123 lines
5.2 KiB
Python
123 lines
5.2 KiB
Python
import json, datetime, textwrap, requests, gradio as gr
|
|
from pathlib import Path
|
|
from collections import deque
|
|
import queue, threading, time
|
|
|
|
# ───────────────────── 基础配置 ─────────────────────
|
|
API_URL = "http://localhost:30000/generate"
|
|
API_KEY = "token-abc123"
|
|
MODEL_PATH = Path("/root/.cradle/Alibaba/Qwen3-30B-A3B-Base")
|
|
|
|
def model_name(path: Path):
|
|
cfg = path / "config.json"
|
|
if cfg.exists():
|
|
data = json.load(cfg.open())
|
|
return data.get("architectures", [None])[0] or data.get("model_type") or path.name
|
|
return path.name
|
|
|
|
MODEL_NAME = model_name(MODEL_PATH)
|
|
now = lambda: datetime.datetime.now().strftime("%H:%M:%S")
|
|
|
|
# ───────────────────── 日志队列 ─────────────────────
|
|
LOG_Q: "queue.Queue[str]" = queue.Queue()
|
|
def log(msg): # 写终端 + 推队列
|
|
print(msg, flush=True)
|
|
LOG_Q.put(msg)
|
|
|
|
def consume_logs(state_txt: str):
|
|
"""供 Interval 调用：把队列里所有新行取出拼接到 state"""
|
|
buf = deque(state_txt.splitlines(), maxlen=400)
|
|
while not LOG_Q.empty():
|
|
buf.append(LOG_Q.get())
|
|
return "\n".join(buf)
|
|
|
|
# ───────────────────── 后端调用 ─────────────────────
|
|
def backend(text, sampling):
|
|
payload = {"model": MODEL_NAME, "text": text, "sampling_params": sampling}
|
|
log(f"\n🟡 [{now()}] payload\n{json.dumps(payload, ensure_ascii=False, indent=2)}")
|
|
try:
|
|
r = requests.post(API_URL,
|
|
headers={"Authorization": f"Bearer {API_KEY}",
|
|
"Content-Type": "application/json"},
|
|
json=payload, timeout=180)
|
|
try:
|
|
data = r.json()
|
|
except Exception:
|
|
data = {}
|
|
fr = data.get("meta_info", {}).get("finish_reason")
|
|
ctok = data.get("meta_info", {}).get("completion_tokens")
|
|
log(f"🟢 [{now()}] HTTP {r.status_code} tokens={ctok} finish={fr}\n"
|
|
f"🟢 resp800={r.text[:800]!r}")
|
|
if r.status_code != 200:
|
|
return f"[HTTP {r.status_code}] {r.text[:300]}"
|
|
return data.get("text", "").strip() or "[⚠ 空]"
|
|
except Exception as e:
|
|
log(f"[❌ 请求异常] {e}")
|
|
return f"[❌ 请求异常] {e}"
|
|
|
|
# ───────────────────── Chat 回调 ─────────────────────
|
|
def chat(
|
|
user, history,
|
|
max_new, temp, top_p, top_k,
|
|
rep_pen, pres_pen, stop_raw,
|
|
log_state
|
|
):
|
|
stop = [s.strip() for s in stop_raw.split(",") if s.strip()] or None
|
|
samp = {
|
|
"max_new_tokens": int(max_new),
|
|
"temperature": temp,
|
|
"top_p": top_p,
|
|
"top_k": int(top_k),
|
|
"repetition_penalty": rep_pen,
|
|
"presence_penalty": pres_pen,
|
|
**({"stop": stop} if stop else {})
|
|
}
|
|
out = backend(user, samp)
|
|
# 返回回答，同时把 log_state 原样带回（不刷新由 Interval 处理）
|
|
return out, log_state
|
|
|
|
# ───────────────────── Gradio UI ─────────────────────
|
|
with gr.Blocks(title="调试界面") as demo:
|
|
gr.Markdown(f"## 💬 调试界面 \n权重 **{MODEL_PATH.name}**")
|
|
|
|
# 采样参数控件
|
|
with gr.Row():
|
|
max_new = gr.Slider(32, 32768, 2048, label="max_new_tokens")
|
|
temp = gr.Slider(0, 1.5, 0.8, step=0.05, label="temperature")
|
|
with gr.Row():
|
|
top_p = gr.Slider(0, 1, 0.95, step=0.01, label="top_p")
|
|
top_k = gr.Slider(0, 200, 50, step=1, label="top_k")
|
|
with gr.Row():
|
|
rep_pen = gr.Slider(0.8, 2, 1.05, step=0.01, label="repetition_penalty")
|
|
pres_pen= gr.Slider(0, 2, 0.0, step=0.05, label="presence_penalty")
|
|
stop_txt = gr.Textbox("", label="stop 序列（逗号分隔）")
|
|
|
|
# dbg_chk = gr.Checkbox(label="📜 显示 Debug 面板", value=True)
|
|
# log_box = gr.Textbox(label="实时日志", lines=20, interactive=False, visible=True)
|
|
# log_state= gr.State("") # 保存全部日志文本
|
|
|
|
# # ────────────── 定时刷新日志 ──────────────
|
|
# timer = gr.Timer(1.0, render=False) # 每秒触发一次
|
|
# timer.tick( # ⬅ 关键是用 .tick
|
|
# fn=consume_logs, # 把队列里的新行合并
|
|
# inputs=[log_box], # 直接取 Textbox 当前内容
|
|
# outputs=[log_box], # 写回同一个 Textbox
|
|
# )
|
|
|
|
# # 显示到 log_box
|
|
# log_state.change(lambda txt: gr.update(value=txt), log_state, log_box)
|
|
|
|
# # Debug 面板可见性切换
|
|
# dbg_chk.change(lambda v: gr.update(visible=v), dbg_chk, log_box)
|
|
|
|
# Chatbot
|
|
chatbot = gr.ChatInterface(
|
|
fn=chat,
|
|
additional_inputs=[max_new, temp, top_p, top_k,
|
|
rep_pen, pres_pen, stop_txt, log_state],
|
|
additional_outputs=[log_state], # ★ 必加
|
|
type="messages"
|
|
)
|
|
|
|
demo.launch(server_name="0.0.0.0", server_port=30001)
|