103 lines
5.1 KiB
Python
103 lines
5.1 KiB
Python
#!/usr/bin/env python3
|
||
"""V4.1 HTTP/session plumbing with the sparse zero-weight GGUF, NOT model quality."""
|
||
import argparse
|
||
import json
|
||
from pathlib import Path
|
||
import signal
|
||
import socket
|
||
import subprocess
|
||
import time
|
||
import urllib.error
|
||
import urllib.request
|
||
|
||
|
||
def main():
|
||
parser = argparse.ArgumentParser(description=__doc__)
|
||
parser.add_argument("model", help="sparse zero-weight V4.1 fixture")
|
||
parser.add_argument("--output", required=True, type=Path)
|
||
parser.add_argument("--port", default=0, type=int)
|
||
args = parser.parse_args()
|
||
args.output.mkdir(parents=True, exist_ok=True)
|
||
with socket.socket() as listener:
|
||
listener.bind(("127.0.0.1", args.port))
|
||
args.port = listener.getsockname()[1]
|
||
base = f"http://127.0.0.1:{args.port}"
|
||
trace = args.output / "trace.log"
|
||
model = "deepseek-v4.1-flash"
|
||
|
||
def request(path, body=None):
|
||
req = urllib.request.Request(base + path,
|
||
data=None if body is None else json.dumps(body).encode(),
|
||
headers={"Content-Type": "application/json"})
|
||
try:
|
||
with urllib.request.urlopen(req, timeout=180) as response:
|
||
data = response.read().decode()
|
||
except urllib.error.HTTPError as error:
|
||
raise AssertionError(error.read().decode()) from error
|
||
return data if body and body.get("stream") else json.loads(data)
|
||
|
||
with (args.output / "server.log").open("w") as log:
|
||
proc = subprocess.Popen(["./ds4-server", "-m", args.model,
|
||
"--ssd-streaming", "--ssd-streaming-cache-experts", "512",
|
||
"--ctx", "512", "--tokens", "4", "--host", "127.0.0.1",
|
||
"--port", str(args.port), "--trace", str(trace),
|
||
"--kv-disk-dir", str(args.output / "kv"), "--kv-disk-space-mb", "128"],
|
||
stdout=log, stderr=subprocess.STDOUT)
|
||
try:
|
||
deadline = time.monotonic() + 180
|
||
while True:
|
||
assert proc.poll() is None, "server exited during startup"
|
||
try:
|
||
models = request("/v1/models")
|
||
break
|
||
except urllib.error.URLError:
|
||
assert time.monotonic() < deadline, "server startup timed out"
|
||
time.sleep(0.2)
|
||
assert any(item["id"] == model for item in models["data"])
|
||
body = {"model": model, "messages": [{"role": "user", "content": "Hello"}],
|
||
"temperature": 0, "reasoning_effort": "none", "max_tokens": 1}
|
||
cold = request("/v1/chat/completions", body)
|
||
warm = request("/v1/chat/completions", body)
|
||
assert cold["choices"][0]["message"] == warm["choices"][0]["message"]
|
||
body["messages"] += [warm["choices"][0]["message"], {"role": "user", "content": "Again"}]
|
||
continued = request("/v1/chat/completions", body)
|
||
assert continued["usage"]["prompt_tokens_details"]["cached_tokens"] > 0
|
||
body["messages"] = [{"role": "user", "content": "Read a.c"},
|
||
{"role": "assistant", "content": "", "tool_calls": [{"id": "call_fixture",
|
||
"type": "function", "function": {"name": "read", "arguments": '{"path":"a.c"}'}}]},
|
||
{"role": "tool", "tool_call_id": "call_fixture", "content": "int main(void) {}"}]
|
||
tool = request("/v1/chat/completions", body)
|
||
assert tool["choices"][0]["finish_reason"] == "length"
|
||
body["stream"] = True
|
||
stream = request("/v1/chat/completions", body)
|
||
assert "data: [DONE]" in stream
|
||
for line in stream.splitlines():
|
||
if line.startswith("data: ") and line != "data: [DONE]":
|
||
json.loads(line[6:])
|
||
response = request("/v1/responses", {"model": model, "input": "Hi",
|
||
"reasoning": {"effort": "none"}, "max_output_tokens": 4, "temperature": 0})
|
||
assert response["object"] == "response"
|
||
anthropic = request("/v1/messages", {"model": model,
|
||
"system": "Be precise.", "messages": [{"role": "user", "content": "Anthropic greeting"}],
|
||
"thinking": {"type": "disabled"}, "max_tokens": 1, "temperature": 0})
|
||
assert anthropic["type"] == "message"
|
||
(args.output / "results.json").write_text(json.dumps({
|
||
"cold": cold, "warm": warm, "continued": continued,
|
||
"tool_replay": tool, "responses": response, "anthropic": anthropic}, indent=2))
|
||
finally:
|
||
if proc.poll() is None:
|
||
proc.send_signal(signal.SIGINT)
|
||
try:
|
||
proc.wait(timeout=60)
|
||
except subprocess.TimeoutExpired:
|
||
proc.kill()
|
||
proc.wait()
|
||
assert proc.returncode == 0, f"server exit status {proc.returncode}"
|
||
assert "<｜DSML｜ calls>" in trace.read_text()
|
||
assert "<｜begin▁of▁sentence｜><｜System｜>Be precise.<｜User｜>Anthropic greeting" in trace.read_text()
|
||
print("PASS: V4.1 model ID, repeat request, continued cache, tool replay, SSE, Responses, clean shutdown")
|
||
|
||
|
||
if __name__ == "__main__":
|
||
main()
|