Upload server.py with huggingface_hub

server.py

@@ -3,27 +3,34 @@ import sys
 import subprocess
 import base64
 import io
-import json
+import traceback
 
-# Install diffusers (latest PyPI supports GLM-Image) + flask for serving
-subprocess.check_call([sys.executable, "-m", "pip", "install", "-q", "--upgrade", "diffusers", "flask", "accelerate", "sentencepiece", "protobuf", "transformers"])
+# Install latest diffusers (supports GLM-Image) + dependencies
+subprocess.check_call([sys.executable, "-m", "pip", "install", "-q", "--upgrade",
+    "diffusers", "flask", "accelerate", "sentencepiece", "protobuf", "transformers", "huggingface_hub"])
 
 import torch
 from flask import Flask, request, jsonify
-from diffusers.pipelines.glm_image import GlmImagePipeline
+
+print(f"[image] torch version: {torch.__version__}", flush=True)
+print(f"[image] CUDA available: {torch.cuda.is_available()}", flush=True)
+if torch.cuda.is_available():
+    print(f"[image] GPU: {torch.cuda.get_device_name(0)}", flush=True)
+    print(f"[image] VRAM: {torch.cuda.get_device_properties(0).total_mem / 1e9:.1f} GB", flush=True)
 
 app = Flask(__name__)
 pipe = None
 
 def load_model():
     global pipe
+    from diffusers.pipelines.glm_image import GlmImagePipeline
     print("[image] Loading GLM-Image pipeline...", flush=True)
     pipe = GlmImagePipeline.from_pretrained(
         "zai-org/GLM-Image",
         torch_dtype=torch.bfloat16,
     )
-    pipe.enable_model_cpu_offload()
-    print("[image] Model loaded with model CPU offload!", flush=True)
+    pipe.to("cuda")
+    print("[image] Model loaded on CUDA!", flush=True)
 
 @app.route("/health", methods=["GET"])
 def health():
@@ -34,20 +41,22 @@ def health():
 @app.route("/", methods=["POST"])
 def generate():
     try:
-        data = request.get_json()
+        data = request.get_json(force=True)
         prompt = data.get("inputs", "")
         params = data.get("parameters", {})
 
         width = int(params.get("width", 1024))
         height = int(params.get("height", 1024))
-        steps = int(params.get("num_inference_steps", 50))
+        steps = int(params.get("num_inference_steps", 30))
         guidance = float(params.get("guidance_scale", 1.5))
 
         # GLM-Image requires dimensions divisible by 32
         width = (width // 32) * 32
         height = (height // 32) * 32
 
-        print(f"[image] Generating: {prompt[:100]}...", flush=True)
+        print(f"[image] Generating: {prompt[:100]}", flush=True)
+        print(f"[image] Params: {width}x{height}, steps={steps}, guidance={guidance}", flush=True)
+
         image = pipe(
             prompt=prompt,
             height=height,
@@ -59,12 +68,13 @@ def generate():
         buf = io.BytesIO()
         image.save(buf, format="PNG")
         img_b64 = base64.b64encode(buf.getvalue()).decode("utf-8")
-        print(f"[image] Done, image size: {len(buf.getvalue())} bytes", flush=True)
+        print(f"[image] Done, size: {len(buf.getvalue())} bytes", flush=True)
 
         return jsonify({"image": img_b64, "format": "png"})
     except Exception as e:
-        print(f"[image] Error: {e}", flush=True)
-        return jsonify({"error": str(e)}), 500
+        tb = traceback.format_exc()
+        print(f"[image] ERROR: {e}\n{tb}", flush=True)
+        return jsonify({"error": str(e), "traceback": tb}), 500
 
 if __name__ == "__main__":
     load_model()