358 lines
No EOL
13 KiB
Python
358 lines
No EOL
13 KiB
Python
# バージョン情報: Python 3.12+ / FastAPI 0.115.0 / uvicorn 0.30.0 / httpx 0.28.0
|
||
# [2026-02-14] 3060(12GB)戦術支援：Mattermost連携・文字化け完全対策・全ロジック非短縮版
|
||
import argparse
|
||
import asyncio
|
||
import json
|
||
import os
|
||
import re
|
||
import signal
|
||
import sys
|
||
import threading
|
||
from contextlib import asynccontextmanager
|
||
from datetime import datetime
|
||
|
||
import httpx
|
||
import uvicorn
|
||
from fastapi import FastAPI, Request
|
||
from starlette.responses import StreamingResponse
|
||
|
||
# --- 視認性向上のためのカラー定数 ---
|
||
C_GRAY, C_CYAN, C_GREEN, C_YELLOW, C_RED, C_WHITE, C_RESET = (
|
||
"\033[90m",
|
||
"\033[96m",
|
||
"\033[92m",
|
||
"\033[93m",
|
||
"\033[91m",
|
||
"\033[97m",
|
||
"\033[0m",
|
||
)
|
||
|
||
CONFIG = {
|
||
"remote_port": 11432,
|
||
"url": "http://127.0.0.1:11432",
|
||
"timeout": httpx.Timeout(600.0, connect=10.0),
|
||
"dump_next": False,
|
||
"models_cache": [],
|
||
"loop": None,
|
||
"vram_total": 12.0,
|
||
# --- Mattermost 連携設定 ---
|
||
"webhook_url": "https://mm.ka.sugeee.com/hooks/ctjisw6ugjg85nhddo9ox9zt4c",
|
||
"enable_webhook": True,
|
||
"max_log_len": 4000
|
||
}
|
||
|
||
# 高速抽出用正規表現
|
||
RE_CONTENT = re.compile(r'"(?:content|response)":\s*"(.*?)(?<!\\)"')
|
||
|
||
|
||
def get_ts():
|
||
return f"{C_GRAY}[{datetime.now().strftime('%H:%M:%S')}]{C_RESET}"
|
||
|
||
|
||
def pulse(char, color=C_RESET):
|
||
print(f"{color}{char}{C_RESET}", end="", flush=True)
|
||
|
||
|
||
# --- 🛰️ Mattermost 送信エンジン (文字化け対策済み) ---
|
||
async def post_to_mattermost(model_name, path, req_body, resp_text):
|
||
"""通信完了後にバックグラウンドでMattermostへログを飛ばす。生日本語を保持する。"""
|
||
if not CONFIG["webhook_url"] or not CONFIG["enable_webhook"]:
|
||
return
|
||
try:
|
||
# リクエストのデコード（生日本語を維持）
|
||
try:
|
||
req_obj = json.loads(req_body)
|
||
req_clean = json.dumps(req_obj, indent=2, ensure_ascii=False)
|
||
except:
|
||
req_clean = req_body.decode('utf-8', errors='replace')
|
||
|
||
# レスポンスのカットオフ
|
||
resp_display = (
|
||
resp_text if len(resp_text) < CONFIG["max_log_len"]
|
||
else resp_text[:CONFIG["max_log_len"]] + "\n\n*(長文のため中略...)*"
|
||
)
|
||
|
||
payload = {
|
||
"username": f"oproxy:{model_name}",
|
||
"icon_url": "https://ollama.com/public/ollama.png",
|
||
"text": (
|
||
f"### 🛡️ LLM Traffic Log: `{path}`\n"
|
||
f"- **Model:** `{model_name}` | **Time:** {datetime.now().strftime('%H:%M:%S')}\n"
|
||
f"#### 📥 Request preview\n```json\n{req_clean[:800]}\n```\n"
|
||
f"#### 📤 Response\n{resp_display}"
|
||
)
|
||
}
|
||
async with httpx.AsyncClient(timeout=15.0) as client:
|
||
# ensure_ascii=False で Mattermost に生の日本語を届ける
|
||
encoded_payload = json.dumps(payload, ensure_ascii=False).encode('utf-8')
|
||
await client.post(CONFIG["webhook_url"], content=encoded_payload, headers={"Content-Type": "application/json"})
|
||
except Exception as e:
|
||
print(f"\n{C_RED}[Webhook Fail] {e}{C_RESET}")
|
||
|
||
|
||
# --- 🧠 戦略的モデル分析 (2行分割レイアウト・非短縮版) ---
|
||
async def fetch_detailed_models():
|
||
"""Ollamaの内部情報を深掘りし、メタデータを完全取得する"""
|
||
try:
|
||
async with httpx.AsyncClient(timeout=5.0) as client:
|
||
res = await client.get(f"{CONFIG['url']}/api/tags")
|
||
if res.status_code != 200:
|
||
return
|
||
models = res.json().get("models", [])
|
||
|
||
for m in models:
|
||
try:
|
||
show_res = await client.post(
|
||
f"{CONFIG['url']}/api/show", json={"name": m["name"]}
|
||
)
|
||
if show_res.status_code == 200:
|
||
data = show_res.json()
|
||
templ = data.get("template", "").lower()
|
||
params = data.get("parameters", "")
|
||
|
||
# エージェント/補完タイプの判定
|
||
m_type = "inst"
|
||
if any(x in templ for x in ["tool", "parameters", "call"]):
|
||
m_type = "agent"
|
||
elif any(
|
||
x in m["name"].lower()
|
||
for x in ["acp", "fim", "base", "code"]
|
||
):
|
||
m_type = "acp"
|
||
|
||
# コンテキスト長の抽出
|
||
ctx = "Def"
|
||
if "num_ctx" in params:
|
||
ctx = params.split("num_ctx")[-1].strip().split()[0]
|
||
|
||
m["meta"] = {
|
||
"type": m_type,
|
||
"ctx": ctx,
|
||
"q": m.get("details", {}).get("quantization_level", "?"),
|
||
}
|
||
except:
|
||
m["meta"] = {"type": "inst", "ctx": "???", "q": "?"}
|
||
CONFIG["models_cache"] = models
|
||
except:
|
||
pass
|
||
|
||
|
||
def print_analysis(detail_mode=False):
|
||
"""ガタつきを排した重厚な2行リスト表示"""
|
||
print(
|
||
f"\n{C_CYAN}--- Strategic Agent Analysis (Target: {CONFIG['vram_total']}GB VRAM) ---{C_RESET}"
|
||
)
|
||
print(
|
||
f"{C_GRAY}{'STATUS':<7} | {'MODEL IDENTIFIER':<45} | {'TYPE':<7} | {'CTX':<7} | {'SIZE'}{C_RESET}"
|
||
)
|
||
print(f"{C_GRAY}{'-' * 88}{C_RESET}")
|
||
|
||
for m in CONFIG["models_cache"]:
|
||
gb = m.get("size", 0) / (1024**3)
|
||
meta = m.get("meta", {"type": "inst", "ctx": "???", "q": "?"})
|
||
|
||
# 3060(12GB)用 VRAM判定
|
||
if gb <= CONFIG["vram_total"] * 0.75:
|
||
color, status = C_GREEN, "SAFE"
|
||
elif gb <= CONFIG["vram_total"]:
|
||
color, status = C_YELLOW, "LIMIT"
|
||
else:
|
||
color, status = C_RED, "OVER"
|
||
|
||
type_c = (
|
||
C_CYAN
|
||
if meta["type"] == "agent"
|
||
else (C_YELLOW if meta["type"] == "acp" else C_WHITE)
|
||
)
|
||
|
||
full_name = m["name"]
|
||
if "/" in full_name:
|
||
parts = full_name.rsplit("/", 1)
|
||
org_path, model_base = parts[0] + "/", parts[1]
|
||
else:
|
||
org_path, model_base = "", full_name
|
||
|
||
# 1行目: ステータス、組織パス、タイプ、CTX、サイズ
|
||
print(
|
||
f" {color}{status:<6}{C_RESET} | {C_GRAY}{org_path:<45}{C_RESET} | {type_c}{meta['type']:<7}{C_RESET} | {C_WHITE}{str(meta['ctx']):>7}{C_RESET} | {C_WHITE}{gb:>6.2f}GB{C_RESET}"
|
||
)
|
||
# 2行目: モデル本体名 + 量子化詳細 (detail_mode時)
|
||
q_info = f" {C_GRAY}[{meta['q']}]{C_RESET}" if detail_mode else ""
|
||
print(f" {'':<6} | {color}{model_base:<45}{C_RESET} | {q_info}")
|
||
|
||
print(f"{C_CYAN}{'=' * 88}{C_RESET}")
|
||
|
||
|
||
# --- 📊 VRAM リアルタイムステータス ---
|
||
def show_vram_status():
|
||
async def _fetch():
|
||
try:
|
||
async with httpx.AsyncClient(timeout=2.0) as client:
|
||
res = await client.get(f"{CONFIG['url']}/api/ps")
|
||
if res.status_code == 200:
|
||
models = res.json().get("models", [])
|
||
print(f"\n{get_ts()} {C_CYAN}--- Active VRAM Usage ---{C_RESET}")
|
||
if not models:
|
||
print(
|
||
f"{get_ts()} {C_GRAY}No models currently in VRAM.{C_RESET}"
|
||
)
|
||
for m in models:
|
||
v_gb, t_gb = (
|
||
m.get("size_vram", 0) / (1024**3),
|
||
m.get("size", 0) / (1024**3),
|
||
)
|
||
v_color = (
|
||
C_RED if v_gb > 11 else (C_YELLOW if v_gb > 8 else C_GREEN)
|
||
)
|
||
print(
|
||
f"{get_ts()} {C_WHITE}{m['name']:<25}{C_RESET} {v_color}{v_gb:.2f}{C_RESET} / {t_gb:.2f} GB"
|
||
)
|
||
except:
|
||
print(f"\n{C_RED}[Error] Could not fetch VRAM status.{C_RESET}")
|
||
|
||
if CONFIG["loop"]:
|
||
asyncio.run_coroutine_threadsafe(_fetch(), CONFIG["loop"])
|
||
|
||
|
||
# --- 🚀 Proxy Core (高速リアルタイムDUMP実装) ---
|
||
@asynccontextmanager
|
||
async def lifespan(app: FastAPI):
|
||
CONFIG["loop"] = asyncio.get_running_loop()
|
||
asyncio.create_task(fetch_detailed_models())
|
||
yield
|
||
|
||
|
||
app = FastAPI(lifespan=lifespan)
|
||
|
||
|
||
@app.api_route("/{path:path}", methods=["GET", "POST", "PUT", "DELETE"])
|
||
async def sticky_proxy(path: str, request: Request):
|
||
do_dump = CONFIG["dump_next"]
|
||
if do_dump:
|
||
CONFIG["dump_next"] = False
|
||
|
||
print(f"{get_ts()} {C_WHITE}/{path: <10}{C_RESET} ", end="", flush=True)
|
||
body = await request.body()
|
||
|
||
# 送信時にどのモデルか判別するための抽出
|
||
model_name = "unknown"
|
||
try:
|
||
model_name = json.loads(body).get("model", "unknown")
|
||
except:
|
||
pass
|
||
|
||
if do_dump:
|
||
print(
|
||
f"\n{C_YELLOW}{'=' * 60}\n[DUMP REQUEST: {model_name}]\n{'=' * 60}{C_RESET}"
|
||
)
|
||
try:
|
||
print(json.dumps(json.loads(body), indent=2, ensure_ascii=False))
|
||
except:
|
||
print(f"{C_GRAY}Body (Raw): {body[:500]!r}{C_RESET}")
|
||
print(f"\n{C_GREEN}[REALTIME AI RESPONSE CONTENT]{C_RESET}")
|
||
|
||
# 送信パルス
|
||
for _ in range(max(1, min(len(body) // 512, 12))):
|
||
pulse("^", C_CYAN)
|
||
pulse("|", C_YELLOW)
|
||
|
||
async def stream_response():
|
||
async with httpx.AsyncClient(timeout=CONFIG["timeout"]) as client:
|
||
try:
|
||
headers = {
|
||
k: v
|
||
for k, v in request.headers.items()
|
||
if k.lower() not in ["host", "content-length", "connection"]
|
||
}
|
||
async with client.stream(
|
||
request.method,
|
||
f"{CONFIG['url']}/{path}",
|
||
content=body,
|
||
headers=headers,
|
||
) as response:
|
||
pulse("v", C_GREEN)
|
||
|
||
full_text_buffer = []
|
||
async for chunk in response.aiter_bytes():
|
||
# チャンクから正規表現で高速テキスト抽出
|
||
raw_data = chunk.decode(errors="ignore")
|
||
matches = RE_CONTENT.findall(raw_data)
|
||
for m in matches:
|
||
try:
|
||
# 【最重要】Unicodeエスケープを戻し、化けないようにUTF-8で再構成
|
||
text = m.encode('utf-8').decode('unicode_escape').encode('latin1').decode('utf-8')
|
||
full_text_buffer.append(text)
|
||
if do_dump:
|
||
print(f"{C_WHITE}{text}{C_RESET}", end="", flush=True)
|
||
except:
|
||
# 万が一デコードに失敗した場合はマッチした文字列をそのまま保持
|
||
full_text_buffer.append(m)
|
||
|
||
pulse("v", C_GREEN)
|
||
yield chunk
|
||
|
||
# 通信完了後に非同期タスクとしてWebhook送信
|
||
if full_text_buffer:
|
||
combined_text = "".join(full_text_buffer)
|
||
asyncio.create_task(post_to_mattermost(
|
||
model_name, path, body, combined_text
|
||
))
|
||
|
||
if do_dump:
|
||
print(f"\n{C_GREEN}{'=' * 60}{C_RESET}")
|
||
pulse("*", C_YELLOW)
|
||
except Exception as e:
|
||
print(f" {C_RED}[Proxy Error] {e}{C_RESET}")
|
||
finally:
|
||
print("", flush=True)
|
||
|
||
return StreamingResponse(stream_response())
|
||
|
||
|
||
# --- ⌨️ コマンド入力ハンドラ ---
|
||
def input_handler():
|
||
print(
|
||
f"\n{C_GREEN}oproxy: Full Tactical Suite (RTX 3060 12GB Edition) Active{C_RESET}"
|
||
)
|
||
print(
|
||
f"{C_GRAY}Commands: [d]Dump [s]VRAM [l]List [ll]Detail [digit]Port [q]Exit{C_RESET}\n"
|
||
)
|
||
while True:
|
||
try:
|
||
line = sys.stdin.readline().strip().lower()
|
||
if not line:
|
||
continue
|
||
if line == "q":
|
||
os.kill(os.getpid(), signal.SIGINT)
|
||
elif line in ["d", "dd"]:
|
||
CONFIG["dump_next"] = not CONFIG["dump_next"]
|
||
print(
|
||
f"{get_ts()} DUMP Reservation: {'ON' if CONFIG['dump_next'] else 'OFF'}"
|
||
)
|
||
elif line == "s":
|
||
show_vram_status()
|
||
elif line in ["l", "ll"]:
|
||
asyncio.run_coroutine_threadsafe(
|
||
fetch_detailed_models(), CONFIG["loop"]
|
||
)
|
||
print_analysis(detail_mode=(line == "ll"))
|
||
elif line.isdigit():
|
||
p = int(line)
|
||
CONFIG["remote_port"], CONFIG["url"] = p, f"http://127.0.0.1:{p}"
|
||
print(f"{get_ts()} Target Port Switched -> {p}")
|
||
except:
|
||
pass
|
||
|
||
|
||
if __name__ == "__main__":
|
||
parser = argparse.ArgumentParser()
|
||
parser.add_argument("-r", "--remote", type=int, default=11432)
|
||
parser.add_argument("-l", "--local", type=int, default=11434)
|
||
args = parser.parse_args()
|
||
CONFIG.update(
|
||
{"remote_port": args.remote, "url": f"http://127.0.0.1:{args.remote}"}
|
||
)
|
||
|
||
# メインスレッドでUvicorn、サブスレッドで入力待ち
|
||
threading.Thread(target=input_handler, daemon=True).start()
|
||
uvicorn.run(app, host="127.0.0.1", port=args.local, log_level="error") |