Deploy audited RTX 6000D Flash-Next 128K multimodal baseline

2026-09-18 10:47:26 +08:00
commit 7e53821920
45 changed files with 1745 additions and 0 deletions
@@ -0,0 +1,65 @@
+"""Small live API acceptance suite; writes evidence without API credentials."""
+import argparse
+import json
+import time
+import urllib.request
+from pathlib import Path
+
+p = argparse.ArgumentParser()
+p.add_argument('--url', default='http://127.0.0.1:8000')
+p.add_argument('--key-file', type=Path, default=Path(__file__).resolve().parents[1] / 'secrets/api-key')
+p.add_argument('--output', type=Path, required=True)
+args = p.parse_args()
+key = args.key_file.read_text().strip()
+
+def request(path, body=None):
+    payload = None if body is None else json.dumps(body, ensure_ascii=False).encode()
+    req = urllib.request.Request(args.url + path, data=payload,
+        headers={'Authorization': 'Bearer ' + key, 'Content-Type': 'application/json'})
+    return urllib.request.urlopen(req, timeout=600)
+
+args.output.parent.mkdir(parents=True, exist_ok=True)
+with args.output.open('x') as out:
+    def record(data):
+        line = json.dumps(data, ensure_ascii=False)
+        out.write(line + '\n'); out.flush(); print(line, flush=True)
+
+    record({'models': json.load(request('/v1/models'))})
+    for name, text, expected in [
+        ('math', 'Compute 17 * 19. Reply with only the integer.', '323'),
+        ('chinese', '请只回复：模型已就绪', '模型已就绪'),
+    ]:
+        body = dict(model='qwen3.8-flash-next', messages=[dict(role='user', content=text)],
+                    temperature=0, reasoning_effort='low', max_tokens=1024,
+                    stream=True, stream_options={'include_usage': True})
+        start = time.monotonic(); first = None; content = ''; reasoning = ''; usage = {}; finish = None
+        with request('/v1/chat/completions', body) as response:
+            for raw in response:
+                if not raw.startswith(b'data: '): continue
+                data = raw[6:].strip()
+                if data == b'[DONE]': break
+                item = json.loads(data)
+                if item.get('usage'): usage = item['usage']
+                for choice in item.get('choices', []):
+                    delta = choice.get('delta', {})
+                    c = delta.get('content') or ''
+                    r = delta.get('reasoning') or delta.get('reasoning_content') or ''
+                    if (c or r) and first is None: first = time.monotonic()
+                    content += c; reasoning += r
+                    finish = choice.get('finish_reason') or finish
+        elapsed = time.monotonic() - start
+        passed = content.strip() == expected and finish == 'stop'
+        record(dict(test=name, passed=passed, content=content, reasoning_chars=len(reasoning),
+                    ttft_s=None if first is None else first-start, elapsed_s=elapsed,
+                    finish_reason=finish, usage=usage))
+        if not passed: raise SystemExit('Acceptance failure: ' + name)
+    body = dict(model='qwen3.8-flash-next', temperature=0, reasoning_effort='low', max_tokens=1024,
+        messages=[dict(role='user',content='Use get_weather to check the weather in Shanghai.')],
+        tools=[dict(type='function',function=dict(name='get_weather',description='Get current weather',
+            parameters=dict(type='object',properties={'city':{'type':'string'}},required=['city'])))],
+        tool_choice='auto')
+    result = json.load(request('/v1/chat/completions',body))
+    calls = result['choices'][0]['message'].get('tool_calls') or []
+    passed = bool(calls) and calls[0]['function']['name']=='get_weather'
+    record(dict(test='tool',passed=passed,response=result))
+    if not passed: raise SystemExit('Tool acceptance failure')