documentlanguagemodel Public

Watch 0 Fork 0 Star 0
Python · 10775 bytes Raw Blame History
1 """Sprint 40 registry refresh — first 2026 text-family entries."""
2
3 from __future__ import annotations
4
5 import pytest
6
7 from dlm.base_models import BASE_MODELS
8
9 _QWEN3_KEYS = ("qwen3-1.7b", "qwen3-4b", "qwen3-8b")
10
11
12 class TestQwen3RegistryEntries:
13 @pytest.mark.parametrize("key", _QWEN3_KEYS)
14 def test_entry_present(self, key: str) -> None:
15 assert key in BASE_MODELS
16
17 @pytest.mark.parametrize("key", _QWEN3_KEYS)
18 def test_qwen3_family_uses_expected_architecture(self, key: str) -> None:
19 spec = BASE_MODELS[key]
20 assert spec.architecture == "Qwen3ForCausalLM"
21 assert spec.template == "chatml"
22 assert spec.gguf_arch == "qwen3"
23 assert spec.tokenizer_pre == "qwen2"
24
25 @pytest.mark.parametrize("key", _QWEN3_KEYS)
26 def test_qwen3_family_is_open_and_reasoning_tuned(self, key: str) -> None:
27 spec = BASE_MODELS[key]
28 assert spec.license_spdx == "Apache-2.0"
29 assert spec.requires_acceptance is False
30 assert spec.redistributable is True
31 assert spec.reasoning_tuned is True
32 assert spec.suggested_prompt_temperature == pytest.approx(0.6)
33
34 def test_qwen3_8b_uses_upper_bound_size_hint(self) -> None:
35 spec = BASE_MODELS["qwen3-8b"]
36 assert spec.params == 8_000_000_000
37 assert spec.size_gb_fp16 == pytest.approx(16.0)
38
39
40 class TestQwen3ThinkingRegistryEntry:
41 def test_entry_present(self) -> None:
42 assert "qwen3-1.7b-thinking" in BASE_MODELS
43
44 def test_reuses_live_qwen3_weights_with_reasoning_profile(self) -> None:
45 spec = BASE_MODELS["qwen3-1.7b-thinking"]
46 assert spec.hf_id == "Qwen/Qwen3-1.7B"
47 assert spec.architecture == "Qwen3ForCausalLM"
48 assert spec.template == "qwen3thinking"
49 assert spec.gguf_arch == "qwen3"
50 assert spec.tokenizer_pre == "qwen2"
51
52 def test_reasoning_profile_keeps_open_license_and_cooler_default(self) -> None:
53 spec = BASE_MODELS["qwen3-1.7b-thinking"]
54 assert spec.license_spdx == "Apache-2.0"
55 assert spec.requires_acceptance is False
56 assert spec.redistributable is True
57 assert spec.reasoning_tuned is True
58 assert spec.suggested_prompt_temperature == pytest.approx(0.6)
59
60
61 class TestLlama33RegistryEntry:
62 def test_entry_present(self) -> None:
63 assert "llama-3.3-8b-instruct" in BASE_MODELS
64
65 def test_keeps_existing_llama_policy_surface(self) -> None:
66 spec = BASE_MODELS["llama-3.3-8b-instruct"]
67 assert spec.hf_id == "allura-forge/Llama-3.3-8B-Instruct"
68 assert spec.architecture == "LlamaForCausalLM"
69 assert spec.template == "llama3"
70 assert spec.gguf_arch == "llama"
71 assert spec.requires_acceptance is True
72 assert spec.redistributable is False
73 assert spec.license_spdx == "Other"
74
75 def test_refreshes_against_mirror_plus_official_provenance_page(self) -> None:
76 spec = BASE_MODELS["llama-3.3-8b-instruct"]
77 assert spec.refresh_check_hf_gating is False
78 assert spec.provenance_url == (
79 "https://about.fb.com/br/news/2025/04/tudo-o-que-anunciamos-no-nosso-primeiro-llamacon/"
80 )
81 assert spec.provenance_match_text == "novo modelo Llama 3.3 8B"
82
83 def test_effective_context_hint_is_lower_than_nominal(self) -> None:
84 spec = BASE_MODELS["llama-3.3-8b-instruct"]
85 assert spec.context_length == 131_072
86 assert spec.context_length_effective == 8_192
87 assert spec.effective_context_length == 8_192
88
89 def test_license_url_uses_current_meta_landing_page(self) -> None:
90 spec = BASE_MODELS["llama-3.3-8b-instruct"]
91 assert spec.license_url == "https://llama.meta.com/llama3/license"
92
93
94 class TestSmolLM3RegistryEntry:
95 def test_entry_present(self) -> None:
96 assert "smollm3-3b" in BASE_MODELS
97
98 def test_live_hf_id_and_architecture_match_current_release(self) -> None:
99 spec = BASE_MODELS["smollm3-3b"]
100 assert spec.hf_id == "HuggingFaceTB/SmolLM3-3B"
101 assert spec.architecture == "SmolLM3ForCausalLM"
102 assert spec.template == "smollm3"
103 assert spec.gguf_arch == "llama"
104 assert spec.tokenizer_pre == "smollm"
105
106 def test_entry_tracks_reasoning_first_defaults(self) -> None:
107 spec = BASE_MODELS["smollm3-3b"]
108 assert spec.reasoning_tuned is True
109 assert spec.suggested_prompt_temperature == pytest.approx(0.6)
110
111 def test_entry_is_open_and_long_context(self) -> None:
112 spec = BASE_MODELS["smollm3-3b"]
113 assert spec.license_spdx == "Apache-2.0"
114 assert spec.requires_acceptance is False
115 assert spec.redistributable is True
116 assert spec.context_length == 65_536
117 assert spec.recommended_seq_len == 4096
118
119
120 class TestOlmo2RegistryEntry:
121 def test_entry_present(self) -> None:
122 assert "olmo-2-7b-instruct" in BASE_MODELS
123
124 def test_live_hf_id_and_architecture_match_current_release(self) -> None:
125 spec = BASE_MODELS["olmo-2-7b-instruct"]
126 assert spec.hf_id == "allenai/OLMo-2-1124-7B-Instruct"
127 assert spec.architecture == "Olmo2ForCausalLM"
128 assert spec.template == "olmo2"
129 assert spec.gguf_arch == "olmo2"
130 assert spec.tokenizer_pre == "superbpe"
131
132 def test_entry_is_open_without_reasoning_override(self) -> None:
133 spec = BASE_MODELS["olmo-2-7b-instruct"]
134 assert spec.license_spdx == "Apache-2.0"
135 assert spec.requires_acceptance is False
136 assert spec.redistributable is True
137 assert spec.reasoning_tuned is False
138
139 def test_entry_uses_model_card_context_and_size_hints(self) -> None:
140 spec = BASE_MODELS["olmo-2-7b-instruct"]
141 assert spec.params == 7_000_000_000
142 assert spec.size_gb_fp16 == pytest.approx(14.6)
143 assert spec.context_length == 4096
144 assert spec.recommended_seq_len == 2048
145
146
147 class TestGemma2RegistryEntry:
148 @pytest.mark.parametrize("key", ["gemma-2-2b-it", "gemma-2-9b-it"])
149 def test_entry_present(self, key: str) -> None:
150 assert key in BASE_MODELS
151
152 @pytest.mark.parametrize(
153 ("key", "hf_id", "params", "size_gb_fp16"),
154 [
155 ("gemma-2-2b-it", "google/gemma-2-2b-it", 2_600_000_000, 5.2),
156 ("gemma-2-9b-it", "google/gemma-2-9b-it", 9_000_000_000, 18.0),
157 ],
158 )
159 def test_live_hf_id_and_architecture_match_current_release(
160 self,
161 key: str,
162 hf_id: str,
163 params: int,
164 size_gb_fp16: float,
165 ) -> None:
166 spec = BASE_MODELS[key]
167 assert spec.hf_id == hf_id
168 assert spec.architecture == "Gemma2ForCausalLM"
169 assert spec.template == "gemma2"
170 assert spec.gguf_arch == "gemma2"
171 assert spec.tokenizer_pre == "gemma"
172 assert spec.params == params
173 assert spec.size_gb_fp16 == pytest.approx(size_gb_fp16)
174
175 @pytest.mark.parametrize("key", ["gemma-2-2b-it", "gemma-2-9b-it"])
176 def test_entry_follows_gemma_gating_pattern(self, key: str) -> None:
177 spec = BASE_MODELS[key]
178 assert spec.license_spdx == "Gemma"
179 assert spec.license_url == "https://ai.google.dev/gemma/terms"
180 assert spec.requires_acceptance is True
181 assert spec.redistributable is False
182
183 @pytest.mark.parametrize("key", ["gemma-2-2b-it", "gemma-2-9b-it"])
184 def test_entry_uses_model_card_context_hints(self, key: str) -> None:
185 spec = BASE_MODELS[key]
186 assert spec.context_length == 8192
187 assert spec.recommended_seq_len == 2048
188
189
190 class TestPhi4MiniReasoningRegistryEntry:
191 def test_entry_present(self) -> None:
192 assert "phi-4-mini-reasoning" in BASE_MODELS
193
194 def test_live_hf_id_and_architecture_match_current_release(self) -> None:
195 spec = BASE_MODELS["phi-4-mini-reasoning"]
196 assert spec.hf_id == "microsoft/Phi-4-mini-reasoning"
197 assert spec.architecture == "Phi3ForCausalLM"
198 assert spec.template == "phi4mini"
199 assert spec.gguf_arch == "phi3"
200 assert spec.tokenizer_pre == "phi-2"
201
202 def test_entry_is_reasoning_tuned_and_open(self) -> None:
203 spec = BASE_MODELS["phi-4-mini-reasoning"]
204 assert spec.license_spdx == "MIT"
205 assert spec.requires_acceptance is False
206 assert spec.redistributable is True
207 assert spec.reasoning_tuned is True
208 assert spec.suggested_prompt_temperature == pytest.approx(0.6)
209
210 def test_entry_uses_model_card_context_and_size_hints(self) -> None:
211 spec = BASE_MODELS["phi-4-mini-reasoning"]
212 assert spec.params == 3_800_000_000
213 assert spec.size_gb_fp16 == pytest.approx(7.6)
214 assert spec.context_length == 131_072
215 assert spec.recommended_seq_len == 2048
216
217
218 class TestMixtralRegistryEntry:
219 def test_entry_present(self) -> None:
220 assert "mixtral-8x7b-instruct" in BASE_MODELS
221
222 def test_entry_is_open_sparse_moe(self) -> None:
223 spec = BASE_MODELS["mixtral-8x7b-instruct"]
224 assert spec.license_spdx == "Apache-2.0"
225 assert spec.requires_acceptance is False
226 assert spec.redistributable is True
227 assert spec.modality == "text-moe"
228
229 def test_entry_matches_live_family_shape(self) -> None:
230 spec = BASE_MODELS["mixtral-8x7b-instruct"]
231 assert spec.hf_id == "mistralai/Mixtral-8x7B-Instruct-v0.1"
232 assert spec.architecture == "MixtralForCausalLM"
233 assert spec.template == "mistral"
234 assert spec.gguf_arch == "llama"
235 assert spec.tokenizer_pre == "llama-bpe"
236
237 def test_entry_uses_total_parameter_and_context_hints(self) -> None:
238 spec = BASE_MODELS["mixtral-8x7b-instruct"]
239 assert spec.params == 46_700_000_000
240 assert spec.size_gb_fp16 == pytest.approx(93.4)
241 assert spec.context_length == 32_768
242 assert spec.recommended_seq_len == 2048
243
244
245 class TestInternVL3RegistryEntry:
246 def test_entry_present(self) -> None:
247 assert "internvl3-2b" in BASE_MODELS
248
249 def test_entry_keeps_remote_code_contract_explicit(self) -> None:
250 spec = BASE_MODELS["internvl3-2b"]
251 assert spec.hf_id == "OpenGVLab/InternVL3-2B"
252 assert spec.architecture == "InternVLChatModel"
253 assert spec.template == "internvl2"
254 assert spec.trust_remote_code is True
255
256 def test_entry_is_registry_visible_but_not_pretending_runtime_is_generic(self) -> None:
257 spec = BASE_MODELS["internvl3-2b"]
258 assert spec.license_spdx == "Apache-2.0"
259 assert spec.requires_acceptance is False
260 assert spec.redistributable is True
261 assert spec.modality == "vision-language"
262 assert spec.vl_preprocessor_plan is not None
263 assert spec.vl_preprocessor_plan.resize_policy == "dynamic"
264 assert spec.vl_preprocessor_plan.image_token == "<image>"