mjpsm's picture
Deploy Progression API (llama-cpp-python)
63d4e3a verified
Raw History Blame Contribute Delete
4.23 kB
import json
from tests.conftest import VALID_OUTPUT
GENERATE = "/v1/progression/generate"
OLD_METADATA_FIELDS = {
"metadata",
"model_repo",
"model_file",
"version",
"quantization",
"runtime",
"elapsed_seconds",
}
def test_health_ready(client):
assert client.get("/health").json() == {
"status": "ok",
"model_ready": True,
}
def test_health_degraded(client, fake_runtime):
fake_runtime.is_ready = False
assert client.get("/health").json() == {
"status": "degraded",
"model_ready": False,
}
def test_models(client):
entry = client.get("/models").json()["progression"]["v1"]
assert entry["runtime"] == "llama-cpp-python"
assert entry["quantization"] == "Q8_0"
assert entry["prompt_format"] == "raw_sft_completion"
assert entry["status"] == "ready"
def test_generate_success(client, fake_runtime, valid_request):
response = client.post(GENERATE, json=valid_request)
assert response.status_code == 200
body = response.json()
assert set(body) == {"progression", "usage"}
assert body["progression"] == VALID_OUTPUT["progression"]
assert set(body["usage"]) == {"input_tokens", "output_tokens"}
assert body["usage"]["input_tokens"] > 0
assert body["usage"]["output_tokens"] > 0
assert not OLD_METADATA_FIELDS & set(body)
# The exact raw SFT prompt is what reaches the model.
assert len(fake_runtime.prompts) == 1
assert fake_runtime.prompts[0].endswith("\n\nOUTPUT:\n")
assert (
'"previous_activity":{"title":"Practice the skill"}'
in fake_runtime.prompts[0]
)
def test_generation_field_rejected(client, fake_runtime, valid_request):
valid_request["generation"] = {"temperature": 1.0, "seed": 7}
response = client.post(GENERATE, json=valid_request)
assert response.status_code == 422
assert fake_runtime.prompts == []
def test_public_schema_has_no_generation_field(client):
schema = client.get("/openapi.json").json()
properties = schema["components"]["schemas"]["ProgressionRequest"][
"properties"
]
assert set(properties) == {
"version",
"knowledge_submission",
"wisdom",
"village_goal",
"previous_activity_title",
}
def test_old_wisdom_schema_rejected(client, valid_request):
valid_request["wisdom"] = {
"title": "Past learning",
"description": "Older record",
"content": "I learned something before.",
}
assert client.post(GENERATE, json=valid_request).status_code == 422
def test_wisdom_capitalization_enforced(client, valid_request):
wisdom = valid_request["wisdom"]
wisdom["bookType"] = wisdom.pop("BookType")
assert client.post(GENERATE, json=valid_request).status_code == 422
def test_unknown_version(client, fake_runtime, valid_request):
valid_request["version"] = "v2"
response = client.post(GENERATE, json=valid_request)
assert response.status_code == 404
assert "v2" in response.json()["detail"]
assert fake_runtime.prompts == []
def test_model_not_ready(client, fake_runtime, valid_request):
fake_runtime.is_ready = False
assert client.post(GENERATE, json=valid_request).status_code == 503
def test_text_before_json_is_accepted(client, fake_runtime, valid_request):
fake_runtime.output = "Sure:\n" + json.dumps(VALID_OUTPUT)
response = client.post(GENERATE, json=valid_request)
assert response.status_code == 200
assert response.json()["progression"] == VALID_OUTPUT["progression"]
def test_malformed_json_returns_502(client, fake_runtime, valid_request):
fake_runtime.output = '{"progression": {"knowledge_state": "COMP'
response = client.post(GENERATE, json=valid_request)
assert response.status_code == 502
assert response.json()["detail"]["raw_output"] == fake_runtime.output
def test_invalid_schema_returns_502(client, fake_runtime, valid_request):
fake_runtime.output = json.dumps(
{"progression": {"knowledge_state": "DONE"}}
)
response = client.post(GENERATE, json=valid_request)
assert response.status_code == 502
assert "validation_error" in response.json()["detail"]