@@ -0,0 +1,174 @@ |
| 1 | +"""Integration test: ``cluster_kl`` end-to-end on a real tiny model. |
| 2 | + |
| 3 | +Mirrors ``test_external_perplexity_e2e`` — the sprint file for S16 |
| 4 | +explicitly lists this as a DoD item but the fixture was never shipped. |
| 5 | + |
| 6 | +The test: |
| 7 | + |
| 8 | +1. Builds a small random LoRA on SmolLM2-135M (same template as |
| 9 | + ``test_external_perplexity_e2e``). |
| 10 | +2. Runs ``cluster_kl`` with a 16-prompt two-topic set (animals + |
| 11 | + programming) — split the ft signal across topics so the specificity |
| 12 | + ratio has a chance to be meaningfully non-0.5. |
| 13 | +3. Asserts the probe terminates in a non-ERROR verdict, the specificity |
| 14 | + is finite and in ``[0, 1]``, and when preceded by ``null_adapter`` in |
| 15 | + a suite the z-score field is populated. |
| 16 | + |
| 17 | +Needs the ``[semsim]`` extra at runtime (sentence-transformers + |
| 18 | +scikit-learn). We assume integration runners install those; skip |
| 19 | +gracefully when they don't. |
| 20 | +""" |
| 21 | + |
| 22 | +from __future__ import annotations |
| 23 | + |
| 24 | +import math |
| 25 | +from collections.abc import Iterator |
| 26 | +from pathlib import Path |
| 27 | + |
| 28 | +import pytest |
| 29 | + |
| 30 | +from dlm_sway.backends.hf import HuggingFaceDifferentialBackend |
| 31 | +from dlm_sway.core.model import ModelSpec |
| 32 | +from dlm_sway.core.result import Verdict |
| 33 | +from dlm_sway.probes.base import RunContext, build_probe |
| 34 | +from dlm_sway.suite.runner import run as run_suite |
| 35 | +from dlm_sway.suite.spec import SwaySpec |
| 36 | + |
| 37 | +pytestmark = [pytest.mark.slow, pytest.mark.online] |
| 38 | + |
| 39 | + |
| 40 | +# 16 prompts split 8/8 across two obvious topics. |
| 41 | +_PROMPTS = [ |
| 42 | + # Animals (topic A) |
| 43 | + "The cat chased the mouse around the house.", |
| 44 | + "Dogs wag their tails when they are happy.", |
| 45 | + "Elephants never forget a face they have seen.", |
| 46 | + "Lions hunt in packs called prides.", |
| 47 | + "Horses gallop across open fields.", |
| 48 | + "Sharks have rows of sharp teeth.", |
| 49 | + "Bees pollinate flowers as they gather nectar.", |
| 50 | + "Owls hunt small rodents at night.", |
| 51 | + # Programming (topic B) |
| 52 | + "Write a Python decorator that logs every call.", |
| 53 | + "Implement binary search in Rust.", |
| 54 | + "Debug a segmentation fault in C++ pointer arithmetic.", |
| 55 | + "Explain ownership semantics in Rust.", |
| 56 | + "Refactor this JavaScript callback hell into promises.", |
| 57 | + "Optimize the SQL query by adding an index.", |
| 58 | + "Profile the memory usage of a Go program.", |
| 59 | + "Write unit tests for a REST API endpoint.", |
| 60 | +] |
| 61 | + |
| 62 | + |
| 63 | +def _build_random_lora_adapter(base_dir: Path, out_dir: Path) -> None: |
| 64 | + import torch |
| 65 | + from peft import LoraConfig, get_peft_model |
| 66 | + from transformers import AutoModelForCausalLM, AutoTokenizer |
| 67 | + |
| 68 | + torch.manual_seed(0) |
| 69 | + tokenizer = AutoTokenizer.from_pretrained(str(base_dir)) |
| 70 | + if tokenizer.pad_token_id is None: |
| 71 | + tokenizer.pad_token = tokenizer.eos_token |
| 72 | + base = AutoModelForCausalLM.from_pretrained(str(base_dir), torch_dtype=torch.float32) |
| 73 | + cfg = LoraConfig( |
| 74 | + r=8, |
| 75 | + lora_alpha=16, |
| 76 | + target_modules=["q_proj", "v_proj"], |
| 77 | + lora_dropout=0.0, |
| 78 | + bias="none", |
| 79 | + task_type="CAUSAL_LM", |
| 80 | + ) |
| 81 | + peft_model = get_peft_model(base, cfg) |
| 82 | + with torch.no_grad(): |
| 83 | + for name, param in peft_model.named_parameters(): |
| 84 | + if "lora_B" in name: |
| 85 | + param.copy_(torch.randn_like(param) * 0.05) |
| 86 | + peft_model.save_pretrained(str(out_dir)) |
| 87 | + tokenizer.save_pretrained(str(out_dir)) |
| 88 | + |
| 89 | + |
| 90 | +@pytest.fixture(scope="module") |
| 91 | +def random_adapter(tiny_model_dir: Path, tmp_path_factory: pytest.TempPathFactory) -> Path: |
| 92 | + adapter_dir = tmp_path_factory.mktemp("cluster-kl-random-adapter") |
| 93 | + _build_random_lora_adapter(tiny_model_dir, adapter_dir) |
| 94 | + return adapter_dir |
| 95 | + |
| 96 | + |
| 97 | +@pytest.fixture(scope="module") |
| 98 | +def hf_backend( |
| 99 | + tiny_model_dir: Path, random_adapter: Path |
| 100 | +) -> Iterator[HuggingFaceDifferentialBackend]: |
| 101 | + backend = HuggingFaceDifferentialBackend( |
| 102 | + base_spec=ModelSpec(base=str(tiny_model_dir), kind="hf", dtype="fp32", device="cpu"), |
| 103 | + adapter_path=random_adapter, |
| 104 | + ) |
| 105 | + yield backend |
| 106 | + backend.close() |
| 107 | + |
| 108 | + |
| 109 | +def test_probe_runs_on_real_backend(hf_backend: HuggingFaceDifferentialBackend) -> None: |
| 110 | + pytest.importorskip("sklearn") |
| 111 | + pytest.importorskip("sentence_transformers") |
| 112 | + |
| 113 | + probe, spec = build_probe( |
| 114 | + { |
| 115 | + "name": "ck", |
| 116 | + "kind": "cluster_kl", |
| 117 | + "prompts": _PROMPTS, |
| 118 | + "num_clusters": 2, |
| 119 | + "min_prompts": 16, |
| 120 | + } |
| 121 | + ) |
| 122 | + ctx = RunContext(backend=hf_backend) |
| 123 | + result = probe.run(spec, ctx) |
| 124 | + |
| 125 | + assert result.verdict != Verdict.ERROR, f"probe errored: {result.message}" |
| 126 | + # Under a small random LoRA we don't know the specificity sign; |
| 127 | + # just pin that it's finite and in [0, 1]. |
| 128 | + assert result.raw is not None |
| 129 | + assert math.isfinite(result.raw) |
| 130 | + assert 0.0 <= result.raw <= 1.0 |
| 131 | + assert result.evidence["num_clusters"] == 2 |
| 132 | + assert result.evidence["num_prompts"] == 16 |
| 133 | + per_cluster = result.evidence["per_cluster_mean_kl"] |
| 134 | + assert len(per_cluster) == 2 |
| 135 | + |
| 136 | + |
| 137 | +def test_null_calibration_lights_up_zscore( |
| 138 | + hf_backend: HuggingFaceDifferentialBackend, |
| 139 | +) -> None: |
| 140 | + """null_adapter → cluster_kl produces a z_score end-to-end.""" |
| 141 | + pytest.importorskip("sklearn") |
| 142 | + pytest.importorskip("sentence_transformers") |
| 143 | + |
| 144 | + raw_spec = SwaySpec.model_validate( |
| 145 | + { |
| 146 | + "version": 1, |
| 147 | + "models": { |
| 148 | + "base": {"base": "placeholder"}, |
| 149 | + "ft": {"base": "placeholder", "adapter": "/tmp/placeholder"}, |
| 150 | + }, |
| 151 | + "suite": [ |
| 152 | + {"name": "null", "kind": "null_adapter", "runs": 2, "cache": False}, |
| 153 | + { |
| 154 | + "name": "ck", |
| 155 | + "kind": "cluster_kl", |
| 156 | + "prompts": _PROMPTS, |
| 157 | + "num_clusters": 2, |
| 158 | + "min_prompts": 16, |
| 159 | + "assert_z_gte": -100.0, # permissive — just want z populated |
| 160 | + }, |
| 161 | + ], |
| 162 | + } |
| 163 | + ) |
| 164 | + result = run_suite(raw_spec, hf_backend) |
| 165 | + assert len(result.probes) == 2 |
| 166 | + null_result = result.probes[0] |
| 167 | + ck_result = result.probes[1] |
| 168 | + assert null_result.verdict == Verdict.PASS |
| 169 | + assert ck_result.verdict != Verdict.ERROR |
| 170 | + assert ck_result.z_score is not None, ( |
| 171 | + f"cluster_kl should have z-scored against null baseline; " |
| 172 | + f"evidence={ck_result.evidence}, message={ck_result.message}" |
| 173 | + ) |
| 174 | + assert math.isfinite(ck_result.z_score) |
