CALL = [{"type": "function", "function": {"name": "get_weather", "arguments": {"city": "Paris"}}}]
PLAIN = [
{"role": "system", "content": "Be brief."},
{"role": "user", "content": "Hi"},
{"role": "assistant", "content": "Hello!"},
{"role": "user", "content": "Bye"},
{"role": "assistant", "content": "Bye!"},
]
TOOL_FLOW = [
{"role": "user", "content": "weather?"},
{"role": "assistant", "content": None, "tool_calls": CALL},
{"role": "tool", "content": "sunny"},
{"role": "assistant", "content": "It is sunny."},
]
@pytest.fixture(autouse = True)
def _scratch_cwd(tmp_path, monkeypatch):
# get_chat_template writes a scratch dir under CWD; isolate it per test and xdist worker.
monkeypatch.chdir(tmp_path)
def _tokenizer(name):
vocab = {"<unk>": 0, "<s>": 1, "</s>": 2, "<pad>": 3}
tokenizer = PreTrainedTokenizerFast(
tokenizer_object = Tokenizer(models.WordLevel(vocab, unk_token = "<unk>")),
bos_token = "<s>",
eos_token = "</s>",
unk_token = "<unk>",
pad_token = "<pad>",
)
return get_chat_template(tokenizer, name)
def _render(tokenizer, messages, add_generation_prompt):
return tokenizer.apply_chat_template(
messages, tokenize = False, add_generation_prompt = add_generation_prompt
)
@pytest.mark.parametrize("filler", [None, []], ids = ["none", "empty_list"])
@pytest.mark.parametrize("add_generation_prompt", [False, True])
@pytest.mark.parametrize("conversation", [PLAIN, TOOL_FLOW], ids = ["plain", "tool_flow"])
@pytest.mark.parametrize("name", TEMPLATES)
def test_blank_tool_calls_match_a_missing_key(name, conversation, add_generation_prompt, filler):
tokenizer = _tokenizer(name)
padded = [dict(m, tool_calls = m.get("tool_calls", filler)) for m in conversation]
expected = _render(tokenizer, conversation, add_generation_prompt)
assert ("get_weather" in expected) == (conversation is TOOL_FLOW)
assert _render(tokenizer, padded, add_generation_prompt) == expected
def test_gpt_oss_drops_the_tool_call_analysis_when_a_final_answer_follows():
conversation = [
TOOL_FLOW[0],
{"role": "assistant", "content": None, "thinking": "Check it.", "tool_calls": CALL},
TOOL_FLOW[2],
TOOL_FLOW[3],
]
padded = [dict(m, tool_calls = m.get("tool_calls")) for m in conversation]
tokenizer = _tokenizer("gpt-oss")
expected = _render(tokenizer, conversation, False)
assert "Check it." not in expected
assert _render(tokenizer, padded, False) == expected