@@ -0,0 +1,109 @@ |
| | 1 | +"""Integration test: multi_turn_coherence_decay end-to-end on a tiny LoRA. |
| | 2 | + |
| | 3 | +Builds a tiny random LoRA on SmolLM2-135M-Instruct (which has a real |
| | 4 | +chat_template) and runs the probe through 4 turns of synthetic |
| | 5 | +dialogue. The intent isn't to assert specific KL values — they |
| | 6 | +depend on the random adapter — but to exercise the full code path |
| | 7 | +on a real backend so a regression in the chat-template wiring, |
| | 8 | +turn-loop, or curve-fit plumbing surfaces in slow CI. |
| | 9 | + |
| | 10 | +Marked ``slow + online``. |
| | 11 | +""" |
| | 12 | + |
| | 13 | +from __future__ import annotations |
| | 14 | + |
| | 15 | +from pathlib import Path |
| | 16 | + |
| | 17 | +import pytest |
| | 18 | + |
| | 19 | +pytestmark = [pytest.mark.slow, pytest.mark.online] |
| | 20 | + |
| | 21 | + |
| | 22 | +def _build_random_lora_adapter(base_dir: Path, out_dir: Path) -> None: |
| | 23 | + """Same shape as the other slow-lane backend tests.""" |
| | 24 | + import torch |
| | 25 | + from peft import LoraConfig, get_peft_model |
| | 26 | + from transformers import AutoModelForCausalLM, AutoTokenizer |
| | 27 | + |
| | 28 | + torch.manual_seed(0) |
| | 29 | + tokenizer = AutoTokenizer.from_pretrained(str(base_dir)) |
| | 30 | + if tokenizer.pad_token_id is None: |
| | 31 | + tokenizer.pad_token = tokenizer.eos_token |
| | 32 | + base = AutoModelForCausalLM.from_pretrained(str(base_dir), torch_dtype=torch.float32) |
| | 33 | + cfg = LoraConfig( |
| | 34 | + r=8, |
| | 35 | + lora_alpha=16, |
| | 36 | + target_modules=["q_proj", "v_proj"], |
| | 37 | + lora_dropout=0.0, |
| | 38 | + bias="none", |
| | 39 | + task_type="CAUSAL_LM", |
| | 40 | + ) |
| | 41 | + peft_model = get_peft_model(base, cfg) |
| | 42 | + with torch.no_grad(): |
| | 43 | + for name, param in peft_model.named_parameters(): |
| | 44 | + if "lora_B" in name: |
| | 45 | + # Tiny perturbation — base != ft, but generations stay sane |
| | 46 | + # enough to thread through 4 dialogue turns without |
| | 47 | + # collapsing to junk. |
| | 48 | + param.copy_(torch.randn_like(param) * 0.02) |
| | 49 | + peft_model.save_pretrained(str(out_dir)) |
| | 50 | + tokenizer.save_pretrained(str(out_dir)) |
| | 51 | + |
| | 52 | + |
| | 53 | +@pytest.fixture(scope="module") |
| | 54 | +def random_adapter(tiny_model_dir: Path, tmp_path_factory: pytest.TempPathFactory) -> Path: |
| | 55 | + out = tmp_path_factory.mktemp("multi-turn-coherence-adapter") |
| | 56 | + _build_random_lora_adapter(tiny_model_dir, out) |
| | 57 | + return out |
| | 58 | + |
| | 59 | + |
| | 60 | +def test_probe_runs_end_to_end_on_real_adapter(tiny_model_dir: Path, random_adapter: Path) -> None: |
| | 61 | + """Smoke: HF backend + chat_template-equipped tokenizer + real |
| | 62 | + multi-turn dialogue produces a finalized result with the |
| | 63 | + documented evidence keys + finite per-turn KLs.""" |
| | 64 | + from dlm_sway.backends.hf import HuggingFaceDifferentialBackend |
| | 65 | + from dlm_sway.core.model import ModelSpec |
| | 66 | + from dlm_sway.core.result import Verdict |
| | 67 | + from dlm_sway.probes.base import RunContext, build_probe |
| | 68 | + |
| | 69 | + backend = HuggingFaceDifferentialBackend( |
| | 70 | + base_spec=ModelSpec(base=str(tiny_model_dir), kind="hf", dtype="fp32", device="cpu"), |
| | 71 | + adapter_path=random_adapter, |
| | 72 | + ) |
| | 73 | + try: |
| | 74 | + probe, spec = build_probe( |
| | 75 | + { |
| | 76 | + "name": "mtc_smoke", |
| | 77 | + "kind": "multi_turn_coherence_decay", |
| | 78 | + "prompts": [ |
| | 79 | + "What's the difference between TCP and UDP?", |
| | 80 | + "Explain how a neural network learns.", |
| | 81 | + ], |
| | 82 | + "max_turns": 3, |
| | 83 | + "max_new_tokens": 32, # keep CPU runtime under control |
| | 84 | + } |
| | 85 | + ) |
| | 86 | + ctx = RunContext(backend=backend, seed=0, top_k=64) |
| | 87 | + result = probe.run(spec, ctx) |
| | 88 | + finally: |
| | 89 | + backend.close() |
| | 90 | + |
| | 91 | + # Shape: any verdict that isn't ERROR is fine. We don't pin |
| | 92 | + # PASS/FAIL because the random adapter's actual decay shape isn't |
| | 93 | + # under our control. |
| | 94 | + assert result.verdict in { |
| | 95 | + Verdict.PASS, |
| | 96 | + Verdict.FAIL, |
| | 97 | + Verdict.WARN, |
| | 98 | + }, result.message |
| | 99 | + assert result.evidence["max_turns"] == 3 |
| | 100 | + assert result.evidence["num_prompts"] == 2 |
| | 101 | + per_turn = result.evidence["per_turn_kls"] |
| | 102 | + assert len(per_turn) == 2 # max_turns - 1 |
| | 103 | + for kl in per_turn: |
| | 104 | + assert isinstance(kl, float) |
| | 105 | + assert kl >= 0.0 # KL is non-negative |
| | 106 | + assert result.evidence["fit_status"] in {"ok", "stable", "non_monotonic", "degenerate"} |
| | 107 | + sparkline = result.evidence["sparkline"] |
| | 108 | + assert isinstance(sparkline, str) |
| | 109 | + assert len(sparkline) == 2 |