build-small-hackathon/hackathon-advisor

1from pathlib import Path2import sys3from types import ModuleType4 5from hackathon_advisor.data import DEFAULT_EMBEDDING_MODEL_FILE, DEFAULT_EMBEDDING_MODEL_REPO6from hackathon_advisor.llama_embedding import (7    DEFAULT_N_CTX,8    LlamaCppEmbedder,9    SubprocessLlamaCppEmbedder,10    create_llama_cpp_embedder,11)12 13 14def test_llama_embedder_uses_q8_defaults_and_configured_context(15    monkeypatch,16    tmp_path: Path,17) -> None:18    model_path = tmp_path / "embedding.gguf"19    model_path.write_bytes(b"gguf")20    captured: dict = {}21 22    hub = ModuleType("huggingface_hub")23 24    def fake_hf_hub_download(repo_id: str, filename: str, repo_type: str) -> str:25        captured["download"] = {26            "repo_id": repo_id,27            "filename": filename,28            "repo_type": repo_type,29        }30        return str(model_path)31 32    hub.hf_hub_download = fake_hf_hub_download33    llama_cpp = ModuleType("llama_cpp")34    llama_cpp.LLAMA_POOLING_TYPE_MEAN = 135 36    class FakeLlama:37        def __init__(self, **kwargs) -> None:38            captured["llama_kwargs"] = kwargs39 40        def embed(self, text: str, normalize: bool) -> list[float]:41            captured["embed"] = {"text": text, "normalize": normalize}42            return [1.0, 0.0]43 44    llama_cpp.Llama = FakeLlama45    monkeypatch.setitem(sys.modules, "huggingface_hub", hub)46    monkeypatch.setitem(sys.modules, "llama_cpp", llama_cpp)47 48    vector = LlamaCppEmbedder().embed("private archive")49 50    assert vector == [1.0, 0.0]51    assert captured["download"] == {52        "repo_id": DEFAULT_EMBEDDING_MODEL_REPO,53        "filename": DEFAULT_EMBEDDING_MODEL_FILE,54        "repo_type": "model",55    }56    assert captured["llama_kwargs"]["n_ctx"] == DEFAULT_N_CTX57    assert captured["llama_kwargs"]["n_batch"] == DEFAULT_N_CTX58    assert captured["llama_kwargs"]["n_ubatch"] == DEFAULT_N_CTX59    assert captured["embed"] == {"text": "private archive", "normalize": True}60 61 62def test_create_llama_embedder_accepts_explicit_batch(monkeypatch) -> None:63    monkeypatch.setenv("ADVISOR_EMBEDDING_BATCH", "256")64 65    embedder = create_llama_cpp_embedder({"dimensions": 768})66 67    assert embedder.n_batch == 25668 69 70def test_create_llama_embedder_can_isolate_native_runtime(monkeypatch) -> None:71    monkeypatch.setenv("ADVISOR_EMBEDDING_SUBPROCESS", "1")72 73    embedder = create_llama_cpp_embedder({"dimensions": 768})74 75    assert isinstance(embedder, SubprocessLlamaCppEmbedder)76    embedder.close()77 78 79def test_create_llama_embedder_isolates_macos_minicpm_runtime(monkeypatch) -> None:80    monkeypatch.delenv("ADVISOR_EMBEDDING_SUBPROCESS", raising=False)81    monkeypatch.setenv("ADVISOR_MODEL_BACKEND", "minicpm-transformers")82    monkeypatch.setattr("hackathon_advisor.llama_embedding.platform.system", lambda: "Darwin")83 84    embedder = create_llama_cpp_embedder({"dimensions": 768})85 86    assert isinstance(embedder, SubprocessLlamaCppEmbedder)87    embedder.close()88 89 90def test_create_llama_embedder_keeps_in_process_when_isolation_disabled(monkeypatch) -> None:91    monkeypatch.setenv("ADVISOR_EMBEDDING_SUBPROCESS", "0")92    monkeypatch.setenv("ADVISOR_MODEL_BACKEND", "minicpm-transformers")93    monkeypatch.setattr("hackathon_advisor.llama_embedding.platform.system", lambda: "Darwin")94 95    embedder = create_llama_cpp_embedder({"dimensions": 768})96 97    assert isinstance(embedder, LlamaCppEmbedder)98