Dumitru Picalau
9ec21566f3
Settings › AI › Advanced › Sampling sets top_p, top_k, min_p and repeat/presence penalties per provider slot (empty keeps the server's default). They go only to local and OpenAI-compatible servers. The chat eval gains --seed for a fixed sampling seed per scenario and run.
49 lines
2.4 KiB
Python
49 lines
2.4 KiB
Python
"""Optional samplers (Settings › AI assistant › Sampling) reach each server
|
||
in its own spelling, and never Claude or OpenAI itself; the assistant's
|
||
own temperatures stay."""
|
||
from __future__ import annotations
|
||
|
||
from engine.ai import config as ai_config
|
||
from tests.wp_api.test_ai_api import _chat
|
||
from wp_api import _ai_client
|
||
|
||
QWEN = {"top_k": 20, "top_p": 0.8, "min_p": 0, "repeat_penalty": 1.0}
|
||
|
||
|
||
def test_fields_per_server():
|
||
base = {"sampling": dict(QWEN)}
|
||
assert _ai_client.sampling_fields({**base, "api": "openai", "provider": "lmstudio",
|
||
"local": True}) == QWEN
|
||
got = _ai_client.sampling_fields({**base, "api": "openai", "provider": "openrouter",
|
||
"base_url": "https://openrouter.ai/api/v1"})
|
||
assert got["repetition_penalty"] == 1.0 and "repeat_penalty" not in got
|
||
assert _ai_client.sampling_fields({**base, "api": "anthropic"}) == {}
|
||
assert _ai_client.sampling_fields({**base, "api": "openai", "provider": "openai",
|
||
"base_url": "https://api.openai.com/v1"}) == {}
|
||
assert _ai_client.sampling_fields({"api": "openai", "provider": "lmstudio"}) == {}
|
||
|
||
|
||
async def test_a_chat_sends_them_beside_its_own_temperature(wp_client, fake_llm):
|
||
ai_config.update({"enabled": True, "provider": "lmstudio",
|
||
"base_url": str(fake_llm.make_url("/v1")), "model": "qwen3.5-9b",
|
||
"sampling": QWEN})
|
||
fake_llm.chat_steps.append({"message": "Hello.", "calls": []})
|
||
n = len(fake_llm.seen)
|
||
await _chat(wp_client, "hi there, what can you do?")
|
||
asks = [x["body"] for x in fake_llm.seen[n:] if "messages" in x["body"]]
|
||
assert asks
|
||
for body in asks:
|
||
assert {k: body[k] for k in QWEN} == QWEN
|
||
assert body["temperature"] in (0.15, 0.4)
|
||
|
||
|
||
async def test_ollama_gets_them_in_options(wp_client, fake_llm):
|
||
ai_config.update({"enabled": True, "provider": "ollama",
|
||
"base_url": str(fake_llm.make_url("")), "model": "qwen3.5:9b",
|
||
"sampling": {"top_k": 20}})
|
||
fake_llm.chat_steps.append({"message": "Hello.", "calls": []})
|
||
n = len(fake_llm.seen)
|
||
await _chat(wp_client, "hi there, what can you do?")
|
||
opts = [x["body"].get("options") for x in fake_llm.seen[n:] if x["path"] == "/api/chat"]
|
||
assert opts and all(o["top_k"] == 20 and "temperature" in o for o in opts)
|