motif64

m64_e51ffc16a885 / z_image_server.py

119 lines, 4.3 KBraw
1"""OpenAI-compatible /v1/images/generations server for Z-Image (Turbo or Base).23Env vars (identity only — NOT for tuning generation):4  MODEL_DIR              container path to weights      (default /z-image-model)5  PORT                   listen port                    (default 9006)6  SERVED_MODEL_ID        /v1/models id + variant key    (default z-image-turbo-6b)78Generation defaults are documented per-variant in VARIANT_DEFAULTS below. Query9them at runtime via GET /v1/config. Override any value per-request in the10/v1/images/generations body. The response echoes the resolved values under11"params_used".12"""13import base64, io, os, time, torch14from fastapi import FastAPI, HTTPException15from fastapi.responses import JSONResponse16from pydantic import BaseModel, Field17from typing import Optional18import uvicorn1920MODEL_DIR = os.environ.get("MODEL_DIR", "/z-image-model")21PORT = int(os.environ.get("PORT", 9006))22SERVED_MODEL_ID = os.environ.get("SERVED_MODEL_ID", "z-image-turbo-6b")2324# Per-variant generation defaults. Keys = SERVED_MODEL_ID values.25VARIANT_DEFAULTS = {26    "z-image-turbo-6b": {"num_inference_steps": 9,  "guidance_scale": 0.0},  # Turbo MUST be cfg=027    "z-image-6b":       {"num_inference_steps": 50, "guidance_scale": 4.0},28}29DEFAULTS = VARIANT_DEFAULTS.get(SERVED_MODEL_ID, VARIANT_DEFAULTS["z-image-turbo-6b"])3031app = FastAPI()32pipe = None3334def load_model():35    global pipe36    from diffusers import ZImagePipeline37    print(f"Loading {SERVED_MODEL_ID} from {MODEL_DIR}...", flush=True)38    pipe = ZImagePipeline.from_pretrained(39        MODEL_DIR,40        torch_dtype=torch.bfloat16,41        low_cpu_mem_usage=False,42    )43    pipe.to("cuda")44    print("Model ready.", flush=True)4546class ImageRequest(BaseModel):47    prompt: str = Field(..., description="Text prompt")48    n: int = Field(1, description="Number of images to generate")49    size: str = Field("1024x1024", description="WxH, e.g. 1024x1024")50    num_inference_steps: int = Field(51        DEFAULTS["num_inference_steps"],52        description="Diffusion steps. Turbo: ~9. Base: 28-50. More = higher quality, slower.",53    )54    guidance_scale: float = Field(55        DEFAULTS["guidance_scale"],56        description="CFG scale. Turbo MUST be 0.0; Base typically 3.0-5.0.",57    )58    seed: Optional[int] = Field(None, description="Random seed. null = random per call.")59    response_format: str = Field("b64_json", description="Only b64_json supported.")6061def _params_doc():62    """Introspect ImageRequest into a {name: {default, description}} dict."""63    out = {}64    for name, info in ImageRequest.model_fields.items():65        default = info.default66        if default is ...:67            default = None68        out[name] = {"default": default, "description": info.description}69    return out7071@app.get("/health")72def health():73    return {"status": "ok", "model": SERVED_MODEL_ID}7475@app.get("/v1/models")76def models():77    return {"object": "list", "data": [{"id": SERVED_MODEL_ID, "object": "model"}]}7879@app.get("/v1/config")80def config():81    return {"model": SERVED_MODEL_ID, "params": _params_doc()}8283@app.post("/v1/images/generations")84def generate(req: ImageRequest):85    if pipe is None:86        raise HTTPException(503, "Model not loaded")87    w, h = (int(x) for x in req.size.split("x"))88    gen = torch.Generator("cuda").manual_seed(req.seed) if req.seed is not None else None89    t0 = time.time()90    images = pipe(91        prompt=req.prompt,92        width=w, height=h,93        num_inference_steps=req.num_inference_steps,94        guidance_scale=req.guidance_scale,95        num_images_per_prompt=req.n,96        generator=gen,97    ).images98    elapsed = time.time() - t099    print(f"Generated {len(images)} image(s) in {elapsed:.1f}s", flush=True)100    data = []101    for img in images:102        buf = io.BytesIO()103        img.save(buf, format="PNG")104        data.append({"b64_json": base64.b64encode(buf.getvalue()).decode()})105    return JSONResponse({106        "created": int(time.time()),107        "data": data,108        "params_used": {109            "n": req.n,110            "size": req.size,111            "num_inference_steps": req.num_inference_steps,112            "guidance_scale": req.guidance_scale,113            "seed": req.seed,114        },115    })116117if __name__ == "__main__":118    load_model()119    uvicorn.run(app, host="0.0.0.0", port=PORT)