@@ -2,9 +2,13 @@ |
| 2 | 2 | |
| 3 | 3 | from __future__ import annotations |
| 4 | 4 | |
| 5 | +import builtins |
| 6 | +import json |
| 7 | +import sys |
| 8 | +import urllib.error |
| 5 | 9 | from pathlib import Path |
| 6 | | -from types import SimpleNamespace |
| 7 | | -from typing import Any |
| 10 | +from types import ModuleType, SimpleNamespace |
| 11 | +from typing import Any, Literal |
| 8 | 12 | |
| 9 | 13 | import pytest |
| 10 | 14 | |
@@ -15,6 +19,7 @@ from dlm.synth import ( |
| 15 | 19 | InvalidTeacherSpecError, |
| 16 | 20 | OpenAiTeacher, |
| 17 | 21 | SelfTeacher, |
| 22 | + TeacherInvocationError, |
| 18 | 23 | TeacherUnavailableError, |
| 19 | 24 | VllmServerTeacher, |
| 20 | 25 | build_teacher, |
@@ -22,6 +27,13 @@ from dlm.synth import ( |
| 22 | 27 | ) |
| 23 | 28 | |
| 24 | 29 | |
| 30 | +def _module(name: str, **attrs: object) -> ModuleType: |
| 31 | + module = ModuleType(name) |
| 32 | + for key, value in attrs.items(): |
| 33 | + setattr(module, key, value) |
| 34 | + return module |
| 35 | + |
| 36 | + |
| 25 | 37 | class TestTeacherSelectorParsing: |
| 26 | 38 | @pytest.mark.parametrize( |
| 27 | 39 | ("raw", "kind", "target"), |
@@ -46,6 +58,19 @@ class TestTeacherSelectorParsing: |
| 46 | 58 | with pytest.raises(InvalidTeacherSpecError, match="unknown teacher selector"): |
| 47 | 59 | parse_teacher_ref("mystery:thing") |
| 48 | 60 | |
| 61 | + @pytest.mark.parametrize( |
| 62 | + ("raw", "message"), |
| 63 | + [ |
| 64 | + ("hf: ", "hf teacher selector must include a model id"), |
| 65 | + ("openai: ", "openai teacher selector must include a model id"), |
| 66 | + ("anthropic: ", "anthropic teacher selector must include a model id"), |
| 67 | + ("vllm-server: ", "vllm-server teacher selector must include a URL"), |
| 68 | + ], |
| 69 | + ) |
| 70 | + def test_missing_selector_targets_are_refused(self, raw: str, message: str) -> None: |
| 71 | + with pytest.raises(InvalidTeacherSpecError, match=message): |
| 72 | + parse_teacher_ref(raw) |
| 73 | + |
| 49 | 74 | |
| 50 | 75 | class TestBuildTeacher: |
| 51 | 76 | def test_self_requires_dlm_path(self) -> None: |
@@ -103,6 +128,10 @@ class TestSelfTeacher: |
| 103 | 128 | |
| 104 | 129 | |
| 105 | 130 | class TestHfTeacher: |
| 131 | + def test_blank_hf_id_refused(self) -> None: |
| 132 | + with pytest.raises(InvalidTeacherSpecError, match="must include a model id"): |
| 133 | + HfTeacher(" ") |
| 134 | + |
| 106 | 135 | def test_hf_teacher_uses_loader_and_runner(self) -> None: |
| 107 | 136 | seen: dict[str, Any] = {} |
| 108 | 137 | |
@@ -134,8 +163,31 @@ class TestHfTeacher: |
| 134 | 163 | ) |
| 135 | 164 | assert seen["runner"][3:] == (21, 0.5, 0.8, 11) |
| 136 | 165 | |
| 166 | + def test_hf_teacher_reuses_loaded_bundle(self) -> None: |
| 167 | + loads: list[tuple[str, str]] = [] |
| 168 | + |
| 169 | + def _loader(hf_id: str, device: str) -> teachers_mod._LoadedHfTeacher: |
| 170 | + loads.append((hf_id, device)) |
| 171 | + return teachers_mod._LoadedHfTeacher(model="model", tokenizer="tok", device=device) |
| 172 | + |
| 173 | + teacher = HfTeacher( |
| 174 | + "Qwen/Qwen2.5-1.5B-Instruct", |
| 175 | + loader=_loader, |
| 176 | + runner=lambda *_args, **_kwargs: "ok", |
| 177 | + ) |
| 178 | + |
| 179 | + assert teacher.generate("system", "user") == "ok" |
| 180 | + assert teacher.generate("system", "user") == "ok" |
| 181 | + assert loads == [ |
| 182 | + ("Qwen/Qwen2.5-1.5B-Instruct", teachers_mod._resolve_generation_device("auto")) |
| 183 | + ] |
| 184 | + |
| 137 | 185 | |
| 138 | 186 | class TestOpenAiTeacher: |
| 187 | + def test_blank_model_refused(self) -> None: |
| 188 | + with pytest.raises(InvalidTeacherSpecError, match="must include a model id"): |
| 189 | + OpenAiTeacher(" ") |
| 190 | + |
| 139 | 191 | def test_missing_api_key_refused(self, monkeypatch: pytest.MonkeyPatch) -> None: |
| 140 | 192 | monkeypatch.delenv("OPENAI_API_KEY", raising=False) |
| 141 | 193 | teacher = OpenAiTeacher("gpt-4o-mini") |
@@ -145,28 +197,59 @@ class TestOpenAiTeacher: |
| 145 | 197 | def test_openai_teacher_extracts_message_text(self, monkeypatch: pytest.MonkeyPatch) -> None: |
| 146 | 198 | monkeypatch.setenv("OPENAI_API_KEY", "secret") |
| 147 | 199 | |
| 148 | | - captured: dict[str, Any] = {} |
| 200 | + payloads: list[dict[str, Any]] = [] |
| 201 | + factories: list[str] = [] |
| 149 | 202 | |
| 150 | 203 | def _create(**kwargs: Any) -> Any: |
| 151 | | - captured["payload"] = kwargs |
| 204 | + payloads.append(kwargs) |
| 152 | 205 | return SimpleNamespace( |
| 153 | 206 | choices=[SimpleNamespace(message=SimpleNamespace(content=" generated "))] |
| 154 | 207 | ) |
| 155 | 208 | |
| 209 | + def _factory(api_key: str) -> Any: |
| 210 | + factories.append(api_key) |
| 211 | + return client |
| 212 | + |
| 156 | 213 | client = SimpleNamespace( |
| 157 | 214 | chat=SimpleNamespace( |
| 158 | 215 | completions=SimpleNamespace(create=_create), |
| 159 | 216 | ) |
| 160 | 217 | ) |
| 161 | 218 | |
| 162 | | - teacher = OpenAiTeacher("gpt-4o-mini", client_factory=lambda api_key: client) |
| 219 | + teacher = OpenAiTeacher( |
| 220 | + "gpt-4o-mini", |
| 221 | + client_factory=_factory, |
| 222 | + ) |
| 163 | 223 | out = teacher.generate("sys", "usr", max_new_tokens=17, temperature=0.3, top_p=0.7, seed=5) |
| 224 | + second = teacher.generate("sys", "usr") |
| 164 | 225 | assert out == "generated" |
| 165 | | - assert captured["payload"]["model"] == "gpt-4o-mini" |
| 166 | | - assert captured["payload"]["seed"] == 5 |
| 226 | + assert second == "generated" |
| 227 | + assert payloads[0]["model"] == "gpt-4o-mini" |
| 228 | + assert payloads[0]["seed"] == 5 |
| 229 | + assert factories == ["secret"] |
| 230 | + |
| 231 | + def test_openai_teacher_wraps_request_failures(self, monkeypatch: pytest.MonkeyPatch) -> None: |
| 232 | + monkeypatch.setenv("OPENAI_API_KEY", "secret") |
| 233 | + |
| 234 | + def _create(**_kwargs: Any) -> Any: |
| 235 | + raise RuntimeError("boom") |
| 236 | + |
| 237 | + client = SimpleNamespace( |
| 238 | + chat=SimpleNamespace( |
| 239 | + completions=SimpleNamespace(create=_create), |
| 240 | + ) |
| 241 | + ) |
| 242 | + teacher = OpenAiTeacher("gpt-4o-mini", client_factory=lambda _api_key: client) |
| 243 | + |
| 244 | + with pytest.raises(TeacherInvocationError, match="openai:gpt-4o-mini request failed: boom"): |
| 245 | + teacher.generate("sys", "usr") |
| 167 | 246 | |
| 168 | 247 | |
| 169 | 248 | class TestAnthropicTeacher: |
| 249 | + def test_blank_model_refused(self) -> None: |
| 250 | + with pytest.raises(InvalidTeacherSpecError, match="must include a model id"): |
| 251 | + AnthropicTeacher(" ") |
| 252 | + |
| 170 | 253 | def test_missing_api_key_refused(self, monkeypatch: pytest.MonkeyPatch) -> None: |
| 171 | 254 | monkeypatch.delenv("ANTHROPIC_API_KEY", raising=False) |
| 172 | 255 | teacher = AnthropicTeacher("claude-3-5-haiku-latest") |
@@ -176,6 +259,7 @@ class TestAnthropicTeacher: |
| 176 | 259 | def test_anthropic_teacher_extracts_text_blocks(self, monkeypatch: pytest.MonkeyPatch) -> None: |
| 177 | 260 | monkeypatch.setenv("ANTHROPIC_API_KEY", "secret") |
| 178 | 261 | captured: dict[str, Any] = {} |
| 262 | + factories: list[str] = [] |
| 179 | 263 | |
| 180 | 264 | class _Messages: |
| 181 | 265 | @staticmethod |
@@ -183,6 +267,7 @@ class TestAnthropicTeacher: |
| 183 | 267 | captured["payload"] = kwargs |
| 184 | 268 | return SimpleNamespace( |
| 185 | 269 | content=[ |
| 270 | + SimpleNamespace(type="image", text="ignored"), |
| 186 | 271 | SimpleNamespace(type="text", text=" first "), |
| 187 | 272 | SimpleNamespace(type="text", text=" second "), |
| 188 | 273 | ] |
@@ -191,16 +276,51 @@ class TestAnthropicTeacher: |
| 191 | 276 | class _Client: |
| 192 | 277 | messages = _Messages() |
| 193 | 278 | |
| 279 | + def _factory(api_key: str) -> _Client: |
| 280 | + factories.append(api_key) |
| 281 | + return _Client() |
| 282 | + |
| 194 | 283 | teacher = AnthropicTeacher( |
| 195 | 284 | "claude-3-5-haiku-latest", |
| 196 | | - client_factory=lambda api_key: _Client(), |
| 285 | + client_factory=_factory, |
| 197 | 286 | ) |
| 198 | 287 | out = teacher.generate("sys", "usr", max_new_tokens=19, temperature=0.2, top_p=0.6) |
| 288 | + second = teacher.generate("sys", "usr") |
| 199 | 289 | assert out == "first\nsecond" |
| 290 | + assert second == "first\nsecond" |
| 200 | 291 | assert captured["payload"]["model"] == "claude-3-5-haiku-latest" |
| 292 | + assert factories == ["secret"] |
| 293 | + |
| 294 | + def test_anthropic_teacher_wraps_request_failures( |
| 295 | + self, monkeypatch: pytest.MonkeyPatch |
| 296 | + ) -> None: |
| 297 | + monkeypatch.setenv("ANTHROPIC_API_KEY", "secret") |
| 298 | + |
| 299 | + class _Messages: |
| 300 | + @staticmethod |
| 301 | + def create(**_kwargs: Any) -> Any: |
| 302 | + raise RuntimeError("boom") |
| 303 | + |
| 304 | + class _Client: |
| 305 | + messages = _Messages() |
| 306 | + |
| 307 | + teacher = AnthropicTeacher( |
| 308 | + "claude-3-5-haiku-latest", |
| 309 | + client_factory=lambda _api_key: _Client(), |
| 310 | + ) |
| 311 | + |
| 312 | + with pytest.raises( |
| 313 | + TeacherInvocationError, |
| 314 | + match="anthropic:claude-3-5-haiku-latest request failed: boom", |
| 315 | + ): |
| 316 | + teacher.generate("sys", "usr") |
| 201 | 317 | |
| 202 | 318 | |
| 203 | 319 | class TestVllmServerTeacher: |
| 320 | + def test_blank_url_refused(self) -> None: |
| 321 | + with pytest.raises(InvalidTeacherSpecError, match="must include a URL"): |
| 322 | + VllmServerTeacher(" ") |
| 323 | + |
| 204 | 324 | def test_invalid_url_refused(self) -> None: |
| 205 | 325 | with pytest.raises(InvalidTeacherSpecError, match="http\\(s\\)"): |
| 206 | 326 | VllmServerTeacher("localhost:8000") |
@@ -208,10 +328,11 @@ class TestVllmServerTeacher: |
| 208 | 328 | def test_vllm_teacher_queries_model_and_completion( |
| 209 | 329 | self, monkeypatch: pytest.MonkeyPatch |
| 210 | 330 | ) -> None: |
| 211 | | - calls: dict[str, Any] = {} |
| 331 | + model_calls: list[tuple[str, float]] = [] |
| 332 | + completion_calls: list[tuple[Any, ...]] = [] |
| 212 | 333 | |
| 213 | 334 | def _fake_models(base_url: str, *, request_timeout: float) -> str | None: |
| 214 | | - calls["models"] = (base_url, request_timeout) |
| 335 | + model_calls.append((base_url, request_timeout)) |
| 215 | 336 | return "demo-model" |
| 216 | 337 | |
| 217 | 338 | def _fake_completion( |
@@ -225,15 +346,17 @@ class TestVllmServerTeacher: |
| 225 | 346 | seed: int | None, |
| 226 | 347 | request_timeout: float, |
| 227 | 348 | ) -> str: |
| 228 | | - calls["completion"] = ( |
| 229 | | - base_url, |
| 230 | | - model_id, |
| 231 | | - messages, |
| 232 | | - max_new_tokens, |
| 233 | | - temperature, |
| 234 | | - top_p, |
| 235 | | - seed, |
| 236 | | - request_timeout, |
| 349 | + completion_calls.append( |
| 350 | + ( |
| 351 | + base_url, |
| 352 | + model_id, |
| 353 | + messages, |
| 354 | + max_new_tokens, |
| 355 | + temperature, |
| 356 | + top_p, |
| 357 | + seed, |
| 358 | + request_timeout, |
| 359 | + ) |
| 237 | 360 | ) |
| 238 | 361 | return " served " |
| 239 | 362 | |
@@ -242,8 +365,734 @@ class TestVllmServerTeacher: |
| 242 | 365 | |
| 243 | 366 | teacher = VllmServerTeacher("http://127.0.0.1:8000") |
| 244 | 367 | out = teacher.generate("sys", "usr", max_new_tokens=29, temperature=0.4, top_p=0.75, seed=9) |
| 368 | + second = teacher.generate("sys", "usr") |
| 245 | 369 | |
| 246 | 370 | assert out == "served" |
| 247 | | - assert calls["models"] == ("http://127.0.0.1:8000", 30.0) |
| 248 | | - assert calls["completion"][1] == "demo-model" |
| 249 | | - assert calls["completion"][3:] == (29, 0.4, 0.75, 9, 30.0) |
| 371 | + assert second == "served" |
| 372 | + assert model_calls == [("http://127.0.0.1:8000", 30.0)] |
| 373 | + assert completion_calls[0][1] == "demo-model" |
| 374 | + assert completion_calls[0][3:] == (29, 0.4, 0.75, 9, 30.0) |
| 375 | + |
| 376 | + |
| 377 | +class TestTeacherHelpers: |
| 378 | + def test_flatten_teacher_prompt_handles_partial_inputs(self) -> None: |
| 379 | + assert teachers_mod._flatten_teacher_prompt("system", "user").startswith("System:\n") |
| 380 | + assert teachers_mod._flatten_teacher_prompt("", "user") == "user" |
| 381 | + assert teachers_mod._flatten_teacher_prompt("system", "") == "system" |
| 382 | + |
| 383 | + def test_require_non_empty_teacher_output_refuses_blank_text(self) -> None: |
| 384 | + with pytest.raises(TeacherInvocationError, match="self returned empty output"): |
| 385 | + teachers_mod._require_non_empty_teacher_output(" ", teacher="self") |
| 386 | + |
| 387 | + def test_extract_openai_message_text_handles_list_content_and_errors(self) -> None: |
| 388 | + response = { |
| 389 | + "choices": [ |
| 390 | + { |
| 391 | + "message": { |
| 392 | + "content": [ |
| 393 | + {"text": " first "}, |
| 394 | + {"text": " second "}, |
| 395 | + ] |
| 396 | + } |
| 397 | + } |
| 398 | + ] |
| 399 | + } |
| 400 | + assert teachers_mod._extract_openai_message_text(response) == "first\nsecond" |
| 401 | + |
| 402 | + with pytest.raises(TeacherInvocationError, match="missing choices"): |
| 403 | + teachers_mod._extract_openai_message_text({}) |
| 404 | + |
| 405 | + with pytest.raises(TeacherInvocationError, match="missing choices\\[0\\]\\.message"): |
| 406 | + teachers_mod._extract_openai_message_text({"choices": [{}]}) |
| 407 | + |
| 408 | + with pytest.raises(TeacherInvocationError, match="missing non-empty message content"): |
| 409 | + teachers_mod._extract_openai_message_text({"choices": [{"message": {"content": None}}]}) |
| 410 | + |
| 411 | + def test_extract_anthropic_text_handles_errors(self) -> None: |
| 412 | + with pytest.raises(TeacherInvocationError, match="missing content blocks"): |
| 413 | + teachers_mod._extract_anthropic_text({}) |
| 414 | + |
| 415 | + with pytest.raises(TeacherInvocationError, match="missing non-empty text blocks"): |
| 416 | + teachers_mod._extract_anthropic_text( |
| 417 | + {"content": [{"type": "image", "text": "ignored"}, {"type": "text", "text": " "}]} |
| 418 | + ) |
| 419 | + |
| 420 | + def test_normalize_chat_content_and_obj_get_helpers(self) -> None: |
| 421 | + assert teachers_mod._normalize_chat_content(" hello ") == "hello" |
| 422 | + assert ( |
| 423 | + teachers_mod._normalize_chat_content([{"text": " one "}, {"text": " two "}]) |
| 424 | + == "one\ntwo" |
| 425 | + ) |
| 426 | + assert teachers_mod._normalize_chat_content([{"text": " "}]) is None |
| 427 | + assert teachers_mod._normalize_chat_content(123) is None |
| 428 | + assert teachers_mod._obj_get({"name": "value"}, "name") == "value" |
| 429 | + assert teachers_mod._obj_get(SimpleNamespace(name="value"), "name") == "value" |
| 430 | + |
| 431 | + def test_openai_compat_url_helpers_normalize_suffixes(self) -> None: |
| 432 | + assert ( |
| 433 | + teachers_mod._normalize_openai_compat_base_url( |
| 434 | + "http://127.0.0.1:8000/v1/chat/completions" |
| 435 | + ) |
| 436 | + == "http://127.0.0.1:8000" |
| 437 | + ) |
| 438 | + assert ( |
| 439 | + teachers_mod._normalize_openai_compat_base_url("http://127.0.0.1:8000/chat/completions") |
| 440 | + == "http://127.0.0.1:8000" |
| 441 | + ) |
| 442 | + assert teachers_mod._openai_compat_models_url("http://127.0.0.1:8000/v1") == ( |
| 443 | + "http://127.0.0.1:8000/v1/models" |
| 444 | + ) |
| 445 | + assert teachers_mod._openai_compat_models_url("http://127.0.0.1:8000") == ( |
| 446 | + "http://127.0.0.1:8000/v1/models" |
| 447 | + ) |
| 448 | + assert teachers_mod._openai_compat_chat_url("http://127.0.0.1:8000/v1") == ( |
| 449 | + "http://127.0.0.1:8000/v1/chat/completions" |
| 450 | + ) |
| 451 | + assert teachers_mod._openai_compat_chat_url("http://127.0.0.1:8000") == ( |
| 452 | + "http://127.0.0.1:8000/v1/chat/completions" |
| 453 | + ) |
| 454 | + |
| 455 | + |
| 456 | +class TestTeacherRuntimeHelpers: |
| 457 | + def test_resolve_generation_device_prefers_requested_or_detected_backends( |
| 458 | + self, |
| 459 | + monkeypatch: pytest.MonkeyPatch, |
| 460 | + ) -> None: |
| 461 | + assert teachers_mod._resolve_generation_device("mps") == "mps" |
| 462 | + |
| 463 | + monkeypatch.delitem(sys.modules, "torch", raising=False) |
| 464 | + real_import = builtins.__import__ |
| 465 | + |
| 466 | + def _missing_torch( |
| 467 | + name: str, |
| 468 | + globals: dict[str, object] | None = None, |
| 469 | + locals: dict[str, object] | None = None, |
| 470 | + fromlist: tuple[str, ...] = (), |
| 471 | + level: int = 0, |
| 472 | + ) -> object: |
| 473 | + if name == "torch": |
| 474 | + raise ImportError("no torch") |
| 475 | + return real_import(name, globals, locals, fromlist, level) |
| 476 | + |
| 477 | + monkeypatch.setattr(builtins, "__import__", _missing_torch) |
| 478 | + assert teachers_mod._resolve_generation_device("auto") == "cpu" |
| 479 | + |
| 480 | + monkeypatch.setattr(builtins, "__import__", real_import) |
| 481 | + monkeypatch.setitem( |
| 482 | + sys.modules, |
| 483 | + "torch", |
| 484 | + SimpleNamespace( |
| 485 | + cuda=SimpleNamespace(is_available=lambda: True), |
| 486 | + backends=SimpleNamespace(mps=SimpleNamespace(is_available=lambda: False)), |
| 487 | + ), |
| 488 | + ) |
| 489 | + assert teachers_mod._resolve_generation_device("auto") == "cuda" |
| 490 | + |
| 491 | + monkeypatch.setitem( |
| 492 | + sys.modules, |
| 493 | + "torch", |
| 494 | + SimpleNamespace( |
| 495 | + cuda=SimpleNamespace(is_available=lambda: False), |
| 496 | + backends=SimpleNamespace(mps=SimpleNamespace(is_available=lambda: True)), |
| 497 | + ), |
| 498 | + ) |
| 499 | + assert teachers_mod._resolve_generation_device("auto") == "mps" |
| 500 | + |
| 501 | + monkeypatch.setitem( |
| 502 | + sys.modules, |
| 503 | + "torch", |
| 504 | + SimpleNamespace( |
| 505 | + cuda=SimpleNamespace(is_available=lambda: False), |
| 506 | + backends=SimpleNamespace(mps=SimpleNamespace(is_available=lambda: False)), |
| 507 | + ), |
| 508 | + ) |
| 509 | + assert teachers_mod._resolve_generation_device("auto") == "cpu" |
| 510 | + |
| 511 | + def test_default_openai_client_validates_import_surface( |
| 512 | + self, |
| 513 | + monkeypatch: pytest.MonkeyPatch, |
| 514 | + ) -> None: |
| 515 | + def _raise_import(name: str) -> object: |
| 516 | + raise ImportError(name) |
| 517 | + |
| 518 | + monkeypatch.setattr("dlm.synth.teachers.importlib.import_module", _raise_import) |
| 519 | + with pytest.raises(TeacherUnavailableError, match="requires the openai package"): |
| 520 | + teachers_mod._default_openai_client("secret") |
| 521 | + |
| 522 | + monkeypatch.setattr( |
| 523 | + "dlm.synth.teachers.importlib.import_module", lambda _name: SimpleNamespace() |
| 524 | + ) |
| 525 | + with pytest.raises(TeacherUnavailableError, match="does not expose OpenAI client"): |
| 526 | + teachers_mod._default_openai_client("secret") |
| 527 | + |
| 528 | + captured: list[str] = [] |
| 529 | + |
| 530 | + class _OpenAI: |
| 531 | + def __init__(self, *, api_key: str) -> None: |
| 532 | + captured.append(api_key) |
| 533 | + |
| 534 | + monkeypatch.setattr( |
| 535 | + "dlm.synth.teachers.importlib.import_module", |
| 536 | + lambda _name: SimpleNamespace(OpenAI=_OpenAI), |
| 537 | + ) |
| 538 | + client = teachers_mod._default_openai_client("secret") |
| 539 | + assert isinstance(client, _OpenAI) |
| 540 | + assert captured == ["secret"] |
| 541 | + |
| 542 | + def test_default_anthropic_client_validates_import_surface( |
| 543 | + self, |
| 544 | + monkeypatch: pytest.MonkeyPatch, |
| 545 | + ) -> None: |
| 546 | + def _raise_import(name: str) -> object: |
| 547 | + raise ImportError(name) |
| 548 | + |
| 549 | + monkeypatch.setattr("dlm.synth.teachers.importlib.import_module", _raise_import) |
| 550 | + with pytest.raises(TeacherUnavailableError, match="requires the anthropic package"): |
| 551 | + teachers_mod._default_anthropic_client("secret") |
| 552 | + |
| 553 | + monkeypatch.setattr( |
| 554 | + "dlm.synth.teachers.importlib.import_module", lambda _name: SimpleNamespace() |
| 555 | + ) |
| 556 | + with pytest.raises(TeacherUnavailableError, match="does not expose Anthropic client"): |
| 557 | + teachers_mod._default_anthropic_client("secret") |
| 558 | + |
| 559 | + captured: list[str] = [] |
| 560 | + |
| 561 | + class _Anthropic: |
| 562 | + def __init__(self, *, api_key: str) -> None: |
| 563 | + captured.append(api_key) |
| 564 | + |
| 565 | + monkeypatch.setattr( |
| 566 | + "dlm.synth.teachers.importlib.import_module", |
| 567 | + lambda _name: SimpleNamespace(Anthropic=_Anthropic), |
| 568 | + ) |
| 569 | + client = teachers_mod._default_anthropic_client("secret") |
| 570 | + assert isinstance(client, _Anthropic) |
| 571 | + assert captured == ["secret"] |
| 572 | + |
| 573 | + def test_fetch_openai_compat_model_id_handles_success_empty_and_errors( |
| 574 | + self, |
| 575 | + monkeypatch: pytest.MonkeyPatch, |
| 576 | + ) -> None: |
| 577 | + class _Response: |
| 578 | + def __init__(self, payload: object) -> None: |
| 579 | + self._payload = payload |
| 580 | + |
| 581 | + def __enter__(self) -> _Response: |
| 582 | + return self |
| 583 | + |
| 584 | + def __exit__(self, *_args: object) -> Literal[False]: |
| 585 | + return False |
| 586 | + |
| 587 | + def read(self) -> bytes: |
| 588 | + return json.dumps(self._payload).encode("utf-8") |
| 589 | + |
| 590 | + monkeypatch.setattr( |
| 591 | + "dlm.synth.teachers.urllib.request.urlopen", |
| 592 | + lambda *_args, **_kwargs: _Response({"data": [{"id": "demo-model"}]}), |
| 593 | + ) |
| 594 | + assert ( |
| 595 | + teachers_mod._fetch_openai_compat_model_id( |
| 596 | + "http://127.0.0.1:8000", |
| 597 | + request_timeout=1.0, |
| 598 | + ) |
| 599 | + == "demo-model" |
| 600 | + ) |
| 601 | + |
| 602 | + monkeypatch.setattr( |
| 603 | + "dlm.synth.teachers.urllib.request.urlopen", |
| 604 | + lambda *_args, **_kwargs: _Response({"data": []}), |
| 605 | + ) |
| 606 | + assert ( |
| 607 | + teachers_mod._fetch_openai_compat_model_id( |
| 608 | + "http://127.0.0.1:8000", |
| 609 | + request_timeout=1.0, |
| 610 | + ) |
| 611 | + is None |
| 612 | + ) |
| 613 | + |
| 614 | + monkeypatch.setattr( |
| 615 | + "dlm.synth.teachers.urllib.request.urlopen", |
| 616 | + lambda *_args, **_kwargs: _Response({"data": [{"id": " "}]}), |
| 617 | + ) |
| 618 | + assert ( |
| 619 | + teachers_mod._fetch_openai_compat_model_id( |
| 620 | + "http://127.0.0.1:8000", |
| 621 | + request_timeout=1.0, |
| 622 | + ) |
| 623 | + is None |
| 624 | + ) |
| 625 | + |
| 626 | + def _raise_url_error(*_args: object, **_kwargs: object) -> object: |
| 627 | + raise urllib.error.URLError("boom") |
| 628 | + |
| 629 | + monkeypatch.setattr("dlm.synth.teachers.urllib.request.urlopen", _raise_url_error) |
| 630 | + with pytest.raises(TeacherUnavailableError, match="could not query models"): |
| 631 | + teachers_mod._fetch_openai_compat_model_id( |
| 632 | + "http://127.0.0.1:8000", |
| 633 | + request_timeout=1.0, |
| 634 | + ) |
| 635 | + |
| 636 | + def test_request_openai_compat_completion_handles_success_and_failures( |
| 637 | + self, |
| 638 | + monkeypatch: pytest.MonkeyPatch, |
| 639 | + ) -> None: |
| 640 | + class _Response: |
| 641 | + def __init__(self, payload: object) -> None: |
| 642 | + self._payload = payload |
| 643 | + |
| 644 | + def __enter__(self) -> _Response: |
| 645 | + return self |
| 646 | + |
| 647 | + def __exit__(self, *_args: object) -> Literal[False]: |
| 648 | + return False |
| 649 | + |
| 650 | + def read(self) -> bytes: |
| 651 | + return json.dumps(self._payload).encode("utf-8") |
| 652 | + |
| 653 | + monkeypatch.setattr( |
| 654 | + "dlm.synth.teachers.urllib.request.urlopen", |
| 655 | + lambda *_args, **_kwargs: _Response( |
| 656 | + {"choices": [{"message": {"content": [{"text": " served "}]}}]} |
| 657 | + ), |
| 658 | + ) |
| 659 | + assert ( |
| 660 | + teachers_mod._request_openai_compat_completion( |
| 661 | + "http://127.0.0.1:8000", |
| 662 | + model_id="demo-model", |
| 663 | + messages=[{"role": "user", "content": "hello"}], |
| 664 | + max_new_tokens=11, |
| 665 | + temperature=0.2, |
| 666 | + top_p=0.8, |
| 667 | + seed=5, |
| 668 | + request_timeout=1.0, |
| 669 | + ) |
| 670 | + == "served" |
| 671 | + ) |
| 672 | + |
| 673 | + monkeypatch.setattr( |
| 674 | + "dlm.synth.teachers.urllib.request.urlopen", |
| 675 | + lambda *_args, **_kwargs: _Response({"choices": []}), |
| 676 | + ) |
| 677 | + with pytest.raises(TeacherInvocationError, match="response missing choices"): |
| 678 | + teachers_mod._request_openai_compat_completion( |
| 679 | + "http://127.0.0.1:8000", |
| 680 | + model_id=None, |
| 681 | + messages=[{"role": "user", "content": "hello"}], |
| 682 | + max_new_tokens=11, |
| 683 | + temperature=0.2, |
| 684 | + top_p=None, |
| 685 | + seed=None, |
| 686 | + request_timeout=1.0, |
| 687 | + ) |
| 688 | + |
| 689 | + monkeypatch.setattr( |
| 690 | + "dlm.synth.teachers.urllib.request.urlopen", |
| 691 | + lambda *_args, **_kwargs: _Response({"choices": [{}]}), |
| 692 | + ) |
| 693 | + with pytest.raises( |
| 694 | + TeacherInvocationError, match="response missing choices\\[0\\]\\.message" |
| 695 | + ): |
| 696 | + teachers_mod._request_openai_compat_completion( |
| 697 | + "http://127.0.0.1:8000", |
| 698 | + model_id=None, |
| 699 | + messages=[{"role": "user", "content": "hello"}], |
| 700 | + max_new_tokens=11, |
| 701 | + temperature=0.2, |
| 702 | + top_p=None, |
| 703 | + seed=None, |
| 704 | + request_timeout=1.0, |
| 705 | + ) |
| 706 | + |
| 707 | + monkeypatch.setattr( |
| 708 | + "dlm.synth.teachers.urllib.request.urlopen", |
| 709 | + lambda *_args, **_kwargs: _Response( |
| 710 | + {"choices": [{"message": {"content": [{"text": " "}]}}]} |
| 711 | + ), |
| 712 | + ) |
| 713 | + with pytest.raises(TeacherInvocationError, match="missing non-empty message content"): |
| 714 | + teachers_mod._request_openai_compat_completion( |
| 715 | + "http://127.0.0.1:8000", |
| 716 | + model_id=None, |
| 717 | + messages=[{"role": "user", "content": "hello"}], |
| 718 | + max_new_tokens=11, |
| 719 | + temperature=0.2, |
| 720 | + top_p=None, |
| 721 | + seed=None, |
| 722 | + request_timeout=1.0, |
| 723 | + ) |
| 724 | + |
| 725 | + def _raise_url_error(*_args: object, **_kwargs: object) -> object: |
| 726 | + raise urllib.error.URLError("boom") |
| 727 | + |
| 728 | + monkeypatch.setattr("dlm.synth.teachers.urllib.request.urlopen", _raise_url_error) |
| 729 | + with pytest.raises(TeacherInvocationError, match="request to http://127.0.0.1:8000 failed"): |
| 730 | + teachers_mod._request_openai_compat_completion( |
| 731 | + "http://127.0.0.1:8000", |
| 732 | + model_id=None, |
| 733 | + messages=[{"role": "user", "content": "hello"}], |
| 734 | + max_new_tokens=11, |
| 735 | + temperature=0.2, |
| 736 | + top_p=None, |
| 737 | + seed=None, |
| 738 | + request_timeout=1.0, |
| 739 | + ) |
| 740 | + |
| 741 | + |
| 742 | +def _install_self_loader_modules( |
| 743 | + monkeypatch: pytest.MonkeyPatch, |
| 744 | + *, |
| 745 | + manifest_exists: bool = True, |
| 746 | + license_acceptance: object | None = "accepted", |
| 747 | + load_manifest_error: str | None = None, |
| 748 | + resolve_error: str | None = None, |
| 749 | + select_error: str | None = None, |
| 750 | + backend_load_error: str | None = None, |
| 751 | +) -> dict[str, object]: |
| 752 | + calls: dict[str, object] = {} |
| 753 | + spec = object() |
| 754 | + caps = object() |
| 755 | + parsed = SimpleNamespace( |
| 756 | + frontmatter=SimpleNamespace( |
| 757 | + dlm_id="01KPQ9X1000000000000000000", |
| 758 | + base_model="smollm2-135m", |
| 759 | + ) |
| 760 | + ) |
| 761 | + manifest = SimpleNamespace(exists=lambda: manifest_exists) |
| 762 | + store = SimpleNamespace(manifest=manifest) |
| 763 | + |
| 764 | + class GatedModelError(Exception): |
| 765 | + pass |
| 766 | + |
| 767 | + class AdapterNotFoundError(Exception): |
| 768 | + pass |
| 769 | + |
| 770 | + class UnsupportedBackendError(Exception): |
| 771 | + pass |
| 772 | + |
| 773 | + class ManifestCorruptError(Exception): |
| 774 | + pass |
| 775 | + |
| 776 | + class _Backend: |
| 777 | + def load(self, spec_arg: object, store_arg: object) -> None: |
| 778 | + calls["load"] = (spec_arg, store_arg) |
| 779 | + if backend_load_error is not None: |
| 780 | + raise AdapterNotFoundError(backend_load_error) |
| 781 | + |
| 782 | + backend = _Backend() |
| 783 | + |
| 784 | + def _resolve(base_model: str, *, accept_license: bool) -> object: |
| 785 | + calls["resolve"] = (base_model, accept_license) |
| 786 | + if resolve_error is not None: |
| 787 | + raise GatedModelError(resolve_error) |
| 788 | + return spec |
| 789 | + |
| 790 | + def _load_manifest(_path: object) -> object: |
| 791 | + calls["load_manifest"] = True |
| 792 | + if load_manifest_error is not None: |
| 793 | + raise ManifestCorruptError(load_manifest_error) |
| 794 | + return SimpleNamespace(license_acceptance=license_acceptance) |
| 795 | + |
| 796 | + def _select_backend(backend_name: str, capabilities: object) -> str: |
| 797 | + calls["select_backend"] = (backend_name, capabilities) |
| 798 | + if select_error is not None: |
| 799 | + raise UnsupportedBackendError(select_error) |
| 800 | + return "stub-backend" |
| 801 | + |
| 802 | + def _build_backend(name: str, capabilities: object) -> object: |
| 803 | + calls["build_backend"] = (name, capabilities) |
| 804 | + return backend |
| 805 | + |
| 806 | + monkeypatch.setitem( |
| 807 | + sys.modules, "dlm.base_models", _module("dlm.base_models", resolve=_resolve) |
| 808 | + ) |
| 809 | + monkeypatch.setitem( |
| 810 | + sys.modules, |
| 811 | + "dlm.base_models.errors", |
| 812 | + _module("dlm.base_models.errors", GatedModelError=GatedModelError), |
| 813 | + ) |
| 814 | + monkeypatch.setitem( |
| 815 | + sys.modules, |
| 816 | + "dlm.doc.parser", |
| 817 | + _module("dlm.doc.parser", parse_file=lambda _path: parsed), |
| 818 | + ) |
| 819 | + monkeypatch.setitem( |
| 820 | + sys.modules, |
| 821 | + "dlm.hardware", |
| 822 | + _module("dlm.hardware", doctor=lambda: SimpleNamespace(capabilities=caps)), |
| 823 | + ) |
| 824 | + monkeypatch.setitem( |
| 825 | + sys.modules, |
| 826 | + "dlm.inference", |
| 827 | + _module("dlm.inference", AdapterNotFoundError=AdapterNotFoundError), |
| 828 | + ) |
| 829 | + monkeypatch.setitem( |
| 830 | + sys.modules, |
| 831 | + "dlm.inference.backends", |
| 832 | + _module( |
| 833 | + "dlm.inference.backends", build_backend=_build_backend, select_backend=_select_backend |
| 834 | + ), |
| 835 | + ) |
| 836 | + monkeypatch.setitem( |
| 837 | + sys.modules, |
| 838 | + "dlm.inference.backends.select", |
| 839 | + _module("dlm.inference.backends.select", UnsupportedBackendError=UnsupportedBackendError), |
| 840 | + ) |
| 841 | + monkeypatch.setitem( |
| 842 | + sys.modules, |
| 843 | + "dlm.store.errors", |
| 844 | + _module("dlm.store.errors", ManifestCorruptError=ManifestCorruptError), |
| 845 | + ) |
| 846 | + monkeypatch.setitem( |
| 847 | + sys.modules, |
| 848 | + "dlm.store.manifest", |
| 849 | + _module("dlm.store.manifest", load_manifest=_load_manifest), |
| 850 | + ) |
| 851 | + monkeypatch.setitem( |
| 852 | + sys.modules, |
| 853 | + "dlm.store.paths", |
| 854 | + _module("dlm.store.paths", for_dlm=lambda _dlm_id: store), |
| 855 | + ) |
| 856 | + |
| 857 | + calls["caps"] = caps |
| 858 | + calls["store"] = store |
| 859 | + calls["spec"] = spec |
| 860 | + calls["errors"] = { |
| 861 | + "gated": GatedModelError, |
| 862 | + "adapter": AdapterNotFoundError, |
| 863 | + "unsupported": UnsupportedBackendError, |
| 864 | + "manifest": ManifestCorruptError, |
| 865 | + } |
| 866 | + return calls |
| 867 | + |
| 868 | + |
| 869 | +class TestTeacherLoaderHelpers: |
| 870 | + def test_load_self_backend_wraps_import_error(self, monkeypatch: pytest.MonkeyPatch) -> None: |
| 871 | + real_import = builtins.__import__ |
| 872 | + |
| 873 | + def _raise_on_base_models( |
| 874 | + name: str, |
| 875 | + globals: dict[str, object] | None = None, |
| 876 | + locals: dict[str, object] | None = None, |
| 877 | + fromlist: tuple[str, ...] = (), |
| 878 | + level: int = 0, |
| 879 | + ) -> object: |
| 880 | + if name.startswith("dlm.base_models"): |
| 881 | + raise ImportError("boom") |
| 882 | + return real_import(name, globals, locals, fromlist, level) |
| 883 | + |
| 884 | + monkeypatch.setattr(builtins, "__import__", _raise_on_base_models) |
| 885 | + with pytest.raises(TeacherUnavailableError, match="requires the local inference stack"): |
| 886 | + teachers_mod._load_self_backend(Path("/tmp/doc.dlm"), "auto") |
| 887 | + |
| 888 | + def test_load_self_backend_uses_recorded_license_acceptance( |
| 889 | + self, |
| 890 | + monkeypatch: pytest.MonkeyPatch, |
| 891 | + ) -> None: |
| 892 | + calls = _install_self_loader_modules(monkeypatch, license_acceptance="accepted") |
| 893 | + |
| 894 | + backend = teachers_mod._load_self_backend(Path("/tmp/doc.dlm"), "auto") |
| 895 | + |
| 896 | + assert backend is not None |
| 897 | + assert calls["resolve"] == ("smollm2-135m", True) |
| 898 | + assert calls["select_backend"] == ("auto", calls["caps"]) |
| 899 | + assert calls["build_backend"] == ("stub-backend", calls["caps"]) |
| 900 | + assert calls["load"] == (calls["spec"], calls["store"]) |
| 901 | + |
| 902 | + def test_load_self_backend_tolerates_manifest_read_failure( |
| 903 | + self, |
| 904 | + monkeypatch: pytest.MonkeyPatch, |
| 905 | + ) -> None: |
| 906 | + calls = _install_self_loader_modules( |
| 907 | + monkeypatch, |
| 908 | + load_manifest_error="bad manifest", |
| 909 | + ) |
| 910 | + |
| 911 | + teachers_mod._load_self_backend(Path("/tmp/doc.dlm"), "auto") |
| 912 | + |
| 913 | + assert calls["resolve"] == ("smollm2-135m", False) |
| 914 | + |
| 915 | + def test_load_self_backend_wraps_gated_backend_and_adapter_failures( |
| 916 | + self, |
| 917 | + monkeypatch: pytest.MonkeyPatch, |
| 918 | + ) -> None: |
| 919 | + _install_self_loader_modules(monkeypatch, resolve_error="gated") |
| 920 | + with pytest.raises(TeacherUnavailableError, match="cannot resolve gated base"): |
| 921 | + teachers_mod._load_self_backend(Path("/tmp/doc.dlm"), "auto") |
| 922 | + |
| 923 | + _install_self_loader_modules(monkeypatch, select_error="unsupported backend") |
| 924 | + with pytest.raises(TeacherUnavailableError, match="unsupported backend"): |
| 925 | + teachers_mod._load_self_backend(Path("/tmp/doc.dlm"), "auto") |
| 926 | + |
| 927 | + _install_self_loader_modules(monkeypatch, backend_load_error="missing adapter") |
| 928 | + with pytest.raises(TeacherUnavailableError, match="requires a trained adapter"): |
| 929 | + teachers_mod._load_self_backend(Path("/tmp/doc.dlm"), "auto") |
| 930 | + |
| 931 | + def test_default_hf_loader_wraps_import_error(self, monkeypatch: pytest.MonkeyPatch) -> None: |
| 932 | + real_import = builtins.__import__ |
| 933 | + |
| 934 | + def _raise_transformers( |
| 935 | + name: str, |
| 936 | + globals: dict[str, object] | None = None, |
| 937 | + locals: dict[str, object] | None = None, |
| 938 | + fromlist: tuple[str, ...] = (), |
| 939 | + level: int = 0, |
| 940 | + ) -> object: |
| 941 | + if name == "transformers": |
| 942 | + raise ImportError("boom") |
| 943 | + return real_import(name, globals, locals, fromlist, level) |
| 944 | + |
| 945 | + monkeypatch.setattr(builtins, "__import__", _raise_transformers) |
| 946 | + with pytest.raises(TeacherUnavailableError, match="requires transformers"): |
| 947 | + teachers_mod._default_hf_loader("hf/model", "cpu") |
| 948 | + |
| 949 | + def test_default_hf_loader_moves_model_and_sets_eval( |
| 950 | + self, |
| 951 | + monkeypatch: pytest.MonkeyPatch, |
| 952 | + ) -> None: |
| 953 | + seen: dict[str, object] = {} |
| 954 | + |
| 955 | + class _Model: |
| 956 | + def to(self, device: str) -> _Model: |
| 957 | + seen["device"] = device |
| 958 | + return self |
| 959 | + |
| 960 | + def eval(self) -> None: |
| 961 | + seen["eval"] = True |
| 962 | + |
| 963 | + model = _Model() |
| 964 | + |
| 965 | + class AutoModelForCausalLM: |
| 966 | + @staticmethod |
| 967 | + def from_pretrained(hf_id: str) -> _Model: |
| 968 | + seen["model_id"] = hf_id |
| 969 | + return model |
| 970 | + |
| 971 | + class AutoTokenizer: |
| 972 | + @staticmethod |
| 973 | + def from_pretrained(hf_id: str) -> str: |
| 974 | + seen["tokenizer_id"] = hf_id |
| 975 | + return "tok" |
| 976 | + |
| 977 | + monkeypatch.setitem( |
| 978 | + sys.modules, |
| 979 | + "transformers", |
| 980 | + _module( |
| 981 | + "transformers", |
| 982 | + AutoModelForCausalLM=AutoModelForCausalLM, |
| 983 | + AutoTokenizer=AutoTokenizer, |
| 984 | + ), |
| 985 | + ) |
| 986 | + |
| 987 | + loaded = teachers_mod._default_hf_loader("hf/model", "cuda") |
| 988 | + |
| 989 | + assert loaded.model is model |
| 990 | + assert loaded.tokenizer == "tok" |
| 991 | + assert loaded.device == "cuda" |
| 992 | + assert seen == { |
| 993 | + "model_id": "hf/model", |
| 994 | + "tokenizer_id": "hf/model", |
| 995 | + "device": "cuda", |
| 996 | + "eval": True, |
| 997 | + } |
| 998 | + |
| 999 | + def test_default_hf_generate_seeds_torch_and_calls_runner( |
| 1000 | + self, |
| 1001 | + monkeypatch: pytest.MonkeyPatch, |
| 1002 | + ) -> None: |
| 1003 | + manual: list[int] = [] |
| 1004 | + manual_all: list[int] = [] |
| 1005 | + calls: dict[str, object] = {} |
| 1006 | + |
| 1007 | + def _generate( |
| 1008 | + model: object, |
| 1009 | + tokenizer: object, |
| 1010 | + prompt: str, |
| 1011 | + *, |
| 1012 | + max_new_tokens: int, |
| 1013 | + temperature: float, |
| 1014 | + top_p: float | None, |
| 1015 | + ) -> str: |
| 1016 | + calls["args"] = (model, tokenizer, prompt, max_new_tokens, temperature, top_p) |
| 1017 | + return "ok" |
| 1018 | + |
| 1019 | + monkeypatch.setitem( |
| 1020 | + sys.modules, |
| 1021 | + "dlm.inference.generate", |
| 1022 | + _module("dlm.inference.generate", generate=_generate), |
| 1023 | + ) |
| 1024 | + monkeypatch.setitem( |
| 1025 | + sys.modules, |
| 1026 | + "torch", |
| 1027 | + SimpleNamespace( |
| 1028 | + manual_seed=lambda seed: manual.append(seed), |
| 1029 | + cuda=SimpleNamespace( |
| 1030 | + is_available=lambda: True, |
| 1031 | + manual_seed_all=lambda seed: manual_all.append(seed), |
| 1032 | + ), |
| 1033 | + ), |
| 1034 | + ) |
| 1035 | + |
| 1036 | + out = teachers_mod._default_hf_generate( |
| 1037 | + "model", |
| 1038 | + "tokenizer", |
| 1039 | + "prompt", |
| 1040 | + max_new_tokens=17, |
| 1041 | + temperature=0.3, |
| 1042 | + top_p=0.8, |
| 1043 | + seed=7, |
| 1044 | + ) |
| 1045 | + |
| 1046 | + assert out == "ok" |
| 1047 | + assert manual == [7] |
| 1048 | + assert manual_all == [7] |
| 1049 | + assert calls["args"] == ("model", "tokenizer", "prompt", 17, 0.3, 0.8) |
| 1050 | + |
| 1051 | + def test_default_hf_generate_tolerates_missing_torch_when_seeding( |
| 1052 | + self, |
| 1053 | + monkeypatch: pytest.MonkeyPatch, |
| 1054 | + ) -> None: |
| 1055 | + real_import = builtins.__import__ |
| 1056 | + |
| 1057 | + def _generate( |
| 1058 | + model: object, |
| 1059 | + tokenizer: object, |
| 1060 | + prompt: str, |
| 1061 | + *, |
| 1062 | + max_new_tokens: int, |
| 1063 | + temperature: float, |
| 1064 | + top_p: float | None, |
| 1065 | + ) -> str: |
| 1066 | + _ = model, tokenizer, prompt, max_new_tokens, temperature, top_p |
| 1067 | + return "ok" |
| 1068 | + |
| 1069 | + def _raise_torch( |
| 1070 | + name: str, |
| 1071 | + globals: dict[str, object] | None = None, |
| 1072 | + locals: dict[str, object] | None = None, |
| 1073 | + fromlist: tuple[str, ...] = (), |
| 1074 | + level: int = 0, |
| 1075 | + ) -> object: |
| 1076 | + if name == "torch": |
| 1077 | + raise ImportError("no torch") |
| 1078 | + return real_import(name, globals, locals, fromlist, level) |
| 1079 | + |
| 1080 | + monkeypatch.setitem( |
| 1081 | + sys.modules, |
| 1082 | + "dlm.inference.generate", |
| 1083 | + _module("dlm.inference.generate", generate=_generate), |
| 1084 | + ) |
| 1085 | + monkeypatch.delitem(sys.modules, "torch", raising=False) |
| 1086 | + monkeypatch.setattr(builtins, "__import__", _raise_torch) |
| 1087 | + |
| 1088 | + out = teachers_mod._default_hf_generate( |
| 1089 | + "model", |
| 1090 | + "tokenizer", |
| 1091 | + "prompt", |
| 1092 | + max_new_tokens=17, |
| 1093 | + temperature=0.3, |
| 1094 | + top_p=0.8, |
| 1095 | + seed=7, |
| 1096 | + ) |
| 1097 | + |
| 1098 | + assert out == "ok" |
