from __future__ import annotations """ Hugging Face Spaces 入口(Gradio)。 目标: - 复用本项目现有“加载 checkpoint -> DDPM/DDIM 采样 -> 输出图片”的逻辑; - 增加“伪私有”鉴权:所有推理请求必须提供正确的 API_KEY,否则拒绝; - 支持从 Hugging Face Hub 下载 checkpoint(便于 Spaces 部署,不必把大文件提交到 Space 仓库)。 重要说明(伪私有的边界): - Space 本身仍然是公开的(你选择 public 的情况下),只是把推理能力用 API_KEY 做访问控制; - 拿到 API_KEY 的人仍可调用(这是预期)。 """ import hmac import os from typing import Optional, Tuple, List import torch import gradio as gr from huggingface_hub import hf_hub_download from PIL import Image # 复用现有逻辑(不重复造轮子) from webui import _load_cached, _tensor_to_pil # type: ignore from blockdiffusion.train.guidance import CFGuidanceModel from blockdiffusion.text.simple_tokenizer import tokenize from blockdiffusion.utils.seed import seed_everything, get_device def _get_env(name: str, default: str = "") -> str: v = os.getenv(name, default) return (v or "").strip() def _require_space_api_key() -> str: """ Space 的“伪私有”Key:建议在 Hugging Face Space Settings -> Secrets 里配置。 """ # 注意:Hugging Face 网页端对 Secret 名称有时会限制为字母数字且字母开头, # 导致带下划线的 `SPACE_API_KEY` 无法创建;因此这里兼容 `SPACEAPIKEY`。 key = _get_env("SPACE_API_KEY", "") or _get_env("SPACEAPIKEY", "") if not key: # 失败更安全:避免误部署成“无鉴权公开推理” raise RuntimeError("缺少环境变量 SPACE_API_KEY/SPACEAPIKEY(请在 Space Secrets 中配置)") return key _RESOLVED_CKPT_PATH: Optional[str] = None def _resolve_ckpt_path() -> str: """ 获取 checkpoint 路径: - 优先使用本地路径 BD_CKPT_PATH - 否则从 Hugging Face Hub 下载:BD_CKPT_REPO + BD_CKPT_FILENAME(可选 BD_CKPT_REVISION) 说明: - 若模型仓库是私有/受限,需要在 Space Secrets 设置 HF_TOKEN(或 HUGGINGFACE_HUB_TOKEN) """ global _RESOLVED_CKPT_PATH if _RESOLVED_CKPT_PATH and os.path.isfile(_RESOLVED_CKPT_PATH): return _RESOLVED_CKPT_PATH local_path = _get_env("BD_CKPT_PATH", "") if local_path: if not os.path.isfile(local_path): raise RuntimeError(f"BD_CKPT_PATH 指向的文件不存在:{local_path}") _RESOLVED_CKPT_PATH = local_path return _RESOLVED_CKPT_PATH repo_id = _get_env("BD_CKPT_REPO", "") # 约定优先使用生态常见的权重名,减少歧义 filename = _get_env("BD_CKPT_FILENAME", "pytorch_model.bin") revision = _get_env("BD_CKPT_REVISION", "") or None if not repo_id: raise RuntimeError("未配置 checkpoint:请设置 BD_CKPT_PATH 或 BD_CKPT_REPO/BD_CKPT_FILENAME") token = _get_env("HF_TOKEN", "") or _get_env("HUGGINGFACE_HUB_TOKEN", "") or None try: path = hf_hub_download(repo_id=repo_id, filename=filename, revision=revision, token=token) except Exception as e: raise RuntimeError(f"从Hub下载checkpoint失败:repo={repo_id} file={filename} rev={revision} err={e!r}") from e if not os.path.isfile(path): raise RuntimeError(f"下载完成但文件不存在:{path}") _RESOLVED_CKPT_PATH = path return path def _check_api_key(user_key: str) -> None: """ 伪私有鉴权:对外只暴露“成功/失败”,避免泄漏真实 Key。 """ # 若 Space 端未配置 Key,给出明确错误,避免客户端只能看到“上游异常但无verbose” try: expected = _require_space_api_key() except Exception: raise gr.Error("服务端未配置SPACEAPIKEY/SPACE_API_KEY:请到Space Settings->Secrets 新建后重启") user_key = (user_key or "").strip() if not user_key or not hmac.compare_digest(user_key, expected): raise gr.Error("未授权:API_KEY错误") @torch.no_grad() def _infer( api_key: str, prompt: str, cfg_scale: float, use_ema: bool, use_ddim: bool, ddim_steps: int, ddim_eta: float, n_images: int, batch: int, seed: int, preview_upscale: int, ) -> Tuple[str, Image.Image, List[Image.Image]]: # 1) 鉴权 _check_api_key(api_key) # 2) 参数整理 n_images = int(max(1, n_images)) batch = int(max(1, batch)) preview_upscale = int(max(1, preview_upscale)) seed_everything(int(seed), deterministic=False) # 3) 定位 ckpt + device ckpt_path = _resolve_ckpt_path() # Spaces 上尽量自动选择(有GPU则用GPU) device = get_device(None) device_text = str(device) # 4) 复用 webui 的缓存加载(包含:读取meta、构建模型、EMA覆盖、构建diffusion) loaded = _load_cached(ckpt_path=ckpt_path, device_text=device_text, use_ema=bool(use_ema)) # 5) 条件采样(prompt + CFG) sampler_model: torch.nn.Module = loaded.model prompt = (prompt or "").strip() if prompt: if loaded.model_type != "text_cond": raise gr.Error("该checkpoint为无条件模型,不支持prompt/CFG") if loaded.vocab is None or loaded.text_cfg is None: raise gr.Error("条件模型缺少tokenizer/text_cfg") tok = torch.tensor( [loaded.vocab.encode(prompt, max_len=loaded.text_cfg.max_text_len)], dtype=torch.long, device=device, ) m = tok != loaded.vocab.pad_id cond_kwargs = {"tokens": tok, "mask": m} un_tok = torch.full_like(tok, fill_value=loaded.vocab.pad_id) un_m = torch.zeros_like(m) uncond_kwargs = {"tokens": un_tok, "mask": un_m} sampler_model = CFGuidanceModel( base_model=loaded.model, cond_kwargs=cond_kwargs, uncond_kwargs=uncond_kwargs, scale=float(cfg_scale), ).to(device) sampler_model.eval() # 6) 采样(分批) all_imgs: List[torch.Tensor] = [] remaining = n_images while remaining > 0: b = min(batch, remaining) shape = (b, loaded.channels, loaded.image_size, loaded.image_size) if use_ddim: x = loaded.diffusion.ddim_sample_loop( sampler_model, shape=shape, device=device, steps=int(ddim_steps), eta=float(ddim_eta), model_kwargs=None, progress=False, ) else: x = loaded.diffusion.p_sample_loop( sampler_model, shape=shape, device=device, model_kwargs=None, progress=False, ) x = (x.clamp(-1, 1) + 1) * 0.5 # [0,1] all_imgs.append(x.detach().cpu()) remaining -= b imgs = torch.cat(all_imgs, dim=0) # (N,C,H,W) on CPU # 7) 输出:网格图 + gallery(按像素纹理预览做NEAREST放大) # 为了不重复引入 torchvision/make_grid,这里直接拼接成单行网格(N很小时足够)。 # 如果你需要更复杂网格,可直接复用 webui.py 的 make_grid 逻辑。 gallery: List[Image.Image] = [] for i in range(n_images): im = _tensor_to_pil(imgs[i]) if preview_upscale > 1: w, h = im.size im = im.resize((w * preview_upscale, h * preview_upscale), resample=Image.NEAREST) gallery.append(im) # 简单拼图:横向拼接(避免引入额外依赖/逻辑) total_w = sum(im.size[0] for im in gallery) max_h = max(im.size[1] for im in gallery) grid = Image.new("RGBA", (total_w, max_h)) x0 = 0 for im in gallery: grid.paste(im, (x0, 0)) x0 += im.size[0] # 提示词token诊断:是否大量(仅用于展示) unk_info = "" if prompt and loaded.vocab is not None: toks = tokenize(prompt) unk = [t for t in toks if int(loaded.vocab.token_to_id.get(t, loaded.vocab.unk_id)) == int(loaded.vocab.unk_id)] if unk: unk_info = f" unk_tokens={len(unk)}" info = ( f"ckpt={os.path.basename(ckpt_path)} device={device_text} " f"type={loaded.model_type} size={loaded.image_size} ch={loaded.channels} " f"sampler={'DDIM' if use_ddim else 'DDPM'} n={n_images} batch={batch} seed={int(seed)} " f"prompt={'(none)' if not prompt else prompt[:80]}{unk_info}" ) return info, grid, gallery def build_demo() -> gr.Blocks: """ Gradio UI(同时提供 /api/predict 的在线调用能力)。 """ with gr.Blocks(title="BlockDiffusion (Pseudo-Private)") as demo: gr.Markdown("### BlockDiffusion Spaces(伪私有:需API_KEY)") with gr.Row(): api_key = gr.Textbox(label="API_KEY(必填)", type="password") with gr.Row(): prompt = gr.Textbox(label="Prompt(条件模型可用;留空=无条件)", value="", lines=2) cfg_scale = gr.Slider(minimum=0.0, maximum=12.0, value=3.0, step=0.1, label="CFG scale") with gr.Row(): use_ema = gr.Checkbox(value=True, label="使用EMA权重(推荐)") use_ddim = gr.Checkbox(value=True, label="使用DDIM(更快)") ddim_steps = gr.Slider(minimum=10, maximum=200, value=50, step=1, label="DDIM steps") ddim_eta = gr.Slider(minimum=0.0, maximum=1.0, value=0.0, step=0.01, label="DDIM eta") with gr.Row(): n_images = gr.Slider(minimum=1, maximum=8, value=1, step=1, label="生成张数N(Spaces建议<=8)") batch = gr.Slider(minimum=1, maximum=8, value=1, step=1, label="采样batch(Spaces建议<=8)") seed = gr.Number(value=123, precision=0, label="seed") preview_upscale = gr.Slider(minimum=1, maximum=64, value=16, step=1, label="预览放大倍数(NEAREST)") run = gr.Button("生成") info = gr.Textbox(label="信息", lines=2) grid = gr.Image(label="网格图", type="pil") gallery = gr.Gallery(label="单张结果") def _run(*args): # 把非 gr.Error 的异常也转换成可见错误信息(不暴露堆栈) try: return _infer(*args) except gr.Error: raise except Exception as e: raise gr.Error(f"推理失败:{e}") run.click( _run, inputs=[ api_key, prompt, cfg_scale, use_ema, use_ddim, ddim_steps, ddim_eta, n_images, batch, seed, preview_upscale, ], outputs=[info, grid, gallery], ) return demo demo = build_demo() def main() -> None: """ Hugging Face Spaces 会直接执行 `python app.py`; 如果不显式 launch,进程会立刻退出(Exit code 0)并被判定为 Runtime error。 """ port = int(os.getenv("PORT", "7860")) # 监听 0.0.0.0 才能被 Space 反向代理访问到 demo.launch(server_name="0.0.0.0", server_port=port, share=False, show_error=False) if __name__ == "__main__": main()