Files
Jony ba920f2b66 Publish current ComfyUI O1Key code baseline
Replace the prior release tree with the current plugin, frontend, tests, and documentation. Document retired node IDs and the public Gitea update source.
2026-09-24 19:56:48 +08:00

629 lines
26 KiB
Python
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
"""
提示词专家节点
ComfyUI 自定义节点,通过 OpenAI 兼容协议调用市面上主流的 AI 对话大模型
支持多模态(图片输入),单轮对话,非流式输出
API 密钥和地址通过插件统一配置(环境变量或 .config 文件),与 Google Gemini 节点一致
"""
import os
import time
import base64
import json
from io import BytesIO
from typing import Optional, Tuple, List
import torch
from PIL import Image
from comfy_api.latest import io
from ..utils.image_utils import tensor_to_pil
from ..utils.config import get_api_key_or_raise, get_base_url_by_route
from ..utils.file_types import FileList
# ============================================================================
# 模型配置
# ============================================================================
DEFAULT_MODEL = "gpt-6-sol"
SUPPORTED_MODELS = [
DEFAULT_MODEL,
"gpt-6-astra",
"gpt-5.6-sol",
"gpt-5.5",
"gemini-3.1-pro-preview",
"deepseek-v4-pro",
"claude-opus-5",
"doubao-seed-2.0-pro",
]
MAX_IMAGE_INPUTS = 9
REASONING_DEPTH_OPTIONS = ["低", "中", "高"]
REASONING_DEPTH_VALUE_MAP = {
"低": "low",
"中": "medium",
"高": "high",
"low": "low",
"medium": "medium",
"high": "high",
}
# 节点内实时 token 预览开关。设为 True 即可恢复原打字机效果。
ENABLE_NODE_TYPEWRITER_PREVIEW = False
# 图片缩放最大尺寸
MAX_IMAGE_DIMENSION = 1568
# 图片最大文件大小(20MB
MAX_IMAGE_SIZE = 20 * 1024 * 1024
def _collect_autogrow_inputs(value) -> list:
"""收集已连接的 Autogrow 输入,并兼容单个旧值。"""
if value is None:
return []
if isinstance(value, dict):
return [item for item in value.values() if item is not None]
return [value]
class UniversalLLMChat(io.ComfyNode):
"""
提示词专家
功能:
- 通过 OpenAI 兼容协议调用主流大模型
- 支持多模态(图片输入)
- 单轮对话,非流式输出
- API 密钥和地址继承插件统一配置
"""
@classmethod
def define_schema(cls):
image_inputs = io.Autogrow.Input(
"图片组",
template=io.Autogrow.TemplateNames(
input=io.Image.Input("图片"),
names=[f"图片{i}" for i in range(1, MAX_IMAGE_INPUTS + 1)],
min=0,
),
tooltip=f"连接后自动增加输入端口,合计最多 {MAX_IMAGE_INPUTS} 张图片。",
)
return io.Schema(
node_id="UniversalLLMChat",
display_name="提示词专家",
category="text/generation",
inputs=[
io.Combo.Input(
"模型",
options=SUPPORTED_MODELS,
default=DEFAULT_MODEL,
),
io.Combo.Input(
"思考深度",
options=REASONING_DEPTH_OPTIONS,
default="高",
),
io.Int.Input(
"seed",
default=0,
min=0,
max=2**31 - 1,
step=1,
display_mode=io.NumberDisplay.number,
control_after_generate=io.ControlAfterGenerate.randomize,
),
io.String.Input(
"api(可选)",
default="",
multiline=False,
placeholder="留空则使用默认 API Key",
),
io.String.Input(
"提示词",
default="",
multiline=True,
),
io.Video.Input("视频", optional=True),
io.Custom("FILE_LIST").Input("文件", optional=True),
image_inputs,
],
outputs=[
io.String.Output(display_name="回复"),
],
hidden=[io.Hidden.unique_id],
is_output_node=True,
# 接收旧版固定图片端口及旧令牌字段,避免旧工作流直接失效。
accept_all_inputs=True,
)
def _resize_image(self, img: Image.Image) -> Image.Image:
"""如果图片过长边超过限制,等比缩放"""
w, h = img.size
max_dim = max(w, h)
if max_dim > MAX_IMAGE_DIMENSION:
scale = MAX_IMAGE_DIMENSION / max_dim
new_w, new_h = int(w * scale), int(h * scale)
print(f"提示词专家: 图片缩放 {w}x{h} -> {new_w}x{new_h}")
return img.resize((new_w, new_h), Image.Resampling.LANCZOS)
return img
def _image_to_data_url(self, img: Image.Image) -> str:
"""将 PIL Image 转为 data URLJPEG base64"""
img = self._resize_image(img)
if img.mode in ('RGBA', 'P'):
img = img.convert('RGB')
for quality in [92, 82, 72, 60, 45]:
buf = BytesIO()
img.save(buf, format='JPEG', quality=quality, optimize=True)
data = buf.getvalue()
if len(data) <= MAX_IMAGE_SIZE:
b64 = base64.b64encode(data).decode('utf-8')
return f"data:image/jpeg;base64,{b64}"
b64 = base64.b64encode(data).decode('utf-8')
return f"data:image/jpeg;base64,{b64}"
# 文件大小限制
MAX_FILE_SIZE = 50 * 1024 * 1024 # 单文件 50MB
MAX_TOTAL_FILE_SIZE = 50 * 1024 * 1024 # 所有文件总计 50MB
# 常见 MIME 类型映射
MIME_MAP = {
".pdf": "application/pdf",
".txt": "text/plain",
".md": "text/markdown",
".csv": "text/csv",
".json": "application/json",
".py": "text/x-python",
".js": "text/javascript",
".html": "text/html",
".xml": "application/xml",
".docx": "application/vnd.openxmlformats-officedocument.wordprocessingml.document",
".xlsx": "application/vnd.openxmlformats-officedocument.spreadsheetml.sheet",
".pptx": "application/vnd.openxmlformats-officedocument.presentationml.presentation",
".zip": "application/zip",
}
# 纯文本类型,直接读取内容
TEXT_EXTS = {".txt", ".md", ".csv", ".json", ".py", ".js", ".ts", ".html",
".xml", ".yaml", ".yml", ".toml", ".ini", ".cfg", ".log",
".sh", ".bat", ".sql", ".css", ".scss", ".jsx", ".tsx"}
def _load_files(self, file_paths_str: str) -> List[dict]:
"""读取文件列表,返回 content part 数组"""
if not file_paths_str or not file_paths_str.strip():
return []
paths = [p.strip() for p in file_paths_str.split(",") if p.strip()]
parts = []
total_size = 0
for path in paths:
if not os.path.isfile(path):
raise ValueError(f"文件不存在: {path}")
file_size = os.path.getsize(path)
if file_size > self.MAX_FILE_SIZE:
raise ValueError(f"文件 {os.path.basename(path)} 大小 {file_size / 1024 / 1024:.1f}MB 超过单文件 50MB 限制")
total_size += file_size
if total_size > self.MAX_TOTAL_FILE_SIZE:
raise ValueError(f"所有文件总大小超过 50MB 限制")
ext = os.path.splitext(path)[1].lower()
mime = self.MIME_MAP.get(ext, "application/octet-stream")
filename = os.path.basename(path)
if ext in self.TEXT_EXTS:
# 文本文件直接读取内容
with open(path, "r", encoding="utf-8", errors="replace") as f:
text_content = f.read()
parts.append({
"type": "text",
"text": f"[文件: {filename}]\n```\n{text_content}\n```",
})
else:
# 二进制文件转 base64,使用 file 格式(OpenAI 兼容协议)
with open(path, "rb") as f:
file_data = base64.b64encode(f.read()).decode("utf-8")
parts.append({
"type": "file",
"file": {
"filename": filename,
"file_data": f"data:{mime};base64,{file_data}",
},
})
print(f"提示词专家: 加载文件 {filename} ({file_size / 1024:.1f}KB, {mime})")
return parts
def _build_input(
self,
prompt: str,
image_tensors: Optional[List[torch.Tensor]] = None,
file_paths: str = "",
file_list: Optional[FileList] = None,
video=None,
) -> list:
"""构建 chat/completions 格式的 messages 数组"""
image_data_urls = []
pil_images_cache = [] # 保留 PIL Image 用于总体积重新编码
if image_tensors:
for tensor in image_tensors:
if tensor is None:
continue
for img in tensor_to_pil(tensor):
img_resized = self._resize_image(img)
if img_resized.mode in ('RGBA', 'P'):
img_resized = img_resized.convert('RGB')
pil_images_cache.append(img_resized)
image_data_urls.append(self._image_to_data_url(img_resized))
# 多图总体积控制
if pil_images_cache and len(pil_images_cache) > 1:
total_bytes = sum(
len(base64.b64decode(url.split(',', 1)[1])) for url in image_data_urls
)
if total_bytes > MAX_IMAGE_SIZE:
print(f"提示词专家: 图片总体积 {total_bytes / 1024 / 1024:.2f}MB 超过 {MAX_IMAGE_SIZE // 1024 // 1024}MB 限制,正在压缩...")
# 降质量
compressed = False
for quality in [80, 70, 60, 50, 40, 30, 20]:
new_urls = []
for img in pil_images_cache:
buf = BytesIO()
img.save(buf, format='JPEG', quality=quality, optimize=True)
b64 = base64.b64encode(buf.getvalue()).decode('utf-8')
new_urls.append(f"data:image/jpeg;base64,{b64}")
total_bytes = sum(len(base64.b64decode(u.split(',', 1)[1])) for u in new_urls)
if total_bytes <= MAX_IMAGE_SIZE:
image_data_urls = new_urls
print(f"提示词专家: 图片压缩完成,总体积 {total_bytes / 1024 / 1024:.2f}MB ({len(pil_images_cache)}张图片,质量{quality})")
compressed = True
break
# 降分辨率
if not compressed:
for scale in [0.75, 0.5, 0.35]:
new_urls = []
for img in pil_images_cache:
w, h = img.size
resized = img.resize((int(w * scale), int(h * scale)), Image.Resampling.LANCZOS)
buf = BytesIO()
resized.save(buf, format='JPEG', quality=20, optimize=True)
b64 = base64.b64encode(buf.getvalue()).decode('utf-8')
new_urls.append(f"data:image/jpeg;base64,{b64}")
total_bytes = sum(len(base64.b64decode(u.split(',', 1)[1])) for u in new_urls)
if total_bytes <= MAX_IMAGE_SIZE:
image_data_urls = new_urls
print(f"提示词专家: 图片压缩完成,总体积 {total_bytes / 1024 / 1024:.2f}MB ({len(pil_images_cache)}张图片,缩放{int(scale*100)}%)")
compressed = True
break
if not compressed:
print(f"提示词专家: 无法将 {len(pil_images_cache)} 张图片压缩到 {MAX_IMAGE_SIZE // 1024 // 1024}MB 以内,请减少图片数量或降低分辨率")
raise ValueError(f"图片总体积 {total_bytes / 1024 / 1024:.2f}MB 超过限制,无法压缩到 {MAX_IMAGE_SIZE // 1024 // 1024}MB 以内")
# 处理视频输入(ComfyUI VIDEO 类型)
video_url_str = ""
if video is not None:
# 从 VIDEO 对象中提取文件路径
vp = None
if isinstance(video, dict):
vp = video.get("video") or video.get("path") or video.get("file") or video.get("filename")
if not vp:
for val in video.values():
if isinstance(val, str) and os.path.exists(val):
vp = val
break
elif isinstance(video, str):
vp = video
else:
for attr in ("video", "path", "filename"):
if hasattr(video, attr):
vp = getattr(video, attr)
break
if not vp and hasattr(video, "__dict__"):
for attr_val in video.__dict__.values():
if isinstance(attr_val, str) and os.path.isfile(attr_val):
vp = attr_val
break
if not vp or not os.path.isfile(vp):
raise ValueError(f"视频文件不存在或路径无效: {vp}")
mime_map = {
".mp4": "video/mp4", ".mpeg": "video/mpeg", ".mpg": "video/mpg",
".mov": "video/quicktime", ".avi": "video/x-msvideo",
".flv": "video/x-flv", ".webm": "video/webm",
".wmv": "video/x-ms-wmv", ".mkv": "video/x-matroska",
}
ext = os.path.splitext(vp)[1].lower()
mime = mime_map.get(ext, "video/mp4")
file_size = os.path.getsize(vp)
print(f"提示词专家: 加载视频 {os.path.basename(vp)} ({file_size / 1024 / 1024:.1f}MB, {mime})")
with open(vp, "rb") as f:
b64 = base64.b64encode(f.read()).decode("utf-8")
video_url_str = f"data:{mime};base64,{b64}"
# 加载文件:优先使用 FILE_LIST,其次使用字符串路径
file_parts = []
if file_list:
for fd in file_list:
print(f"提示词专家: 使用文件 {fd.filename}{fd.extension} ({fd.size / 1024:.1f}KB)")
file_parts.append({
"type": "file",
"file": {
"filename": fd.filename + fd.extension,
"file_data": f"data:{fd.mime_type};base64,{fd.data}",
},
})
elif file_paths:
file_parts = self._load_files(file_paths)
# 纯文本,无图片无文件无视频
if not image_data_urls and not file_parts and not video_url_str:
return [{"role": "user", "content": prompt}]
content_parts = []
# 图片
for url in image_data_urls:
content_parts.append({
"type": "image_url",
"image_url": {"url": url},
})
# 视频:用 image_url 类型传 data URLGemini OpenAI 兼容层支持此格式)
# 同时保留 video_url 类型作为备用(其他支持 video_url 的模型)
if video_url_str:
content_parts.append({
"type": "image_url",
"image_url": {"url": video_url_str},
})
# 文件
for fp in file_parts:
content_parts.append(fp)
content_parts.append({
"type": "text",
"text": prompt,
})
return [{"role": "user", "content": content_parts}]
@staticmethod
def _send_stream_token(node_id, token, done=False):
"""通过 PromptServer 向前端推送流式 token"""
try:
from server import PromptServer
PromptServer.instance.send_sync(
"o1key.stream_token",
{"node_id": str(node_id), "token": token, "done": done},
)
except Exception:
pass
@classmethod
def execute(
cls,
模型: str,
思考深度: str = "高",
seed: int = 0,
提示词: str = "",
视频=None,
文件: Optional[FileList] = None,
**kwargs,
) -> io.NodeOutput:
worker = cls()
result = worker.generate(
模型=模型,
思考深度=思考深度,
seed=seed,
提示词=提示词,
视频=视频,
文件=文件,
node_id=str(cls.hidden.unique_id or ""),
**kwargs,
)
return io.NodeOutput(*result)
def generate(
self,
模型: str,
思考深度: str = "高",
seed: int = 0,
提示词: str = "",
视频=None,
文件: Optional[FileList] = None,
node_id: str = "",
**kwargs,
) -> Tuple[str]:
start_time = time.time()
try:
# 用户填写 api 时覆盖默认 API Key;保留旧字段名兼容旧工作流。
api_value = kwargs.get(
"api(可选)",
kwargs.get("分组令牌(可留空)", kwargs.get("令牌", "")),
)
effective_api_key = str(api_value).strip() if api_value else ""
if not effective_api_key:
effective_api_key = get_api_key_or_raise("O1KEY_API_KEY")
base_url = get_base_url_by_route()
reasoning_effort = REASONING_DEPTH_VALUE_MAP.get(思考深度, "medium")
image_tensors = _collect_autogrow_inputs(kwargs.get("图片组"))
if not image_tensors:
# 兼容 Autogrow 改造前的「图片」及「图片1~图片9」端口。
旧图片 = kwargs.get("图片")
if 旧图片 is not None:
image_tensors.append(旧图片)
image_tensors.extend(
kwargs[f"图片{i}"]
for i in range(1, MAX_IMAGE_INPUTS + 1)
if kwargs.get(f"图片{i}") is not None
)
# 构建 input
input_data = self._build_input(提示词, image_tensors, "", 文件, 视频)
img_count = sum(len(tensor_to_pil(t)) for t in image_tensors)
file_count = len(文件) if 文件 else 0
input_desc = "文本"
if img_count: input_desc += f" + {img_count}张图片"
if 视频 is not None: input_desc += " + 视频"
if file_count: input_desc += f" + {file_count}个文件"
print(f"提示词专家: 模型 = {模型}")
print(f"提示词专家: 思考深度 = {思考深度} ({reasoning_effort})")
print(f"提示词专家: seed = {seed}")
print(f"提示词专家: 输入 = {input_desc}")
# 构建请求体(chat/completions 格式)
request_body = {
"model": 模型,
"messages": input_data,
"stream": True,
"reasoning_effort": reasoning_effort,
"seed": seed,
}
# 打印请求体,base64 截断显示
def _truncate_for_log(obj):
if isinstance(obj, dict):
return {k: _truncate_for_log(v) for k, v in obj.items()}
if isinstance(obj, list):
return [_truncate_for_log(i) for i in obj]
if isinstance(obj, str) and obj.startswith("data:"):
return obj[:60] + f"...[{len(obj)}chars]"
return obj
print(f"提示词专家: 请求原始内容 = {json.dumps(_truncate_for_log(request_body), ensure_ascii=False)}")
# 发送请求(在独立线程中运行异步请求,避免与 ComfyUI 事件循环冲突)
import aiohttp
import asyncio
from concurrent.futures import ThreadPoolExecutor
async def _do_request():
headers = {
"Content-Type": "application/json",
"Authorization": f"Bearer {effective_api_key}",
}
url = f"{base_url}/v1/chat/completions"
# 流式接口不设整体 total 上限(否则会掐断高思考深度的长生成),
# 改用连接超时 + 单次读取超时:只要在 sock_read 间隔内有数据返回就不超时。
timeout = aiohttp.ClientTimeout(total=None, sock_connect=30, sock_read=300)
async with aiohttp.ClientSession(timeout=timeout) as session:
async with session.post(url, headers=headers, json=request_body) as resp:
status = resp.status
if status != 200:
body = await resp.text()
print(f"提示词专家: 响应体 = {body}")
try:
err_data = json.loads(body)
err_msg = err_data.get("error", {}).get("message", body[:200])
except Exception:
err_msg = body[:200]
if status == 401:
raise ValueError(f"认证失败:API Key 无效或已过期")
elif status == 403:
raise ValueError(f"无权访问模型 {模型}")
elif status == 429:
raise ValueError(f"请求频率超限,请稍后重试")
elif status == 404:
raise ValueError(f"模型 {模型} 不存在或 API 地址错误")
else:
raise RuntimeError(f"API 错误 ({status}): {err_msg}")
# 流式读取,拼接 delta content
reply_parts = []
response_body_parts = []
async for raw_line in resp.content:
line = raw_line.decode("utf-8").strip()
if not line or not line.startswith("data:"):
continue
data_str = line[len("data:"):].strip()
response_body_parts.append(f"data: {data_str}")
if data_str == "[DONE]":
break
try:
chunk = json.loads(data_str)
except Exception:
continue
stream_error = chunk.get("error")
if stream_error:
response_body = "\n".join(response_body_parts)
print(f"提示词专家: 响应体 = {response_body}")
if isinstance(stream_error, dict):
error_message = stream_error.get("message", "上游服务暂时不可用")
error_type = stream_error.get("type", "upstream_error")
else:
error_message = str(stream_error)
error_type = "upstream_error"
raise RuntimeError(f"上游服务错误 ({error_type}): {error_message}")
choices = chunk.get("choices")
if not choices:
continue
delta = choices[0].get("delta", {})
content = delta.get("content")
if content:
reply_parts.append(content)
if ENABLE_NODE_TYPEWRITER_PREVIEW:
UniversalLLMChat._send_stream_token(node_id, content)
response_body = "\n".join(response_body_parts)
print(f"提示词专家: 响应体 = {response_body}")
if ENABLE_NODE_TYPEWRITER_PREVIEW:
UniversalLLMChat._send_stream_token(node_id, "", done=True)
reply = "".join(reply_parts)
if not reply:
raise RuntimeError("模型未返回有效文本内容")
return reply
def _run_in_thread():
loop = asyncio.new_event_loop()
try:
return loop.run_until_complete(_do_request())
finally:
loop.close()
with ThreadPoolExecutor(max_workers=1) as pool:
reply = pool.submit(_run_in_thread).result()
elapsed = time.time() - start_time
print(f"提示词专家: 生成完成 (耗时: {elapsed:.2f}s)")
if reply:
preview = reply[:100] + "..." if len(reply) > 100 else reply
print(f"提示词专家: 回复预览: {preview}")
return (reply,)
except ValueError as e:
if str(e) == "未授权!":
print("提示词专家: 请联系作者授权后方可使用!")
raise ValueError("未授权!") from None
error_msg = str(e).split('\n')[0]
print(f"提示词专家: ❌ {error_msg}")
raise
except Exception as e:
import asyncio as _asyncio
if isinstance(e, _asyncio.TimeoutError):
error_msg = "请求超时:服务端长时间未返回数据(可能是模型思考过久或网络不稳定),请重试或降低思考深度/图片数量"
else:
error_msg = str(e).split('\n')[0] or f"未知错误({type(e).__name__}"
print(f"提示词专家: ❌ {error_msg}")
raise RuntimeError(error_msg) from None