| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import copy |
| 5 | import json |
| 6 | from pathlib import Path |
| 7 | |
| 8 | import pytest |
| 9 | |
| 10 | from tests.tokenizers_.test_deepseek_v4 import FakeHfTokenizer |
| 11 | from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest |
| 12 | from vllm.tokenizers.deepseek_v41 import get_deepseek_v41_tokenizer |
| 13 | |
| 14 | FIXTURES = Path(__file__).parent / "fixtures" |
| 15 | |
| 16 | |
| 17 | def render(messages, **kwargs): |
| 18 | return get_deepseek_v41_tokenizer(FakeHfTokenizer()).apply_chat_template( |
| 19 | messages, tokenize=False, **kwargs |
| 20 | ) |
| 21 | |
| 22 | |
| 23 | @pytest.mark.parametrize("case_id", [1, 2]) |
| 24 | def test_reference_encoder_fixtures(case_id): |
| 25 | # Expected prompts come from the DeepSeek V4.1 reference encoder. |
| 26 | data = json.loads( |
| 27 | (FIXTURES / "deepseek_v4" / f"test_input_{case_id}.json").read_text() |
| 28 | ) |
| 29 | messages = data["messages"] if isinstance(data, dict) else data |
| 30 | tools = data.get("tools") if isinstance(data, dict) else None |
| 31 | expected = (FIXTURES / "deepseek_v41" / f"test_output_{case_id}.txt").read_text() |
| 32 | assert render(messages, tools=tools) == expected |
| 33 | |
| 34 | |
| 35 | @pytest.mark.parametrize( |
| 36 | ("effort", "budget"), |
| 37 | [ |
| 38 | (None, 75), |
| 39 | ("low", 50), |
| 40 | ("high", 75), |
| 41 | ("xhigh", 75), |
| 42 | ("max", 100), |
| 43 | (1, 1), |
| 44 | (42, 42), |
| 45 | (100, 100), |
| 46 | ], |
| 47 | ) |
| 48 | def test_numeric_reasoning_effort(effort, budget): |
| 49 | assert render( |
| 50 | [{"role": "user", "content": "question"}], reasoning_effort=effort |
| 51 | ) == ( |
| 52 | "<｜begin▁of▁sentence｜><｜System｜>" |
| 53 | f"Reasoning Effort: {budget} (range 1-100, the higher the value, " |
| 54 | "the more thorough the reasoning)\n\n" |
| 55 | "<｜User｜>question<｜Assistant｜><think>" |
| 56 | ) |
| 57 | |
| 58 | |
| 59 | @pytest.mark.parametrize("effort", ["medium", "minimal", -1, 0, 101, True, 1.5, []]) |
| 60 | def test_invalid_effort_is_a_request_error(effort): |
| 61 | with pytest.raises(ValueError, match="reasoning_effort"): |
| 62 | render([{"role": "user", "content": "question"}], reasoning_effort=effort) |
| 63 | |
| 64 | |
| 65 | @pytest.mark.parametrize( |
| 66 | "controls", |
| 67 | [ |
| 68 | {"thinking": False}, |
| 69 | {"enable_thinking": False}, |
| 70 | {"thinking": True, "reasoning_effort": "none"}, |
| 71 | {"enable_thinking": True, "reasoning_effort": "none"}, |
| 72 | ], |
| 73 | ) |
| 74 | def test_chat_mode_has_closed_thinking_prefix(controls): |
| 75 | assert render([{"role": "user", "content": "question"}], **controls) == ( |
| 76 | "<｜begin▁of▁sentence｜><｜User｜>question<｜Assistant｜></think>" |
| 77 | ) |
| 78 | |
| 79 | |
| 80 | def test_raw_text_parts_preserve_reference_separator(): |
| 81 | messages = [ |
| 82 | { |
| 83 | "role": "user", |
| 84 | "content": [ |
| 85 | {"type": "text", "text": "first"}, |
| 86 | {"type": "text", "text": "second"}, |
| 87 | ], |
| 88 | } |
| 89 | ] |
| 90 | original = copy.deepcopy(messages) |
| 91 | assert ( |
| 92 | render( |
| 93 | messages, |
| 94 | conversation=[{"role": "user", "content": "first\nsecond"}], |
| 95 | thinking=False, |
| 96 | ) |
| 97 | == "<｜begin▁of▁sentence｜><｜User｜>first\n\nsecond<｜Assistant｜></think>" |
| 98 | ) |
| 99 | assert messages == original |
| 100 | |
| 101 | |
| 102 | @pytest.mark.parametrize( |
| 103 | ("role", "responses_type"), |
| 104 | [("user", "input_text"), ("assistant", "output_text")], |
| 105 | ) |
| 106 | def test_responses_text_parts_match_chat_text_parts(role, responses_type): |
| 107 | responses_message = { |
| 108 | "role": role, |
| 109 | "content": [{"type": responses_type, "text": "hello"}], |
| 110 | } |
| 111 | chat_message = { |
| 112 | "role": role, |
| 113 | "content": [{"type": "text", "text": "hello"}], |
| 114 | } |
| 115 | assert render([responses_message], thinking=False) == render( |
| 116 | [chat_message], thinking=False |
| 117 | ) |
| 118 | |
| 119 | |
| 120 | def test_mid_system_gets_its_own_marker_and_generation_header(): |
| 121 | assert render( |
| 122 | [ |
| 123 | {"role": "user", "content": "question"}, |
| 124 | {"role": "system", "content": "answer briefly"}, |
| 125 | ], |
| 126 | thinking=False, |
| 127 | ) == ( |
| 128 | "<｜begin▁of▁sentence｜><｜User｜>question" |
| 129 | "<｜System｜>answer briefly<｜Assistant｜></think>" |
| 130 | ) |
| 131 | |
| 132 | |
| 133 | def test_top_level_effort_overrides_template_effort(): |
| 134 | request = ChatCompletionRequest( |
| 135 | model="deepseek-ai/DeepSeek-V4.1-Flash", |
| 136 | messages=[{"role": "user", "content": "question"}], |
| 137 | reasoning_effort="low", |
| 138 | chat_template_kwargs={"reasoning_effort": 100}, |
| 139 | ) |
| 140 | kwargs = request.build_chat_params(None, "auto").get_apply_chat_template_kwargs() |
| 141 | assert "Reasoning Effort: 50 " in render(request.messages, **kwargs) |
| 142 | |
| 143 | |
| 144 | def test_encode_uses_one_bos_and_forwards_truncation(): |
| 145 | tokenizer = get_deepseek_v41_tokenizer(FakeHfTokenizer()) |
| 146 | tokenizer.apply_chat_template( |
| 147 | [{"role": "user", "content": "question"}], |
| 148 | thinking=False, |
| 149 | truncation=True, |
| 150 | max_length=16, |
| 151 | ) |
| 152 | text, add_special_tokens, kwargs = tokenizer.last_encode |
| 153 | assert text.startswith("<｜begin▁of▁sentence｜>") |
| 154 | assert add_special_tokens is False |
| 155 | assert kwargs == {"truncation": True, "max_length": 16} |
| 156 | |
| 157 | |
| 158 | @pytest.mark.parametrize("image_type", ["image_url", "input_image", "image_pil"]) |
| 159 | def test_images_preserve_content_order_and_reference_separator(image_type): |
| 160 | from PIL import Image |
| 161 | |
| 162 | def image_part(color): |
| 163 | if image_type == "image_pil": |
| 164 | return {"type": image_type, "image_pil": Image.new("RGB", (2, 2), color)} |
| 165 | url = f"https://example.com/{color}.png" |
| 166 | return { |
| 167 | "type": image_type, |
| 168 | "image_url": {"url": url} if image_type == "image_url" else url, |
| 169 | } |
| 170 | |
| 171 | parts = [ |
| 172 | {"type": "text", "text": "first"}, |
| 173 | image_part("red"), |
| 174 | {"type": "text", "text": "second"}, |
| 175 | image_part("blue"), |
| 176 | {"type": "text", "text": "compare"}, |
| 177 | ] |
| 178 | assert render([{"role": "user", "content": parts}], thinking=False) == ( |
| 179 | "<｜begin▁of▁sentence｜><｜User｜>first\n\n<｜deepseek_image｜>\n\n" |
| 180 | "second\n\n<｜deepseek_image｜>\n\ncompare<｜Assistant｜></think>" |
| 181 | ) |
| 182 | assert parts[1]["type"] == parts[3]["type"] == image_type |
| 183 | |
| 184 | |
| 185 | def test_unsupported_media_is_a_request_error(): |
| 186 | with pytest.raises(ValueError, match="text and image content only"): |
| 187 | render([{"role": "user", "content": [{"type": "input_audio"}]}]) |