| 1 | """Command implementations for the ``sway`` CLI. |
| 2 | |
| 3 | Each function here is wired to a subcommand in :mod:`dlm_sway.cli.app`. |
| 4 | Commands deliberately do as little as possible themselves — the real |
| 5 | work lives in :mod:`dlm_sway.suite`, :mod:`dlm_sway.backends`, and the |
| 6 | probes package. |
| 7 | """ |
| 8 | |
| 9 | from __future__ import annotations |
| 10 | |
| 11 | import json |
| 12 | import sys |
| 13 | from enum import StrEnum |
| 14 | from pathlib import Path |
| 15 | from typing import Annotated, Any |
| 16 | |
| 17 | import typer |
| 18 | from rich.console import Console |
| 19 | |
| 20 | from dlm_sway import __version__ |
| 21 | from dlm_sway.core.errors import SwayError |
| 22 | from dlm_sway.core.result import SuiteResult, SwayScore, Verdict |
| 23 | |
| 24 | |
| 25 | def run_cmd( |
| 26 | spec: Annotated[Path, typer.Argument(help="Path to a sway.yaml spec.")], |
| 27 | json_out: Annotated[ |
| 28 | Path | None, |
| 29 | typer.Option( |
| 30 | "--json", |
| 31 | "-j", |
| 32 | help="Write the JSON report to this path in addition to the terminal render.", |
| 33 | ), |
| 34 | ] = None, |
| 35 | markdown_out: Annotated[ |
| 36 | Path | None, |
| 37 | typer.Option("--markdown", "-m", help="Write a markdown report to this path."), |
| 38 | ] = None, |
| 39 | weights: Annotated[ |
| 40 | str | None, |
| 41 | typer.Option( |
| 42 | "--weights", |
| 43 | help=( |
| 44 | "Override composite-score category weights. Format: " |
| 45 | "'adherence=0.4,attribution=0.3,calibration=0.2,ablation=0.1'. " |
| 46 | "Unspecified categories keep their defaults." |
| 47 | ), |
| 48 | ), |
| 49 | ] = None, |
| 50 | dry_run: Annotated[ |
| 51 | bool, |
| 52 | typer.Option( |
| 53 | "--dry-run", |
| 54 | help=( |
| 55 | "Validate the spec, list the probes that would run with their " |
| 56 | "category, and exit 0 — no backend is built (D6)." |
| 57 | ), |
| 58 | ), |
| 59 | ] = False, |
| 60 | trace: Annotated[ |
| 61 | Path | None, |
| 62 | typer.Option( |
| 63 | "--trace", |
| 64 | help=( |
| 65 | "Write a forward-pass trace (JSONL) to this path — one event " |
| 66 | "per backend scoring call with probe / view / cache-hit info. " |
| 67 | "Useful for perf investigation; zero overhead when unset." |
| 68 | ), |
| 69 | ), |
| 70 | ] = None, |
| 71 | ) -> None: |
| 72 | """Execute a suite and render a terminal report.""" |
| 73 | if dry_run: |
| 74 | _print_dry_run(spec) |
| 75 | return |
| 76 | try: |
| 77 | weights_override = _parse_weights_flag(weights) |
| 78 | result, score_obj = _execute_spec(spec, weights_override=weights_override, trace_path=trace) |
| 79 | except SwayError as exc: |
| 80 | typer.secho(f"error: {exc}", err=True, fg=typer.colors.RED) |
| 81 | raise typer.Exit(code=2) from exc |
| 82 | |
| 83 | from dlm_sway.suite import report |
| 84 | |
| 85 | console = Console() |
| 86 | report.to_terminal(result, score_obj, console=console) |
| 87 | |
| 88 | if json_out is not None: |
| 89 | json_out.write_text(report.to_json(result, score_obj), encoding="utf-8") |
| 90 | console.print(f"\n[dim]wrote JSON → {json_out}[/dim]") |
| 91 | if markdown_out is not None: |
| 92 | markdown_out.write_text(report.to_markdown(result, score_obj), encoding="utf-8") |
| 93 | console.print(f"[dim]wrote markdown → {markdown_out}[/dim]") |
| 94 | |
| 95 | |
| 96 | def _print_dry_run(spec_path: Path) -> None: |
| 97 | """D6: load + validate the spec, print the probe table, exit cleanly. |
| 98 | |
| 99 | No backend construction — useful for fast feedback on spec edits |
| 100 | before paying for a model load. |
| 101 | """ |
| 102 | from rich.table import Table |
| 103 | |
| 104 | from dlm_sway.probes.base import build_probe, registry, validate_all_probes |
| 105 | from dlm_sway.suite.loader import load_spec |
| 106 | |
| 107 | try: |
| 108 | spec = load_spec(spec_path) |
| 109 | validate_all_probes(spec.suite) |
| 110 | except SwayError as exc: |
| 111 | typer.secho(f"error: {exc}", err=True, fg=typer.colors.RED) |
| 112 | raise typer.Exit(code=2) from exc |
| 113 | |
| 114 | console = Console() |
| 115 | console.print(f"[bold]dry-run for {spec_path}[/bold] — {len(spec.suite)} probe(s)") |
| 116 | console.print() |
| 117 | |
| 118 | table = Table(show_header=True, header_style="bold", box=None, padding=(0, 1)) |
| 119 | table.add_column("#", style="dim") |
| 120 | table.add_column("name", style="cyan") |
| 121 | table.add_column("kind") |
| 122 | table.add_column("category", style="dim") |
| 123 | table.add_column("enabled", style="dim") |
| 124 | registered = registry() |
| 125 | for idx, raw in enumerate(spec.suite, start=1): |
| 126 | probe, probe_spec = build_probe(raw) |
| 127 | cls = registered.get(probe.kind) |
| 128 | category = cls.category if cls is not None else "?" |
| 129 | table.add_row( |
| 130 | str(idx), |
| 131 | probe_spec.name, |
| 132 | probe.kind, |
| 133 | category, |
| 134 | "yes" if probe_spec.enabled else "no", |
| 135 | ) |
| 136 | console.print(table) |
| 137 | |
| 138 | |
| 139 | def list_probes_cmd() -> None: |
| 140 | """List every shipped probe kind with its category + one-line summary (D6).""" |
| 141 | import sys |
| 142 | |
| 143 | from rich.table import Table |
| 144 | |
| 145 | # Make sure every probe module has been imported and registered. |
| 146 | import dlm_sway.probes # noqa: F401 |
| 147 | from dlm_sway.probes.base import registry |
| 148 | |
| 149 | table = Table(show_header=True, header_style="bold", box=None, padding=(0, 1)) |
| 150 | table.add_column("kind", style="cyan") |
| 151 | table.add_column("category", style="dim") |
| 152 | table.add_column("summary") |
| 153 | for kind in sorted(registry()): |
| 154 | cls = registry()[kind] |
| 155 | # Prefer the class-level docstring, then fall back to the |
| 156 | # defining module's module-level docstring. Most probe modules |
| 157 | # lead with a solid one-liner at the top; the class body often |
| 158 | # skips a docstring to avoid repeating it. |
| 159 | summary = _first_doc_line(cls.__doc__) |
| 160 | if not summary: |
| 161 | module = sys.modules.get(cls.__module__) |
| 162 | summary = _first_doc_line(getattr(module, "__doc__", None)) |
| 163 | table.add_row(kind, cls.category, summary) |
| 164 | Console().print(table) |
| 165 | |
| 166 | |
| 167 | def _first_doc_line(doc: str | None) -> str: |
| 168 | """Return the first non-empty line of ``doc``, stripped.""" |
| 169 | if not doc: |
| 170 | return "" |
| 171 | for line in doc.splitlines(): |
| 172 | stripped = line.strip() |
| 173 | if stripped: |
| 174 | return stripped |
| 175 | return "" |
| 176 | |
| 177 | |
| 178 | def gate_cmd( |
| 179 | spec: Annotated[Path, typer.Argument(help="Path to a sway.yaml spec.")], |
| 180 | junit_out: Annotated[ |
| 181 | Path | None, typer.Option("--junit", help="Write JUnit XML for CI ingestion.") |
| 182 | ] = None, |
| 183 | coverage_threshold: Annotated[ |
| 184 | float | None, |
| 185 | typer.Option( |
| 186 | "--threshold", |
| 187 | help="Override the spec's coverage_threshold. Exit non-zero below it.", |
| 188 | ), |
| 189 | ] = None, |
| 190 | weights: Annotated[ |
| 191 | str | None, |
| 192 | typer.Option( |
| 193 | "--weights", |
| 194 | help=( |
| 195 | "Override composite-score category weights. Format: " |
| 196 | "'adherence=0.4,attribution=0.3,calibration=0.2,ablation=0.1'. " |
| 197 | "Unspecified categories keep their defaults." |
| 198 | ), |
| 199 | ), |
| 200 | ] = None, |
| 201 | ) -> None: |
| 202 | """Execute a suite and exit non-zero on failure (CI gate).""" |
| 203 | try: |
| 204 | weights_override = _parse_weights_flag(weights) |
| 205 | result, score_obj = _execute_spec(spec, weights_override=weights_override) |
| 206 | except SwayError as exc: |
| 207 | typer.secho(f"error: {exc}", err=True, fg=typer.colors.RED) |
| 208 | raise typer.Exit(code=2) from exc |
| 209 | |
| 210 | from dlm_sway.suite import report |
| 211 | from dlm_sway.suite.loader import load_spec as _load_spec |
| 212 | |
| 213 | console = Console() |
| 214 | report.to_terminal(result, score_obj, console=console) |
| 215 | |
| 216 | if junit_out is not None: |
| 217 | junit_out.write_text(report.to_junit(result, score_obj), encoding="utf-8") |
| 218 | console.print(f"[dim]wrote JUnit → {junit_out}[/dim]") |
| 219 | |
| 220 | threshold = ( |
| 221 | coverage_threshold |
| 222 | if coverage_threshold is not None |
| 223 | else _load_spec(spec).defaults.coverage_threshold |
| 224 | ) |
| 225 | has_failures = any(p.verdict == Verdict.FAIL for p in result.probes) |
| 226 | below_threshold = score_obj.overall < threshold |
| 227 | if has_failures or below_threshold: |
| 228 | console.print( |
| 229 | f"\n[red]gate FAILED[/red] — overall={score_obj.overall:.2f} < {threshold:.2f}" |
| 230 | if below_threshold |
| 231 | else "\n[red]gate FAILED[/red] — at least one probe reported FAIL" |
| 232 | ) |
| 233 | raise typer.Exit(code=1) |
| 234 | console.print(f"\n[green]gate passed[/green] — overall={score_obj.overall:.2f}") |
| 235 | |
| 236 | |
| 237 | def _infer_base_from_adapter_config(adapter_dir: Path) -> str | None: |
| 238 | """Read ``base_model_name_or_path`` from ``adapter_config.json``. |
| 239 | |
| 240 | Returns ``None`` when the file is missing, malformed, or doesn't |
| 241 | expose the field. Used by ``sway check`` to make ``--base`` optional |
| 242 | in the common case where PEFT already wrote the base id on training |
| 243 | (D4). |
| 244 | """ |
| 245 | cfg_path = adapter_dir / "adapter_config.json" |
| 246 | if not cfg_path.exists(): |
| 247 | return None |
| 248 | try: |
| 249 | data = json.loads(cfg_path.read_text(encoding="utf-8")) |
| 250 | except (OSError, json.JSONDecodeError): |
| 251 | return None |
| 252 | base = data.get("base_model_name_or_path") |
| 253 | if isinstance(base, str) and base: |
| 254 | return base |
| 255 | return None |
| 256 | |
| 257 | |
| 258 | def _check_banner(score_obj: SwayScore, result: SuiteResult) -> tuple[str, str]: |
| 259 | """Compute the (text, rich-style) check verdict banner (D12). |
| 260 | |
| 261 | Calibrated on the delta_kl z-score: ≥3σ is green ("above noise"), |
| 262 | ≥1σ is yellow ("marginal"), and below that is red. When no z-score |
| 263 | is available (no null calibration ran), falls back to the raw |
| 264 | score band. |
| 265 | """ |
| 266 | z = next( |
| 267 | (p.z_score for p in result.probes if p.kind == "delta_kl" and p.z_score is not None), |
| 268 | None, |
| 269 | ) |
| 270 | if z is not None: |
| 271 | if z >= 3.0: |
| 272 | return f"✅ adapter is {z:+.2f}σ above noise", "bold green" |
| 273 | if z >= 1.0: |
| 274 | return f"⚠️ adapter is {z:+.2f}σ above noise — marginal", "bold yellow" |
| 275 | return f"❌ adapter is {z:+.2f}σ — indistinguishable from noise", "bold red" |
| 276 | |
| 277 | # Fallback: composite score band. |
| 278 | if score_obj.overall >= 0.6: |
| 279 | return f"✅ adapter scored {score_obj.overall:.2f} — looks healthy", "bold green" |
| 280 | if score_obj.overall >= 0.3: |
| 281 | return f"⚠️ adapter scored {score_obj.overall:.2f} — partial fit", "bold yellow" |
| 282 | return f"❌ adapter scored {score_obj.overall:.2f} — noise band", "bold red" |
| 283 | |
| 284 | |
| 285 | def check_cmd( |
| 286 | adapter: Annotated[Path, typer.Argument(help="Path to a PEFT adapter directory.")], |
| 287 | base: Annotated[ |
| 288 | str | None, |
| 289 | typer.Option( |
| 290 | "--base", |
| 291 | help=( |
| 292 | "HuggingFace base model id or local path. Inferred from " |
| 293 | "the adapter's ``adapter_config.json`` when omitted (D4)." |
| 294 | ), |
| 295 | ), |
| 296 | ] = None, |
| 297 | prompts: Annotated[ |
| 298 | Path | None, |
| 299 | typer.Option( |
| 300 | "--prompts", |
| 301 | help="File with one prompt per line. Defaults to sway's built-in quick set.", |
| 302 | ), |
| 303 | ] = None, |
| 304 | ) -> None: |
| 305 | """<60s smoke test: "is this adapter doing anything at all?". |
| 306 | |
| 307 | Runs A1 DeltaKL + C2 CalibrationDrift on a small prompt set. No |
| 308 | spec file required. |
| 309 | |
| 310 | **Banner semantics (F20 clarification).** The ``+N.NNσ above noise`` |
| 311 | header appears only when ``null_adapter`` actually calibrated this |
| 312 | run — i.e., when the backend implements ``NullCalibratedBackend``. |
| 313 | Without null calibration (non-HF backends like the HTTP API or MLX |
| 314 | inference), the banner falls back to the composite score band |
| 315 | ("healthy", "partial fit", "noise band") and the σ wording is |
| 316 | suppressed to avoid a false-precision claim. |
| 317 | """ |
| 318 | from dlm_sway.backends import build as build_backend |
| 319 | from dlm_sway.core.model import ModelSpec |
| 320 | from dlm_sway.suite import report |
| 321 | from dlm_sway.suite.runner import run as run_suite |
| 322 | from dlm_sway.suite.score import compute as compute_score |
| 323 | from dlm_sway.suite.spec import SuiteDefaults, SuiteModels, SwaySpec |
| 324 | |
| 325 | # D4: try to infer base model from adapter_config.json before |
| 326 | # erroring out on a missing --base. |
| 327 | if base is None: |
| 328 | inferred = _infer_base_from_adapter_config(adapter) |
| 329 | if inferred is None: |
| 330 | typer.secho( |
| 331 | f"error: --base not given and adapter at {adapter} doesn't carry a " |
| 332 | f"base_model_name_or_path in adapter_config.json. Pass --base " |
| 333 | f"explicitly.", |
| 334 | err=True, |
| 335 | fg=typer.colors.RED, |
| 336 | ) |
| 337 | raise typer.Exit(code=2) |
| 338 | base = inferred |
| 339 | typer.secho(f"(inferred base model: {base})", err=True, fg=typer.colors.CYAN) |
| 340 | |
| 341 | quick_prompts = _load_prompts(prompts) if prompts else _BUILTIN_QUICK_PROMPTS |
| 342 | |
| 343 | base_spec = ModelSpec(base=base, kind="hf") |
| 344 | ft_spec = ModelSpec(base=base, kind="hf", adapter=adapter) |
| 345 | spec = SwaySpec( |
| 346 | version=1, |
| 347 | models=SuiteModels(base=base_spec, ft=ft_spec), |
| 348 | defaults=SuiteDefaults(seed=0), |
| 349 | suite=[ |
| 350 | # S25: pre-run training-health check first. SKIPs cleanly |
| 351 | # when the adapter wasn't produced by dlm (no |
| 352 | # training_state.pt); FAILs loudly on severely-undertrained |
| 353 | # adapters with a banner before the rest of the output. |
| 354 | { |
| 355 | "name": "quick_gradient_ghost", |
| 356 | "kind": "gradient_ghost", |
| 357 | "adapter_path": str(adapter), |
| 358 | }, |
| 359 | # Calibrate first so delta_kl can publish a z-score the |
| 360 | # banner reads off. |
| 361 | {"name": "quick_null", "kind": "null_adapter", "runs": 3}, |
| 362 | { |
| 363 | "name": "quick_delta_kl", |
| 364 | "kind": "delta_kl", |
| 365 | "prompts": list(quick_prompts), |
| 366 | "assert_mean_gte": 0.01, |
| 367 | }, |
| 368 | { |
| 369 | "name": "quick_calibration", |
| 370 | "kind": "calibration_drift", |
| 371 | "items_limit": 10, |
| 372 | }, |
| 373 | ], |
| 374 | ) |
| 375 | try: |
| 376 | backend = build_backend(ft_spec) |
| 377 | except SwayError as exc: |
| 378 | typer.secho(f"error: {exc}", err=True, fg=typer.colors.RED) |
| 379 | raise typer.Exit(code=2) from exc |
| 380 | |
| 381 | try: |
| 382 | result = run_suite(spec, backend, spec_path="<check>") |
| 383 | finally: |
| 384 | _close_if_possible(backend) |
| 385 | score_obj = compute_score(result) |
| 386 | |
| 387 | # D12: top-line banner before the full report so a user looking |
| 388 | # only at the first line still gets the verdict. |
| 389 | console = Console() |
| 390 | |
| 391 | # S25 — pre-flight gradient_ghost banner. Fires BEFORE the verdict |
| 392 | # banner so the user sees "this adapter is undertrained" first; |
| 393 | # the rest of the check output stays for context (the user might |
| 394 | # still want to see how badly the other probes scored). |
| 395 | _emit_gradient_ghost_banner(result, console) |
| 396 | |
| 397 | banner_text, banner_style = _check_banner(score_obj, result) |
| 398 | console.print() |
| 399 | console.print(banner_text, style=banner_style) |
| 400 | console.print() |
| 401 | report.to_terminal(result, score_obj, console=console) |
| 402 | |
| 403 | |
| 404 | def _emit_gradient_ghost_banner(result: object, console: Console) -> None: |
| 405 | """Print a yellow/red ⚠️ banner if gradient_ghost FAILed (S25 P6). |
| 406 | |
| 407 | Reaches into ``result.probes`` for any probe with |
| 408 | ``kind=gradient_ghost`` and verdict FAIL. Informational — no |
| 409 | effect on exit code; the user might still want to inspect the |
| 410 | other probes' verdicts. |
| 411 | """ |
| 412 | probes = getattr(result, "probes", ()) or () |
| 413 | for p in probes: |
| 414 | if getattr(p, "kind", "") != "gradient_ghost": |
| 415 | continue |
| 416 | verdict_str = str(getattr(p, "verdict", "")).lower() |
| 417 | if verdict_str == "fail": |
| 418 | console.print() |
| 419 | console.print( |
| 420 | "⚠️ PRE-RUN ALERT — gradient_ghost flagged severe undertraining", |
| 421 | style="bold red", |
| 422 | ) |
| 423 | msg = getattr(p, "message", "") |
| 424 | if msg: |
| 425 | console.print(f" {msg}", style="red") |
| 426 | console.print( |
| 427 | " The probe scores below may be unreliable. Consider retraining.", |
| 428 | style="dim red", |
| 429 | ) |
| 430 | return |
| 431 | if verdict_str == "warn": |
| 432 | console.print() |
| 433 | console.print( |
| 434 | "⚠️ gradient_ghost: training may not have fully converged", |
| 435 | style="bold yellow", |
| 436 | ) |
| 437 | msg = getattr(p, "message", "") |
| 438 | if msg: |
| 439 | console.print(f" {msg}", style="yellow") |
| 440 | return |
| 441 | |
| 442 | |
| 443 | def diff_cmd( |
| 444 | spec: Annotated[Path, typer.Argument(help="Path to a sway.yaml spec.")], |
| 445 | adapter_a: Annotated[Path, typer.Option("--a", help="First adapter path.")], |
| 446 | adapter_b: Annotated[Path, typer.Option("--b", help="Second adapter path.")], |
| 447 | ) -> None: |
| 448 | """Run the same suite against two adapters and show per-probe deltas.""" |
| 449 | from dlm_sway.backends import build as build_backend |
| 450 | from dlm_sway.suite.loader import load_spec |
| 451 | from dlm_sway.suite.runner import run as run_suite |
| 452 | from dlm_sway.suite.score import compute as compute_score |
| 453 | |
| 454 | sway_spec = load_spec(spec) |
| 455 | console = Console() |
| 456 | |
| 457 | def _score_for(adapter_path: Path) -> tuple[float, dict[str, float]]: |
| 458 | ft_spec = sway_spec.models.ft.model_copy(update={"adapter": adapter_path}) |
| 459 | backend = build_backend(ft_spec) |
| 460 | try: |
| 461 | result = run_suite(sway_spec, backend, spec_path=str(spec)) |
| 462 | finally: |
| 463 | _close_if_possible(backend) |
| 464 | scored = compute_score(result) |
| 465 | per_probe = {p.name: (p.score or 0.0) for p in result.probes} |
| 466 | return scored.overall, per_probe |
| 467 | |
| 468 | try: |
| 469 | overall_a, per_a = _score_for(adapter_a) |
| 470 | overall_b, per_b = _score_for(adapter_b) |
| 471 | except SwayError as exc: |
| 472 | typer.secho(f"error: {exc}", err=True, fg=typer.colors.RED) |
| 473 | raise typer.Exit(code=2) from exc |
| 474 | |
| 475 | console.print(f"[bold]overall[/bold] A: {overall_a:.2f} B: {overall_b:.2f}") |
| 476 | console.print() |
| 477 | console.print("[bold]per-probe[/bold] (A → B, Δ):") |
| 478 | regressed_small = 0 # |Δ| > 0.10 in the wrong direction |
| 479 | regressed_large = 0 # |Δ| > 0.20 in the wrong direction |
| 480 | for name in sorted(per_a.keys() | per_b.keys()): |
| 481 | a = per_a.get(name, 0.0) |
| 482 | b = per_b.get(name, 0.0) |
| 483 | delta = b - a |
| 484 | sign = "+" if delta >= 0 else "" |
| 485 | console.print(f" {name:<30} {a:.2f} → {b:.2f} ({sign}{delta:+.2f})") |
| 486 | if delta < -0.10: |
| 487 | regressed_small += 1 |
| 488 | if delta < -0.20: |
| 489 | regressed_large += 1 |
| 490 | |
| 491 | # D13: regression summary line. The audit's example phrasing was |
| 492 | # "A→B: 3 probes regressed >0.10, 1 regressed >0.20, composite Δ=+0.02". |
| 493 | # Color cue tracks the composite delta: green for any improvement, |
| 494 | # red on regression, yellow on flat-with-regressions. |
| 495 | composite_delta = overall_b - overall_a |
| 496 | if composite_delta > 0.0: |
| 497 | summary_style = "bold green" |
| 498 | elif regressed_small or regressed_large: |
| 499 | summary_style = "bold red" if composite_delta < 0.0 else "bold yellow" |
| 500 | else: |
| 501 | summary_style = "dim" |
| 502 | |
| 503 | console.print() |
| 504 | console.print( |
| 505 | f"A→B: {regressed_small} probe(s) regressed >0.10, " |
| 506 | f"{regressed_large} regressed >0.20, " |
| 507 | f"composite Δ={composite_delta:+.2f}", |
| 508 | style=summary_style, |
| 509 | ) |
| 510 | |
| 511 | |
| 512 | def autogen_cmd( |
| 513 | dlm_path: Annotated[Path, typer.Argument(help="Path to a .dlm file.")], |
| 514 | out: Annotated[ |
| 515 | Path, |
| 516 | typer.Option("--out", "-o", help="Where to write the generated sway.yaml."), |
| 517 | ] = Path("sway.yaml"), |
| 518 | ) -> None: |
| 519 | """Generate a sway.yaml from a .dlm file (requires the ``dlm-sway[dlm]`` extra).""" |
| 520 | import importlib |
| 521 | |
| 522 | try: |
| 523 | autogen_mod = importlib.import_module("dlm_sway.integrations.dlm.autogen") |
| 524 | except ImportError as exc: |
| 525 | typer.secho( |
| 526 | "dlm integration not installed — run: pip install 'dlm-sway[dlm]'", |
| 527 | err=True, |
| 528 | fg=typer.colors.RED, |
| 529 | ) |
| 530 | raise typer.Exit(code=2) from exc |
| 531 | |
| 532 | try: |
| 533 | autogen_mod.write_sway_yaml(dlm_path, out) |
| 534 | except SwayError as exc: |
| 535 | typer.secho(f"error: {exc}", err=True, fg=typer.colors.RED) |
| 536 | raise typer.Exit(code=2) from exc |
| 537 | |
| 538 | typer.echo(f"wrote {out}") |
| 539 | |
| 540 | |
| 541 | _DOCTOR_BACKENDS: dict[str, tuple[str, ...]] = { |
| 542 | "hf": ("torch", "transformers", "peft"), |
| 543 | "mlx": ("mlx", "mlx_lm"), |
| 544 | # ``sklearn`` is S16's cluster_kl dep; shipped under [semsim] so it |
| 545 | # rides the same 80 MB MiniLM load adapter_revert already pulls. |
| 546 | "semsim": ("sentence_transformers", "sklearn"), |
| 547 | "style": ("spacy", "textstat", "nlpaug"), |
| 548 | "dlm": ("dlm",), |
| 549 | # ``plotly`` is the load-bearing dep for ``sway report --format html``; |
| 550 | # S12 docs listed it but doctor never probed it before F04. |
| 551 | "viz": ("matplotlib", "plotly"), |
| 552 | # S13 API backend. |
| 553 | "api": ("httpx", "tenacity"), |
| 554 | "pytest": ("pytest",), |
| 555 | } |
| 556 | |
| 557 | |
| 558 | def _doctor_payload() -> dict[str, Any]: |
| 559 | """Build the JSON-friendly doctor payload (used by both render paths).""" |
| 560 | extras: dict[str, dict[str, str | None]] = {} |
| 561 | for extra, modules in _DOCTOR_BACKENDS.items(): |
| 562 | extras[extra] = {mod: _module_version(mod) for mod in modules} |
| 563 | return { |
| 564 | "sway_version": __version__, |
| 565 | "python": sys.version.split()[0], |
| 566 | "platform": sys.platform, |
| 567 | "extras": extras, |
| 568 | } |
| 569 | |
| 570 | |
| 571 | def _module_version(name: str) -> str | None: |
| 572 | """Return the installed module's ``__version__`` string, or ``None``.""" |
| 573 | import importlib |
| 574 | |
| 575 | try: |
| 576 | mod = importlib.import_module(name) |
| 577 | except ImportError: |
| 578 | return None |
| 579 | return str(getattr(mod, "__version__", "installed")) |
| 580 | |
| 581 | |
| 582 | def doctor_cmd( |
| 583 | json_out: Annotated[ |
| 584 | bool, |
| 585 | typer.Option( |
| 586 | "--json", |
| 587 | help=( |
| 588 | "Emit a machine-readable JSON payload instead of the rich " |
| 589 | "terminal layout (D7). CI-grep-friendly." |
| 590 | ), |
| 591 | ), |
| 592 | ] = False, |
| 593 | ) -> None: |
| 594 | """Print backend availability and version info.""" |
| 595 | payload = _doctor_payload() |
| 596 | if json_out: |
| 597 | typer.echo(json.dumps(payload, indent=2, sort_keys=True)) |
| 598 | return |
| 599 | |
| 600 | console = Console() |
| 601 | console.print(f"[bold]sway[/bold] {payload['sway_version']}") |
| 602 | console.print(f" python: {payload['python']}") |
| 603 | console.print(f" platform: {payload['platform']}") |
| 604 | console.print() |
| 605 | console.print("[bold]backends[/bold]") |
| 606 | for extra, modules in payload["extras"].items(): |
| 607 | parts = [] |
| 608 | for mod, ver in modules.items(): |
| 609 | if ver is None: |
| 610 | parts.append(f"[red]{mod}: missing[/red]") |
| 611 | else: |
| 612 | parts.append(f"[green]{mod}: {ver}[/green]") |
| 613 | console.print(f" {extra:<8} {' '.join(parts)}") |
| 614 | |
| 615 | |
| 616 | class ReportFormat(StrEnum): |
| 617 | """Allowed values for ``sway report --format`` (D11). |
| 618 | |
| 619 | Typer enforces the enum at parse time, so unknown formats produce |
| 620 | a clear ``Invalid value`` error instead of silently falling back |
| 621 | to the terminal renderer. |
| 622 | """ |
| 623 | |
| 624 | TERMINAL = "terminal" |
| 625 | MARKDOWN = "md" |
| 626 | MARKDOWN_LONG = "markdown" # alias kept for muscle memory |
| 627 | JUNIT = "junit" |
| 628 | JSON = "json" |
| 629 | HTML = "html" |
| 630 | |
| 631 | |
| 632 | def report_cmd( |
| 633 | result_json: Annotated[Path, typer.Argument(help="Path to a saved result JSON.")], |
| 634 | format: Annotated[ |
| 635 | ReportFormat, |
| 636 | typer.Option( |
| 637 | "--format", |
| 638 | help="Output format: terminal, md (alias: markdown), junit, json, or html.", |
| 639 | ), |
| 640 | ] = ReportFormat.TERMINAL, |
| 641 | out: Annotated[ |
| 642 | Path | None, |
| 643 | typer.Option( |
| 644 | "--out", |
| 645 | "-o", |
| 646 | help=( |
| 647 | "Write the rendered output to this path instead of stdout. " |
| 648 | "Required for --format html (Plotly's inlined JS is ~3 MB)." |
| 649 | ), |
| 650 | ), |
| 651 | ] = None, |
| 652 | ) -> None: |
| 653 | """Re-render a previously saved run (for history tracking / dashboards). |
| 654 | |
| 655 | The CLI deserializes the JSON back into the canonical |
| 656 | ``(SuiteResult, SwayScore)`` pair via :func:`report.from_json`, |
| 657 | then routes through the same renderers as a fresh ``sway run``. |
| 658 | Single source for every format keeps terminal / md / junit / |
| 659 | json / html output identical regardless of where they came from (B16). |
| 660 | """ |
| 661 | from dlm_sway.suite import report |
| 662 | |
| 663 | raw: dict[str, Any] = json.loads(result_json.read_text(encoding="utf-8")) |
| 664 | |
| 665 | if format is ReportFormat.JSON: |
| 666 | # Pass-through: the saved file *is* the canonical JSON. Re-emit |
| 667 | # via to_json against the round-tripped pair so any schema |
| 668 | # additions land consistently. |
| 669 | suite, score = report.from_json(raw) |
| 670 | _emit(report.to_json(suite, score), out) |
| 671 | return |
| 672 | |
| 673 | suite, score = report.from_json(raw) |
| 674 | if format in (ReportFormat.MARKDOWN, ReportFormat.MARKDOWN_LONG): |
| 675 | _emit(report.to_markdown(suite, score), out) |
| 676 | return |
| 677 | if format is ReportFormat.JUNIT: |
| 678 | _emit(report.to_junit(suite, score), out) |
| 679 | return |
| 680 | if format is ReportFormat.HTML: |
| 681 | try: |
| 682 | from dlm_sway.suite import report_html |
| 683 | except ImportError as exc: # pragma: no cover — graceful install hint |
| 684 | typer.echo(f"sway report --format html: {exc}", err=True) |
| 685 | raise typer.Exit(code=2) from exc |
| 686 | try: |
| 687 | html_text = report_html.to_html(suite, score) |
| 688 | except RuntimeError as exc: |
| 689 | typer.echo(f"sway report --format html: {exc}", err=True) |
| 690 | raise typer.Exit(code=2) from exc |
| 691 | if out is None: |
| 692 | # Refuse to dump 3 MB of HTML to stdout by default — the |
| 693 | # user almost always wants a file. |
| 694 | typer.echo( |
| 695 | "sway report --format html requires --out PATH " |
| 696 | "(Plotly JS bundle is ~3 MB; stdout is not an HTML viewer)", |
| 697 | err=True, |
| 698 | ) |
| 699 | raise typer.Exit(code=2) |
| 700 | out.write_text(html_text, encoding="utf-8") |
| 701 | typer.echo(f"wrote HTML → {out}", err=True) |
| 702 | return |
| 703 | # ReportFormat.TERMINAL. |
| 704 | if out is not None: |
| 705 | typer.echo( |
| 706 | "sway report --format terminal does not support --out; " |
| 707 | "use --format md or --format html for file output.", |
| 708 | err=True, |
| 709 | ) |
| 710 | raise typer.Exit(code=2) |
| 711 | report.to_terminal(suite, score, console=Console()) |
| 712 | |
| 713 | |
| 714 | def _emit(text: str, out: Path | None) -> None: |
| 715 | """Either write to the target path or ``typer.echo`` to stdout.""" |
| 716 | if out is None: |
| 717 | typer.echo(text) |
| 718 | else: |
| 719 | out.write_text(text, encoding="utf-8") |
| 720 | typer.echo(f"wrote {out}", err=True) |
| 721 | |
| 722 | |
| 723 | class CompareFormat(StrEnum): |
| 724 | """Allowed values for ``sway compare --format``.""" |
| 725 | |
| 726 | TERMINAL = "terminal" |
| 727 | MARKDOWN = "md" |
| 728 | MARKDOWN_LONG = "markdown" # alias kept for muscle memory |
| 729 | JSON = "json" |
| 730 | |
| 731 | |
| 732 | def compare_cmd( |
| 733 | result_jsons: Annotated[ |
| 734 | list[Path], |
| 735 | typer.Argument(help="Two or more saved result JSONs, in chronological order."), |
| 736 | ], |
| 737 | format: Annotated[ |
| 738 | CompareFormat, |
| 739 | typer.Option( |
| 740 | "--format", |
| 741 | help="Output format: terminal, md (alias: markdown), or json.", |
| 742 | ), |
| 743 | ] = CompareFormat.TERMINAL, |
| 744 | fail_on_regression: Annotated[ |
| 745 | float, |
| 746 | typer.Option( |
| 747 | "--fail-on-regression", |
| 748 | help=( |
| 749 | "Exit non-zero when any probe's score in the newest run dropped " |
| 750 | "by ≥ this threshold vs the previous run. 0 disables the gate." |
| 751 | ), |
| 752 | ), |
| 753 | ] = 0.0, |
| 754 | ) -> None: |
| 755 | """Compare N saved runs side-by-side (regression dashboard). |
| 756 | |
| 757 | Rehydrates each JSON via :func:`report.from_json`, folds the runs |
| 758 | into a :class:`CompareMatrix`, and renders the score table + delta |
| 759 | columns + composite timeline. Intended for CI: point at a history |
| 760 | directory (``sway-history/*.json``) and pipe the output into the |
| 761 | build's log, or set ``--fail-on-regression`` to make the build red |
| 762 | on a real drop. |
| 763 | """ |
| 764 | from dlm_sway.suite import compare, report |
| 765 | |
| 766 | if len(result_jsons) < 2: |
| 767 | typer.echo("sway compare: need at least two result JSONs", err=True) |
| 768 | raise typer.Exit(code=2) |
| 769 | |
| 770 | pairs: list[tuple[SuiteResult, SwayScore]] = [] |
| 771 | labels: list[str] = [] |
| 772 | for path in result_jsons: |
| 773 | try: |
| 774 | raw: dict[str, Any] = json.loads(path.read_text(encoding="utf-8")) |
| 775 | except (OSError, json.JSONDecodeError) as exc: |
| 776 | typer.echo(f"sway compare: cannot read {path}: {exc}", err=True) |
| 777 | raise typer.Exit(code=2) from exc |
| 778 | pairs.append(report.from_json(raw)) |
| 779 | # Short label — the filename without the ``.json`` suffix. |
| 780 | labels.append(path.stem) |
| 781 | |
| 782 | matrix = compare.build_matrix(pairs, labels=labels) |
| 783 | |
| 784 | if format is CompareFormat.JSON: |
| 785 | typer.echo(compare.render_json(matrix, regression_threshold=fail_on_regression)) |
| 786 | elif format in (CompareFormat.MARKDOWN, CompareFormat.MARKDOWN_LONG): |
| 787 | typer.echo(compare.render_markdown(matrix, regression_threshold=fail_on_regression)) |
| 788 | else: |
| 789 | compare.render_terminal( |
| 790 | matrix, |
| 791 | console=Console(), |
| 792 | regression_threshold=fail_on_regression, |
| 793 | ) |
| 794 | |
| 795 | # Exit-code gate: any probe whose last-run delta is ≤ -threshold is a |
| 796 | # regression. ``fail_on_regression=0`` disables the gate entirely. |
| 797 | if fail_on_regression > 0.0: |
| 798 | regressions = matrix.latest_regressions(fail_on_regression) |
| 799 | if regressions: |
| 800 | raise typer.Exit(code=1) |
| 801 | |
| 802 | |
| 803 | class TraceFormat(StrEnum): |
| 804 | """Allowed values for ``sway trace --format``.""" |
| 805 | |
| 806 | TERMINAL = "terminal" |
| 807 | MARKDOWN = "md" |
| 808 | MARKDOWN_LONG = "markdown" # alias kept for muscle memory |
| 809 | JSON = "json" |
| 810 | |
| 811 | |
| 812 | def trace_cmd( |
| 813 | trace_file: Annotated[ |
| 814 | Path, |
| 815 | typer.Argument( |
| 816 | help=("Path to a forward-pass trace JSONL produced by `sway run --trace <path>`."), |
| 817 | ), |
| 818 | ], |
| 819 | format: Annotated[ |
| 820 | TraceFormat, |
| 821 | typer.Option( |
| 822 | "--format", |
| 823 | help="Output format: terminal, md (alias: markdown), or json.", |
| 824 | ), |
| 825 | ] = TraceFormat.TERMINAL, |
| 826 | slowest: Annotated[ |
| 827 | int, |
| 828 | typer.Option( |
| 829 | "--slowest", |
| 830 | help="How many slowest-events rows to show. 0 hides that table.", |
| 831 | ), |
| 832 | ] = 10, |
| 833 | ) -> None: |
| 834 | """Analyze a forward-pass trace JSONL. |
| 835 | |
| 836 | Reads the per-event JSONL `sway run --trace` writes, aggregates |
| 837 | into per-probe + per-view summaries, and surfaces the top-N |
| 838 | slowest events. Intended for suite-performance investigation: |
| 839 | point at a captured trace and see which probe × view pair |
| 840 | dominated wall time, whether the S07 cache helped, and which |
| 841 | individual prompts took the longest. |
| 842 | """ |
| 843 | from dlm_sway.suite import trace_analysis |
| 844 | |
| 845 | try: |
| 846 | events = trace_analysis.load(trace_file) |
| 847 | except OSError as exc: |
| 848 | typer.echo(f"sway trace: cannot read {trace_file}: {exc}", err=True) |
| 849 | raise typer.Exit(code=2) from exc |
| 850 | if not events: |
| 851 | typer.echo(f"sway trace: no events in {trace_file}", err=True) |
| 852 | raise typer.Exit(code=1) |
| 853 | |
| 854 | report = trace_analysis.build_report(events, slowest_k=max(0, slowest)) |
| 855 | |
| 856 | if format is TraceFormat.JSON: |
| 857 | typer.echo(trace_analysis.render_json(report)) |
| 858 | elif format in (TraceFormat.MARKDOWN, TraceFormat.MARKDOWN_LONG): |
| 859 | typer.echo(trace_analysis.render_markdown(report)) |
| 860 | else: |
| 861 | trace_analysis.render_terminal(report, console=Console()) |
| 862 | |
| 863 | |
| 864 | class MineMode(StrEnum): |
| 865 | """``sway mine`` operates in one of two modes.""" |
| 866 | |
| 867 | PARAPHRASE = "paraphrase" |
| 868 | OUTLIERS = "outliers" |
| 869 | |
| 870 | |
| 871 | def mine_cmd( |
| 872 | spec: Annotated[Path, typer.Argument(help="Path to a sway.yaml spec.")], |
| 873 | mode: Annotated[ |
| 874 | MineMode, |
| 875 | typer.Option( |
| 876 | "--mode", |
| 877 | help=( |
| 878 | "``paraphrase``: sharpen every paraphrase_invariance case with mined " |
| 879 | "adversarial paraphrases. ``outliers``: rank the spec's delta_kl prompts " |
| 880 | "(or a corpus-derived pool) by per-prompt raw." |
| 881 | ), |
| 882 | ), |
| 883 | ] = MineMode.PARAPHRASE, |
| 884 | out: Annotated[ |
| 885 | Path | None, |
| 886 | typer.Option( |
| 887 | "--out", |
| 888 | "-o", |
| 889 | help=( |
| 890 | "Where to write the mined YAML fragment. Defaults to " |
| 891 | "``sway-mined-<mode>.yaml`` in the current directory." |
| 892 | ), |
| 893 | ), |
| 894 | ] = None, |
| 895 | top_k: Annotated[ |
| 896 | int, |
| 897 | typer.Option( |
| 898 | "--top-k", help="Keep the top-K candidates per case (paraphrase) or pool (outliers)." |
| 899 | ), |
| 900 | ] = 10, |
| 901 | n_candidates: Annotated[ |
| 902 | int, |
| 903 | typer.Option( |
| 904 | "--n-candidates", |
| 905 | help=( |
| 906 | "Paraphrase mode only — generate this many raw candidates before the " |
| 907 | "diversity filter. Higher = more coverage at more wall-time cost." |
| 908 | ), |
| 909 | ), |
| 910 | ] = 50, |
| 911 | from_corpus: Annotated[ |
| 912 | str | None, |
| 913 | typer.Option( |
| 914 | "--from-corpus", |
| 915 | help=( |
| 916 | "Outliers mode — draw the candidate pool from a packaged corpus " |
| 917 | "(``public_domain_en``) instead of the spec's own prompts." |
| 918 | ), |
| 919 | ), |
| 920 | ] = None, |
| 921 | seed: Annotated[ |
| 922 | int, |
| 923 | typer.Option( |
| 924 | "--seed", |
| 925 | help=( |
| 926 | "Seed for the generator + probe RNGs. Keep fixed to reproduce a " |
| 927 | "previous mining run — nlpaug's synonym and back-translation picks " |
| 928 | "are deterministic under this seed." |
| 929 | ), |
| 930 | ), |
| 931 | ] = 0, |
| 932 | ) -> None: |
| 933 | """Mine adversarial paraphrases or outlier prompts from a spec. |
| 934 | |
| 935 | **Paraphrase mode** (``--mode paraphrase``). For each |
| 936 | ``paraphrase_invariance`` case in the spec, generate candidate |
| 937 | paraphrases, diversity-filter them, and rank by the gap between |
| 938 | verbatim and paraphrased lift. The emitted YAML fragment contains |
| 939 | updated ``cases:`` that you can paste over the originals in your |
| 940 | spec — a memorizing adapter that passed the hand-written list will |
| 941 | typically fail the mined list. |
| 942 | |
| 943 | **Outliers mode** (``--mode outliers``). Rank the spec's |
| 944 | ``delta_kl`` prompt pool (or a corpus-derived pool via |
| 945 | ``--from-corpus``) by per-prompt raw divergence. Emitted fragment |
| 946 | lists top-K and bottom-K prompts, split into two blocks. |
| 947 | |
| 948 | The mined output is paste-compatible with the spec loader — no |
| 949 | schema bumps. Re-run ``sway gate`` after merging the mined list |
| 950 | to confirm the gate's behavior changed as expected. |
| 951 | """ |
| 952 | import yaml |
| 953 | |
| 954 | from dlm_sway.mining.outlier_miner import corpus_prompts, mine_outliers |
| 955 | from dlm_sway.mining.paraphrase_miner import mine_paraphrases |
| 956 | from dlm_sway.suite.loader import load_spec |
| 957 | |
| 958 | loaded_spec = load_spec(spec) |
| 959 | out_path = out or Path(f"sway-mined-{mode.value}.yaml") |
| 960 | |
| 961 | # Materialize the backend. Reuses ``_execute_spec``'s factory so the |
| 962 | # HF / API / MLX selection matches what ``sway run`` would do. |
| 963 | from dlm_sway.backends import build as build_backend |
| 964 | |
| 965 | backend = build_backend(loaded_spec.models.ft) |
| 966 | |
| 967 | if mode is MineMode.PARAPHRASE: |
| 968 | payload = _mine_paraphrase_payload( |
| 969 | loaded_spec, |
| 970 | backend, |
| 971 | mine_paraphrases, |
| 972 | top_k=top_k, |
| 973 | n_candidates=n_candidates, |
| 974 | seed=seed, |
| 975 | ) |
| 976 | else: |
| 977 | candidate_pool = ( |
| 978 | corpus_prompts(from_corpus) if from_corpus else _collect_delta_kl_prompts(loaded_spec) |
| 979 | ) |
| 980 | if not candidate_pool: |
| 981 | typer.secho( |
| 982 | "sway mine --outliers: no candidate prompts found. Either add delta_kl " |
| 983 | "prompts to the spec or pass --from-corpus.", |
| 984 | err=True, |
| 985 | fg=typer.colors.RED, |
| 986 | ) |
| 987 | raise typer.Exit(code=2) |
| 988 | result = mine_outliers( |
| 989 | probe_kind="delta_kl", |
| 990 | candidate_prompts=candidate_pool, |
| 991 | backend=backend, |
| 992 | top_k=top_k, |
| 993 | seed=seed, |
| 994 | ) |
| 995 | payload = _outlier_result_to_yaml(result) |
| 996 | |
| 997 | out_path.write_text(yaml.safe_dump(payload, sort_keys=False), encoding="utf-8") |
| 998 | typer.echo(f"wrote {out_path}") |
| 999 | |
| 1000 | |
| 1001 | def _mine_paraphrase_payload( |
| 1002 | spec: Any, |
| 1003 | backend: Any, |
| 1004 | miner: Any, |
| 1005 | *, |
| 1006 | top_k: int, |
| 1007 | n_candidates: int, |
| 1008 | seed: int, |
| 1009 | ) -> dict[str, Any]: |
| 1010 | """Run paraphrase mining on every paraphrase_invariance entry; shape into YAML.""" |
| 1011 | out_cases: list[dict[str, Any]] = [] |
| 1012 | for entry in spec.suite: |
| 1013 | if entry.get("kind") != "paraphrase_invariance": |
| 1014 | continue |
| 1015 | for case in entry.get("cases", []): |
| 1016 | prompt = case.get("prompt") |
| 1017 | gold = case.get("gold") |
| 1018 | if not prompt or not gold: |
| 1019 | continue |
| 1020 | mined = miner( |
| 1021 | prompt=prompt, |
| 1022 | gold=gold, |
| 1023 | backend=backend, |
| 1024 | n_candidates=n_candidates, |
| 1025 | top_k=top_k, |
| 1026 | seed=seed, |
| 1027 | ) |
| 1028 | out_cases.append( |
| 1029 | { |
| 1030 | "prompt": mined.seed_prompt, |
| 1031 | "gold": mined.gold, |
| 1032 | "paraphrases": [c.prompt for c in mined.candidates], |
| 1033 | "_mining_meta": { |
| 1034 | "top_gaps": [round(c.gap, 6) for c in mined.candidates], |
| 1035 | "verbatim_lift": ( |
| 1036 | round(mined.candidates[0].verbatim_lift, 6) |
| 1037 | if mined.candidates |
| 1038 | else None |
| 1039 | ), |
| 1040 | }, |
| 1041 | } |
| 1042 | ) |
| 1043 | return {"mined_cases": out_cases} |
| 1044 | |
| 1045 | |
| 1046 | def _collect_delta_kl_prompts(spec: Any) -> list[str]: |
| 1047 | """Pull every ``delta_kl`` entry's prompt pool into one flat list.""" |
| 1048 | seen: set[str] = set() |
| 1049 | out: list[str] = [] |
| 1050 | for entry in spec.suite: |
| 1051 | if entry.get("kind") != "delta_kl": |
| 1052 | continue |
| 1053 | for p in entry.get("prompts", []): |
| 1054 | if p not in seen: |
| 1055 | seen.add(p) |
| 1056 | out.append(p) |
| 1057 | return out |
| 1058 | |
| 1059 | |
| 1060 | def _outlier_result_to_yaml(result: Any) -> dict[str, Any]: |
| 1061 | """Format an :class:`OutlierResult` as a YAML-friendly dict.""" |
| 1062 | return { |
| 1063 | "mined_outliers": { |
| 1064 | "probe_kind": result.probe_kind, |
| 1065 | "top": [ |
| 1066 | {"prompt": c.prompt, "raw": round(c.raw, 6), "index": c.index} for c in result.top |
| 1067 | ], |
| 1068 | "bottom": [ |
| 1069 | {"prompt": c.prompt, "raw": round(c.raw, 6), "index": c.index} |
| 1070 | for c in result.bottom |
| 1071 | ], |
| 1072 | } |
| 1073 | } |
| 1074 | |
| 1075 | |
| 1076 | # -- helpers ----------------------------------------------------------- |
| 1077 | |
| 1078 | |
| 1079 | _BUILTIN_QUICK_PROMPTS: tuple[str, ...] = ( |
| 1080 | "The quick brown fox", |
| 1081 | "Once upon a time", |
| 1082 | "The answer to the question is", |
| 1083 | "One important lesson is", |
| 1084 | "In my opinion,", |
| 1085 | "The first step is to", |
| 1086 | "Remember that", |
| 1087 | "A common mistake is", |
| 1088 | ) |
| 1089 | |
| 1090 | |
| 1091 | def _load_prompts(path: Path) -> tuple[str, ...]: |
| 1092 | return tuple( |
| 1093 | line.strip() for line in path.read_text(encoding="utf-8").splitlines() if line.strip() |
| 1094 | ) |
| 1095 | |
| 1096 | |
| 1097 | def _execute_spec( |
| 1098 | path: Path, |
| 1099 | *, |
| 1100 | weights_override: dict[str, float] | None = None, |
| 1101 | trace_path: Path | None = None, |
| 1102 | ) -> tuple[SuiteResult, SwayScore]: |
| 1103 | """Load a spec, build a backend, run the suite, fold scores. Shared |
| 1104 | by ``run`` and ``gate``. Picks up .dlm-derived sections when the |
| 1105 | spec's ``dlm_source`` is set. |
| 1106 | |
| 1107 | ``weights_override`` takes precedence over ``spec.defaults.score_weights`` |
| 1108 | (which itself takes precedence over the compile-time defaults). The |
| 1109 | CLI hands through ``--weights k=v,k=v`` via this parameter. |
| 1110 | """ |
| 1111 | from dlm_sway.backends import build as build_backend |
| 1112 | from dlm_sway.backends import build_two_separate |
| 1113 | from dlm_sway.probes.base import validate_all_probes |
| 1114 | from dlm_sway.suite.loader import load_spec |
| 1115 | from dlm_sway.suite.runner import run as run_suite |
| 1116 | from dlm_sway.suite.score import compute as compute_score |
| 1117 | |
| 1118 | spec = load_spec(path) |
| 1119 | # B7: validate every probe entry before paying the cost of loading |
| 1120 | # a backend. A user with a typo in `kind:` shouldn't wait minutes |
| 1121 | # for the model to download just to learn they spelled the probe |
| 1122 | # name wrong. |
| 1123 | validate_all_probes(spec.suite) |
| 1124 | sections = None |
| 1125 | doc_text = None |
| 1126 | if spec.dlm_source is not None: |
| 1127 | import importlib |
| 1128 | |
| 1129 | try: |
| 1130 | resolver = importlib.import_module("dlm_sway.integrations.dlm.resolver") |
| 1131 | handle = resolver.resolve_dlm(Path(spec.dlm_source)) |
| 1132 | sections = handle.sections |
| 1133 | doc_text = handle.doc_text |
| 1134 | except ImportError: |
| 1135 | # D8: don't silently swallow. The user wrote ``dlm_source`` |
| 1136 | # in their YAML expecting the bridge to populate sections; |
| 1137 | # warn loudly so they know why downstream attribution |
| 1138 | # probes are SKIPping. |
| 1139 | typer.secho( |
| 1140 | f"warning: spec sets dlm_source={spec.dlm_source!r} but the " |
| 1141 | f"[dlm] extra is not installed — sections not provided " |
| 1142 | f"(pip install 'dlm-sway[dlm]')", |
| 1143 | err=True, |
| 1144 | fg=typer.colors.YELLOW, |
| 1145 | ) |
| 1146 | sections = None |
| 1147 | except SwayError as exc: |
| 1148 | # The bridge imported but failed (no adapter, malformed |
| 1149 | # .dlm, etc). Same surface — warn, don't crash the suite. |
| 1150 | typer.secho( |
| 1151 | f"warning: dlm_source={spec.dlm_source!r} did not resolve: {exc}", |
| 1152 | err=True, |
| 1153 | fg=typer.colors.YELLOW, |
| 1154 | ) |
| 1155 | sections = None |
| 1156 | if spec.defaults.differential: |
| 1157 | backend: Any = build_backend(spec.models.ft) |
| 1158 | else: |
| 1159 | backend = build_two_separate(spec.models) |
| 1160 | try: |
| 1161 | result = run_suite( |
| 1162 | spec, |
| 1163 | backend, |
| 1164 | spec_path=str(path), |
| 1165 | sections=sections, |
| 1166 | doc_text=doc_text, |
| 1167 | trace_path=trace_path, |
| 1168 | ) |
| 1169 | finally: |
| 1170 | _close_if_possible(backend) |
| 1171 | effective_weights = weights_override or spec.defaults.score_weights |
| 1172 | score_obj = compute_score(result, weights=effective_weights) |
| 1173 | return result, score_obj |
| 1174 | |
| 1175 | |
| 1176 | def _parse_weights_flag(raw: str | None) -> dict[str, float] | None: |
| 1177 | """Parse ``--weights k=v,k=v`` into a dict; pydantic validates on use. |
| 1178 | |
| 1179 | Returns ``None`` when the flag is empty / unset. Pydantic's |
| 1180 | ``SuiteDefaults._validate_weights`` is re-invoked indirectly via |
| 1181 | ``SwayScore`` — so any unknown category or negative value surfaces |
| 1182 | the same error whether set in YAML or on the command line. |
| 1183 | """ |
| 1184 | if not raw: |
| 1185 | return None |
| 1186 | out: dict[str, float] = {} |
| 1187 | for pair in raw.split(","): |
| 1188 | pair = pair.strip() |
| 1189 | if not pair: |
| 1190 | continue |
| 1191 | if "=" not in pair: |
| 1192 | raise typer.BadParameter( |
| 1193 | f"--weights: expected 'key=value' pairs, got {pair!r}. " |
| 1194 | "Example: --weights adherence=0.4,attribution=0.3" |
| 1195 | ) |
| 1196 | key, _, value = pair.partition("=") |
| 1197 | key = key.strip() |
| 1198 | try: |
| 1199 | out[key] = float(value.strip()) |
| 1200 | except ValueError as exc: |
| 1201 | raise typer.BadParameter(f"--weights: {value!r} for {key!r} is not a number") from exc |
| 1202 | return out or None |
| 1203 | |
| 1204 | |
| 1205 | def _close_if_possible(backend: object) -> None: |
| 1206 | close = getattr(backend, "close", None) |
| 1207 | if callable(close): |
| 1208 | close() |
| 1209 | |
| 1210 | |
| 1211 | # --- convert-adapter (S24, F01) ------------------------------------------ |
| 1212 | |
| 1213 | |
| 1214 | class ConvertTarget(StrEnum): |
| 1215 | MLX = "mlx" |
| 1216 | |
| 1217 | |
| 1218 | def convert_adapter_cmd( |
| 1219 | src: Annotated[Path, typer.Argument(help="PEFT adapter directory to convert.")], |
| 1220 | dst: Annotated[Path, typer.Argument(help="Output directory for the converted adapter.")], |
| 1221 | target: Annotated[ |
| 1222 | ConvertTarget, |
| 1223 | typer.Option("--target", help="Output adapter format. Currently only 'mlx'."), |
| 1224 | ] = ConvertTarget.MLX, |
| 1225 | overwrite: Annotated[ |
| 1226 | bool, |
| 1227 | typer.Option("--overwrite", help="Replace any existing adapter at dst."), |
| 1228 | ] = False, |
| 1229 | ) -> None: |
| 1230 | """Convert a PEFT LoRA adapter to another backend's format. |
| 1231 | |
| 1232 | Today the only target is ``mlx`` — converts ``adapter_model.safetensors`` + |
| 1233 | ``adapter_config.json`` (PEFT) to ``adapters.safetensors`` + |
| 1234 | ``adapter_config.json`` (mlx-lm). Closes the F01 doc-vs-code gap so |
| 1235 | the MLX backend works on dlm-trained / any PEFT-trained adapters |
| 1236 | without manual conversion. |
| 1237 | """ |
| 1238 | from dlm_sway.backends._mlx_convert import MlxConvertError, convert_peft_to_mlx |
| 1239 | |
| 1240 | if target is not ConvertTarget.MLX: |
| 1241 | raise typer.BadParameter(f"unsupported target {target!r}") |
| 1242 | try: |
| 1243 | report = convert_peft_to_mlx(src, dst, overwrite=overwrite) |
| 1244 | except MlxConvertError as exc: |
| 1245 | typer.secho(f"convert-adapter: {exc}", fg=typer.colors.RED, err=True) |
| 1246 | raise typer.Exit(code=1) from exc |
| 1247 | except SwayError as exc: |
| 1248 | typer.secho(f"convert-adapter: {exc}", fg=typer.colors.RED, err=True) |
| 1249 | raise typer.Exit(code=1) from exc |
| 1250 | |
| 1251 | src_kb = report["src_bytes"] / 1024 |
| 1252 | dst_kb = report["dst_bytes"] / 1024 |
| 1253 | typer.echo( |
| 1254 | f"converted: {src} → {dst} rank={report['rank']} " |
| 1255 | f"scale={report['scale']:.3f} num_keys={report['num_keys']} " |
| 1256 | f"({src_kb:.1f} KB → {dst_kb:.1f} KB)" |
| 1257 | ) |
| 1258 | if report["modules_to_save_skipped"]: |
| 1259 | typer.secho( |
| 1260 | f"warning: {len(report['modules_to_save_skipped'])} modules_to_save tensor(s) " |
| 1261 | f"skipped (mlx-lm's LoRA loader doesn't apply full-weight overrides). " |
| 1262 | f"Sample: {report['modules_to_save_skipped'][:1]!r}", |
| 1263 | fg=typer.colors.YELLOW, |
| 1264 | err=True, |
| 1265 | ) |
| 1266 | |
| 1267 | |
| 1268 | # --- pack / unpack (S26, X3) ---------------------------------------------- |
| 1269 | |
| 1270 | |
| 1271 | def pack_cmd( |
| 1272 | spec_path: Annotated[Path, typer.Argument(help="Path to a sway.yaml to pack.")], |
| 1273 | out: Annotated[ |
| 1274 | Path | None, |
| 1275 | typer.Option( |
| 1276 | "--out", |
| 1277 | "-o", |
| 1278 | help="Output tarball path. Defaults to <spec-stem>.swaypack.tar.gz next to spec.", |
| 1279 | ), |
| 1280 | ] = None, |
| 1281 | include_golden: Annotated[ |
| 1282 | Path | None, |
| 1283 | typer.Option( |
| 1284 | "--include-golden", |
| 1285 | help=( |
| 1286 | "Path to a JSON sway-run report (from `sway run --json` or " |
| 1287 | "`sway report --format json`) to bundle for verification." |
| 1288 | ), |
| 1289 | ), |
| 1290 | ] = None, |
| 1291 | include_null_cache: Annotated[ |
| 1292 | bool, |
| 1293 | typer.Option( |
| 1294 | "--include-null-cache/--no-include-null-cache", |
| 1295 | help=( |
| 1296 | "Bundle ~/.dlm-sway/null-stats/*.json into the pack so the " |
| 1297 | "consumer doesn't need to re-calibrate. Default: include." |
| 1298 | ), |
| 1299 | ), |
| 1300 | ] = True, |
| 1301 | max_size_mb: Annotated[ |
| 1302 | int, |
| 1303 | typer.Option( |
| 1304 | "--max-size-mb", |
| 1305 | help="Refuse to write a pack larger than this (default 50 MB).", |
| 1306 | ), |
| 1307 | ] = 50, |
| 1308 | ) -> None: |
| 1309 | """Bundle a spec + its inputs + null-stats cache into a portable swaypack tarball. |
| 1310 | |
| 1311 | The result is a single ``.swaypack.tar.gz`` you can share with a |
| 1312 | coworker or check into a release repo. The receiver runs |
| 1313 | ``sway unpack <pack>`` and then ``sway run`` against the unpacked |
| 1314 | spec — identical verdict to the original run, no live dlm or |
| 1315 | network needed. |
| 1316 | """ |
| 1317 | from dlm_sway.cli._pack import PackError, pack_spec |
| 1318 | |
| 1319 | if out is None: |
| 1320 | out = spec_path.with_suffix("").with_name( |
| 1321 | f"{spec_path.with_suffix('').name}.swaypack.tar.gz" |
| 1322 | ) |
| 1323 | |
| 1324 | try: |
| 1325 | report = pack_spec( |
| 1326 | spec_path, |
| 1327 | out_path=out, |
| 1328 | include_golden=include_golden, |
| 1329 | include_null_cache=include_null_cache, |
| 1330 | max_size_bytes=max_size_mb * 1024 * 1024, |
| 1331 | ) |
| 1332 | except PackError as exc: |
| 1333 | typer.secho(f"pack: {exc}", fg=typer.colors.RED, err=True) |
| 1334 | raise typer.Exit(code=1) from exc |
| 1335 | except SwayError as exc: |
| 1336 | typer.secho(f"pack: {exc}", fg=typer.colors.RED, err=True) |
| 1337 | raise typer.Exit(code=1) from exc |
| 1338 | |
| 1339 | size_kb = report.size_bytes / 1024 |
| 1340 | typer.echo( |
| 1341 | f"wrote {report.out_path} ({size_kb:.1f} KB) " |
| 1342 | f"sections={report.section_bytes}b " |
| 1343 | f"null_stats={report.null_stats_count} " |
| 1344 | f"golden={'yes' if report.golden_included else 'no'}" |
| 1345 | ) |
| 1346 | |
| 1347 | |
| 1348 | def unpack_cmd( |
| 1349 | pack_path: Annotated[Path, typer.Argument(help="Path to a *.swaypack.tar.gz.")], |
| 1350 | out: Annotated[ |
| 1351 | Path | None, |
| 1352 | typer.Option( |
| 1353 | "--out", |
| 1354 | "-o", |
| 1355 | help="Parent directory to extract into. Default: cwd.", |
| 1356 | ), |
| 1357 | ] = None, |
| 1358 | ) -> None: |
| 1359 | """Extract a swaypack into ``out``; print the next ``sway run`` invocation. |
| 1360 | |
| 1361 | The pack lands at ``<out>/swaypack/`` containing ``sway.yaml`` plus |
| 1362 | bundled artifacts. A ready-to-run command line is printed at the |
| 1363 | end including the ``SWAY_NULL_CACHE_DIR=...`` env var that |
| 1364 | redirects null-stats lookups at the bundled cache instead of the |
| 1365 | user's home directory. |
| 1366 | """ |
| 1367 | from dlm_sway.cli._unpack import UnpackError, unpack_swaypack |
| 1368 | |
| 1369 | target_dir = out if out is not None else Path.cwd() |
| 1370 | try: |
| 1371 | report = unpack_swaypack(pack_path, target_dir=target_dir) |
| 1372 | except UnpackError as exc: |
| 1373 | typer.secho(f"unpack: {exc}", fg=typer.colors.RED, err=True) |
| 1374 | raise typer.Exit(code=1) from exc |
| 1375 | except SwayError as exc: |
| 1376 | typer.secho(f"unpack: {exc}", fg=typer.colors.RED, err=True) |
| 1377 | raise typer.Exit(code=1) from exc |
| 1378 | |
| 1379 | typer.echo(f"extracted: {report.out_dir}") |
| 1380 | typer.echo(f" spec_path: {report.spec_path}") |
| 1381 | typer.echo(f" null_stats: {report.null_stats_dir or '(none in pack)'}") |
| 1382 | typer.echo(f" swaypack_version: {report.manifest.get('swaypack_version')}") |
| 1383 | typer.echo(f" packed_at: {report.manifest.get('packed_at')}") |
| 1384 | typer.echo("") |
| 1385 | typer.echo("To run the bundled spec:") |
| 1386 | if report.null_stats_dir is not None: |
| 1387 | typer.echo(f" SWAY_NULL_CACHE_DIR={report.null_stats_dir} sway run {report.spec_path}") |
| 1388 | else: |
| 1389 | typer.echo(f" sway run {report.spec_path}") |
| 1390 | |
| 1391 | |
| 1392 | def serve_cmd( |
| 1393 | host: Annotated[ |
| 1394 | str, |
| 1395 | typer.Option( |
| 1396 | "--host", |
| 1397 | help=( |
| 1398 | "Interface to bind. Default 127.0.0.1 (localhost only). " |
| 1399 | "Binding to 0.0.0.0 requires --api-key." |
| 1400 | ), |
| 1401 | ), |
| 1402 | ] = "127.0.0.1", |
| 1403 | port: Annotated[ |
| 1404 | int, |
| 1405 | typer.Option("--port", help="TCP port to bind."), |
| 1406 | ] = 8787, |
| 1407 | max_loaded_models: Annotated[ |
| 1408 | int, |
| 1409 | typer.Option( |
| 1410 | "--max-loaded-models", |
| 1411 | help=( |
| 1412 | "How many backends to keep warm in memory. Each loaded " |
| 1413 | "model holds its own VRAM/RAM; default 2 fits a 16 GB GPU " |
| 1414 | "with two ~1.5B fp16 adapters." |
| 1415 | ), |
| 1416 | ), |
| 1417 | ] = 2, |
| 1418 | api_key: Annotated[ |
| 1419 | str | None, |
| 1420 | typer.Option( |
| 1421 | "--api-key", |
| 1422 | help=( |
| 1423 | "Bearer token required on every non-/health request. " |
| 1424 | "Required when --host is not loopback." |
| 1425 | ), |
| 1426 | ), |
| 1427 | ] = None, |
| 1428 | log_level: Annotated[ |
| 1429 | str, |
| 1430 | typer.Option("--log-level", help="uvicorn log level."), |
| 1431 | ] = "info", |
| 1432 | ) -> None: |
| 1433 | """Run the warm-backend HTTP daemon (S36). |
| 1434 | |
| 1435 | First call loads the backend (~15s); subsequent calls reuse it |
| 1436 | (~2s). See ``sway run`` for the equivalent one-shot CLI. |
| 1437 | """ |
| 1438 | try: |
| 1439 | import uvicorn # noqa: F401 — presence check |
| 1440 | except ImportError as exc: |
| 1441 | typer.secho( |
| 1442 | "sway serve requires the [serve] extra: pip install 'dlm-sway[serve]'", |
| 1443 | fg=typer.colors.RED, |
| 1444 | err=True, |
| 1445 | ) |
| 1446 | raise typer.Exit(code=2) from exc |
| 1447 | |
| 1448 | from dlm_sway.serve.app import create_app, parse_host_port |
| 1449 | from dlm_sway.serve.cache import BackendCache |
| 1450 | |
| 1451 | # Public-bind safety — refuse before any uvicorn startup work. |
| 1452 | loopback = host in ("127.0.0.1", "::1", "localhost") |
| 1453 | if not loopback and api_key is None: |
| 1454 | typer.secho( |
| 1455 | f"refusing to bind {host}:{port} without --api-key. " |
| 1456 | "Either pass --api-key <key> or use --host 127.0.0.1.", |
| 1457 | fg=typer.colors.RED, |
| 1458 | err=True, |
| 1459 | ) |
| 1460 | raise typer.Exit(code=2) |
| 1461 | |
| 1462 | parse_host_port(host, port) |
| 1463 | if max_loaded_models < 1: |
| 1464 | typer.secho("--max-loaded-models must be >= 1", fg=typer.colors.RED, err=True) |
| 1465 | raise typer.Exit(code=2) |
| 1466 | |
| 1467 | cache = BackendCache(max_size=max_loaded_models) |
| 1468 | app = create_app(cache=cache, api_key=api_key) |
| 1469 | |
| 1470 | typer.echo(f"sway serve {__version__} listening on http://{host}:{port}") |
| 1471 | typer.echo(f" max_loaded_models={max_loaded_models} auth={'yes' if api_key else 'no'}") |
| 1472 | if not loopback: |
| 1473 | typer.secho( |
| 1474 | " WARNING: bound to a non-loopback interface — anyone on " |
| 1475 | "this network with the API key can drive your GPU.", |
| 1476 | fg=typer.colors.YELLOW, |
| 1477 | ) |
| 1478 | |
| 1479 | import uvicorn as _uvicorn |
| 1480 | |
| 1481 | _uvicorn.run(app, host=host, port=port, log_level=log_level) |