| 1 |
"""`ollama run` wrapper — post-export smoke test. |
| 2 |
|
| 3 |
Default prompt is `"Hello."` — just enough to verify the model loads |
| 4 |
and generates one coherent reply. Empty output or non-zero exit |
| 5 |
raises `OllamaSmokeError`. The runner appends the first line of |
| 6 |
stdout to `manifest.exports[-1].smoke_output_first_line` so the |
| 7 |
smoke is auditable from `dlm show` without re-running inference. |
| 8 |
|
| 9 |
`--no-smoke` skips this entirely at the runner level; this module |
| 10 |
exists only to be called on the happy path. |
| 11 |
""" |
| 12 |
|
| 13 |
from __future__ import annotations |
| 14 |
|
| 15 |
import subprocess # nosec B404 |
| 16 |
from pathlib import Path |
| 17 |
|
| 18 |
from dlm.export.ollama.binary import locate_ollama |
| 19 |
from dlm.export.ollama.errors import OllamaSmokeError |
| 20 |
|
| 21 |
_DEFAULT_TIMEOUT_SECONDS = 120.0 |
| 22 |
_DEFAULT_PROMPT = "Hello." |
| 23 |
|
| 24 |
|
| 25 |
def ollama_run( |
| 26 |
*, |
| 27 |
name: str, |
| 28 |
prompt: str = _DEFAULT_PROMPT, |
| 29 |
binary: Path | None = None, |
| 30 |
timeout: float = _DEFAULT_TIMEOUT_SECONDS, |
| 31 |
) -> str: |
| 32 |
"""Invoke `ollama run <name> <prompt>`, return stdout. |
| 33 |
|
| 34 |
Raises `OllamaSmokeError` on: |
| 35 |
- Non-zero exit code. |
| 36 |
- Empty stdout (the model "succeeded" but produced nothing — a |
| 37 |
runaway-stop scenario this smoke test is specifically guarding against). |
| 38 |
- Subprocess timeout. |
| 39 |
""" |
| 40 |
exe = binary or locate_ollama() |
| 41 |
try: |
| 42 |
result = subprocess.run( # nosec B603 — caller-controlled argv |
| 43 |
[str(exe), "run", name, prompt], |
| 44 |
capture_output=True, |
| 45 |
text=True, |
| 46 |
check=False, |
| 47 |
timeout=timeout, |
| 48 |
) |
| 49 |
except subprocess.TimeoutExpired as exc: |
| 50 |
raise OllamaSmokeError( |
| 51 |
stdout=(exc.stdout or b"").decode("utf-8", errors="replace"), |
| 52 |
stderr=(exc.stderr or b"").decode("utf-8", errors="replace") |
| 53 |
+ f"\n(timed out after {timeout}s)", |
| 54 |
) from exc |
| 55 |
|
| 56 |
if result.returncode != 0: |
| 57 |
raise OllamaSmokeError(stdout=result.stdout, stderr=result.stderr) |
| 58 |
if not result.stdout.strip(): |
| 59 |
raise OllamaSmokeError( |
| 60 |
stdout=result.stdout, |
| 61 |
stderr=result.stderr |
| 62 |
+ "\n(empty stdout — the model produced no output for the smoke prompt)", |
| 63 |
) |
| 64 |
return result.stdout |
| 65 |
|
| 66 |
|
| 67 |
def first_line(text: str) -> str: |
| 68 |
"""Return the first non-empty line of `text`, truncated to 200 chars. |
| 69 |
|
| 70 |
Used to stamp `ExportSummary.smoke_output_first_line` without |
| 71 |
bloating the manifest JSON. |
| 72 |
""" |
| 73 |
for line in text.splitlines(): |
| 74 |
stripped = line.strip() |
| 75 |
if stripped: |
| 76 |
return stripped[:200] |
| 77 |
return "" |