llama.cpp/examples/tests/test_embedding.py

185 lines
6.9 KiB
Python

import os, json, subprocess, hashlib
from pathlib import Path
import numpy as np
import pytest
# ---------------------------------------------------------------------------
# Configuration constants
# ---------------------------------------------------------------------------
EPS = 1e-3
REPO_ROOT = Path(__file__).resolve().parents[2]
EXE = REPO_ROOT / ("build/bin/llama-embedding.exe" if os.name == "nt" else "build/bin/llama-embedding")
DEFAULT_ENV = {**os.environ, "LLAMA_CACHE": os.environ.get("LLAMA_CACHE", "tmp")}
SEED = "42"
# ---------------------------------------------------------------------------
# Model setup helpers
# ---------------------------------------------------------------------------
def get_model_hf_params():
"""Default lightweight embedding model."""
return {
"hf_repo": "ggml-org/embeddinggemma-300M-qat-q4_0-GGUF",
"hf_file": "embeddinggemma-300M-qat-Q4_0.gguf",
}
@pytest.fixture(scope="session")
def embedding_model():
"""Download/cache model once per session."""
exe_path = EXE
if not exe_path.exists():
alt = REPO_ROOT / "build/bin/Release/llama-embedding.exe"
if alt.exists():
exe_path = alt
else:
raise FileNotFoundError(f"llama-embedding binary not found under {REPO_ROOT}/build/bin")
params = get_model_hf_params()
cmd = [
str(exe_path),
"-hfr", params["hf_repo"],
"-hff", params["hf_file"],
"--ctx-size", "16",
"--embd-output-format", "json",
"--no-warmup",
"--threads", "1",
"--seed", SEED,
]
res = subprocess.run(cmd, input="ok", capture_output=True, text=True, env=DEFAULT_ENV)
assert res.returncode == 0, f"model download failed: {res.stderr}"
return params
# ---------------------------------------------------------------------------
# Utility functions
# ---------------------------------------------------------------------------
def run_embedding(text: str, fmt: str = "raw", params=None) -> str:
"""Runs llama-embedding and returns stdout (string)."""
exe_path = EXE
if not exe_path.exists():
raise FileNotFoundError(f"Missing binary: {exe_path}")
params = params or get_model_hf_params()
cmd = [
str(exe_path),
"-hfr", params["hf_repo"],
"-hff", params["hf_file"],
"--ctx-size", "2048",
"--embd-output-format", fmt,
"--threads", "1",
"--seed", SEED,
]
result = subprocess.run(cmd, input=text, capture_output=True, text=True, env=DEFAULT_ENV)
if result.returncode:
raise AssertionError(f"embedding failed ({result.returncode}):\n{result.stderr[:400]}")
out = result.stdout.strip()
assert out, f"empty output for text={text!r}, fmt={fmt}"
return out
def cosine_similarity(a: np.ndarray, b: np.ndarray) -> float:
return float(np.dot(a, b) / (np.linalg.norm(a) * np.linalg.norm(b)))
def embedding_hash(vec: np.ndarray) -> str:
"""Return short deterministic signature for regression tracking."""
return hashlib.sha256(vec[:8].tobytes()).hexdigest()[:16]
# ---------------------------------------------------------------------------
# Tests
# ---------------------------------------------------------------------------
# Register custom mark so pytest doesn't warn about it
pytestmark = pytest.mark.filterwarnings("ignore::pytest.PytestUnknownMarkWarning")
@pytest.mark.slow
@pytest.mark.parametrize("fmt", ["raw", "json"])
@pytest.mark.parametrize("text", ["hello world", "hi 🌎", "line1\nline2\nline3"])
def test_embedding_runs_and_finite(fmt, text, embedding_model):
"""Ensure embeddings run end-to-end and produce finite floats."""
out = run_embedding(text, fmt, embedding_model)
floats = (
np.array(out.split(), float)
if fmt == "raw"
else np.array(json.loads(out)["data"][0]["embedding"], float)
)
assert len(floats) > 100
assert np.all(np.isfinite(floats)), f"non-finite values in {fmt} output"
assert 0.1 < np.linalg.norm(floats) < 10
def test_raw_vs_json_consistency(embedding_model):
"""Compare raw vs JSON embedding output for same text."""
text = "hello world"
raw = np.array(run_embedding(text, "raw", embedding_model).split(), float)
jsn = np.array(json.loads(run_embedding(text, "json", embedding_model))["data"][0]["embedding"], float)
assert raw.shape == jsn.shape
cos = cosine_similarity(raw, jsn)
assert cos > 0.999, f"divergence: cos={cos:.4f}"
assert embedding_hash(raw) == embedding_hash(jsn), "hash mismatch → possible nondeterminism"
def test_empty_input_deterministic(embedding_model):
"""Empty input should yield finite, deterministic vector."""
v1 = np.array(run_embedding("", "raw", embedding_model).split(), float)
v2 = np.array(run_embedding("", "raw", embedding_model).split(), float)
assert np.all(np.isfinite(v1))
cos = cosine_similarity(v1, v2)
assert cos > 0.9999, f"Empty input not deterministic (cos={cos:.5f})"
assert 0.1 < np.linalg.norm(v1) < 10
@pytest.mark.slow
def test_very_long_input_stress(embedding_model):
"""Stress test: large input near context window."""
text = "lorem " * 2000
vec = np.array(run_embedding(text, "raw", embedding_model).split(), float)
assert len(vec) > 100
assert np.isfinite(np.linalg.norm(vec))
@pytest.mark.parametrize(
"text",
[" ", "\n\n\n", "123 456 789"],
)
def test_low_information_inputs_stable(text, embedding_model):
"""Whitespace/numeric inputs should yield stable embeddings."""
v1 = np.array(run_embedding(text, "raw", embedding_model).split(), float)
v2 = np.array(run_embedding(text, "raw", embedding_model).split(), float)
cos = cosine_similarity(v1, v2)
assert cos > 0.999, f"unstable embedding for {text!r}"
@pytest.mark.parametrize("flag", ["--no-such-flag", "--help"])
def test_invalid_or_help_flag(flag):
"""Invalid flags should fail; help should succeed."""
res = subprocess.run([str(EXE), flag], capture_output=True, text=True)
if flag == "--no-such-flag":
assert res.returncode != 0
assert any(k in res.stderr.lower() for k in ("error", "invalid", "unknown"))
else:
assert res.returncode == 0
assert "usage" in (res.stdout.lower() + res.stderr.lower())
@pytest.mark.parametrize("fmt", ["raw", "json"])
@pytest.mark.parametrize("text", ["deterministic test", "deterministic test again"])
def test_repeated_call_consistent(fmt, text, embedding_model):
"""Same input → same hash across repeated runs."""
out1 = run_embedding(text, fmt, embedding_model)
out2 = run_embedding(text, fmt, embedding_model)
if fmt == "json":
v1 = np.array(json.loads(out1)["data"][0]["embedding"], float)
v2 = np.array(json.loads(out2)["data"][0]["embedding"], float)
else:
v1 = np.array(out1.split(), float)
v2 = np.array(out2.split(), float)
assert embedding_hash(v1) == embedding_hash(v2)