from __future__ import annotations """ ModelScope SwingDeploy(FC)UI 函数 app.py(仅适配 text-to-image-synthesis)。 用法: - 将本文件内容复制到 FC 的 model_ui_func 里作为 app.py - 环境变量由部署自动注入: - MODEL_ID / MODEL_VERSION / TASK / API_URL / MODEL_BACKEND 说明: - UI 会把请求转发到 API_URL + "/invoke"(pipeline 模式默认); - 兼容服务端返回 image 的两种形式: - base64 字符串(可带 data:image/... 前缀) - HWC 的嵌套数组(uint8) """ import base64 import io import os import json from typing import Any, Dict, List, Optional, Tuple, Union import gradio as gr import numpy as np import requests from PIL import Image MODEL_ID = os.getenv("MODEL_ID", "") or "" MODEL_VERSION = os.getenv("MODEL_VERSION", "") or "" TASK = os.getenv("TASK", "") or "" API_URL = (os.getenv("API_URL", "") or "").rstrip("/") MODEL_BACKEND = os.getenv("MODEL_BACKEND", "pipeline") or "pipeline" # pipeline 默认推理路径为 /invoke;ollama 会是 /v1(这里只做兼容,不建议用 UI 跑 ollama) INFER_PATH = "v1" if MODEL_BACKEND == "ollama" else "invoke" INFER_URL = (API_URL + "/" + INFER_PATH) if API_URL else "" def _unwrap_response(resp_json: Any) -> Dict[str, Any]: """ SwingDeploy 常见返回: - {"Code":200,"Data":{...},...} 也兼容直接返回 dict 的情况。 """ if isinstance(resp_json, dict): if "Code" in resp_json and resp_json.get("Code") != 200: raise gr.Error(f"推理失败:Code={resp_json.get('Code')} Message={resp_json.get('Message')}") data = resp_json.get("Data", None) if isinstance(data, dict): return data # 有些情况下 Data 不是 dict,就当作整体返回 return resp_json raise gr.Error("推理返回不是JSON对象") def _decode_base64_image(s: str) -> Optional[Image.Image]: s = (s or "").strip() if not s: return None if s.startswith("data:image"): parts = s.split(",", 1) if len(parts) == 2: s = parts[1].strip() try: raw = base64.b64decode(s, validate=False) return Image.open(io.BytesIO(raw)).convert("RGBA") except Exception: return None def _decode_array_image(x: Any) -> Optional[Image.Image]: try: a = np.asarray(x) except Exception: return None if a.ndim != 3 or a.shape[-1] not in (3, 4): return None if a.dtype != np.uint8: try: a = a.astype(np.uint8) except Exception: return None mode = "RGBA" if a.shape[-1] == 4 else "RGB" try: return Image.fromarray(a, mode=mode) except Exception: return None def _extract_images(data: Dict[str, Any]) -> List[Image.Image]: """ 兼容 keys: - output_img / output_imgs - 其他上游包装层(递归查找) """ def pick(obj: Any) -> List[Any]: found: List[Any] = [] if isinstance(obj, dict): for k in ("output_imgs", "output_img", "images", "image", "imgs", "img"): if k in obj and obj[k] is not None: found.append(obj[k]) for v in obj.values(): if isinstance(v, (dict, list)): found.extend(pick(v)) elif isinstance(obj, list): for v in obj: if isinstance(v, (dict, list)): found.extend(pick(v)) return found candidates = pick(data) images: List[Image.Image] = [] def push_one(v: Any) -> None: if v is None: return if isinstance(v, Image.Image): images.append(v) return if isinstance(v, str): im = _decode_base64_image(v) if im is not None: images.append(im) return # 数组(HWC) im = _decode_array_image(v) if im is not None: images.append(im) for c in candidates: # output_imgs 可能是 list[img] if isinstance(c, list): # 也可能是单张 HWC 嵌套数组 im_single = _decode_array_image(c) if im_single is not None: images.append(im_single) continue for one in c: push_one(one) else: push_one(c) if not images: raise gr.Error("推理成功但未在返回中找到图片(output_img/output_imgs)") return images def _post_infer(payload: Dict[str, Any]) -> Dict[str, Any]: if not INFER_URL: raise gr.Error("缺少 API_URL 环境变量,无法调用推理接口") try: r = requests.post(INFER_URL, json=payload, timeout=600) except Exception as e: raise gr.Error(f"请求失败:{e!r}") try: j = r.json() except Exception: raise gr.Error(f"返回非JSON,HTTP={r.status_code}") return _unwrap_response(j) def build_demo() -> gr.Blocks: title = "ModelScope × FC:text-to-image-synthesis" article = ( f"- 模型ID:{MODEL_ID}\n" f"- 模型版本:{MODEL_VERSION}\n" f"- TASK:{TASK}\n" f"- 推理URL:{INFER_URL}\n" ) with gr.Blocks(title=title) as demo: gr.Markdown(f"## {title}") gr.Markdown(article) if TASK != "text-to-image-synthesis": gr.Markdown("### 当前 UI 仅支持 TASK=text-to-image-synthesis") return demo with gr.Row(): prompt = gr.Textbox(label="Prompt", value="stone block texture", lines=2) def build_seed_component(): # 兼容旧版 gradio:可能没有 gr.Number if hasattr(gr, "Number"): return gr.Number(label="seed(-1表示不传)", value=-1, precision=0) return gr.Textbox(label="seed(-1表示不传)", value="-1") def build_json_component(): # 兼容旧版 gradio:可能没有 gr.JSON if hasattr(gr, "JSON"): return gr.JSON(label="原始输出(Data)") return gr.Textbox(label="原始输出(Data)", lines=12) with gr.Row(): n = gr.Slider(label="n(生成张数)", minimum=1, maximum=8, step=1, value=1) ddim_steps = gr.Slider(label="ddim_steps", minimum=5, maximum=100, step=1, value=50) cfg_scale = gr.Slider(label="cfg_scale(仅 text_cond 生效)", minimum=0.0, maximum=20.0, step=0.5, value=3.0) seed = build_seed_component() run_btn = gr.Button("生成") # 兼容旧版 gradio:Gallery 可能没有 .style() gallery = gr.Gallery(label="结果") raw_json = build_json_component() def handler(p: str, n_v: int, steps_v: int, cfg_v: float, seed_v: Any): payload: Dict[str, Any] = { "input": {"text": p}, "parameters": {"n": int(n_v), "ddim_steps": int(steps_v), "cfg_scale": float(cfg_v)}, } # seed 兼容:Number 会传 float/int,Textbox 会传 str try: seed_int = int(float(seed_v)) except Exception: seed_int = -1 if seed_int >= 0: payload["parameters"]["seed"] = seed_int data = _post_infer(payload) imgs = _extract_images(data) # Gallery 支持 list[PIL.Image] if hasattr(gr, "JSON"): return imgs, data return imgs, json.dumps(data, ensure_ascii=False, indent=2) run_btn.click(handler, inputs=[prompt, n, ddim_steps, cfg_scale, seed], outputs=[gallery, raw_json]) return demo if __name__ == "__main__": demo = build_demo() demo.launch(server_name="0.0.0.0")