Upload server.py with huggingface_hub

server.py

@@ -3,26 +3,34 @@ import sys
 import subprocess
 import base64
 import io
-import json
+import traceback
 
-# Install diffusers (latest PyPI supports WanPipeline) + flask + imageio for video
-subprocess.check_call([sys.executable, "-m", "pip", "install", "-q", "--upgrade", "diffusers", "flask", "accelerate", "sentencepiece", "protobuf", "imageio[ffmpeg]", "transformers"])
+# Install latest diffusers (supports WanPipeline) + dependencies
+subprocess.check_call([sys.executable, "-m", "pip", "install", "-q", "--upgrade",
+    "diffusers", "flask", "accelerate", "sentencepiece", "protobuf", "imageio[ffmpeg]", "transformers", "huggingface_hub"])
 
 import torch
 from flask import Flask, request, jsonify
-from diffusers import AutoencoderKLWan, WanPipeline
-from diffusers.utils import export_to_video
+
+print(f"[video] torch version: {torch.__version__}", flush=True)
+print(f"[video] CUDA available: {torch.cuda.is_available()}", flush=True)
+if torch.cuda.is_available():
+    print(f"[video] GPU: {torch.cuda.get_device_name(0)}", flush=True)
+    print(f"[video] 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 import AutoencoderKLWan, WanPipeline
     print("[video] Loading Wan2.2-TI2V-5B pipeline...", flush=True)
+
+    # VAE in float32 for quality, model in bfloat16 for speed
     vae = AutoencoderKLWan.from_pretrained(
         "Wan-AI/Wan2.2-TI2V-5B-Diffusers",
         subfolder="vae",
-        torch_dtype=torch.bfloat16,
+        torch_dtype=torch.float32,
     )
     pipe = WanPipeline.from_pretrained(
         "Wan-AI/Wan2.2-TI2V-5B-Diffusers",
@@ -41,19 +49,23 @@ def health():
 @app.route("/", methods=["POST"])
 def generate():
     try:
-        data = request.get_json()
+        from diffusers.utils import export_to_video
+
+        data = request.get_json(force=True)
         prompt = data.get("inputs", "")
         params = data.get("parameters", {})
 
-        num_frames = int(params.get("num_frames", 49))
+        num_frames = int(params.get("num_frames", 25))
         height = int(params.get("height", 480))
         width = int(params.get("width", 832))
-        steps = int(params.get("num_inference_steps", 30))
+        steps = int(params.get("num_inference_steps", 15))
         fps = int(params.get("fps", 24))
         guidance = float(params.get("guidance_scale", 5.0))
         negative = params.get("negative_prompt", "low quality, blurry, distorted")
 
-        print(f"[video] Generating {num_frames} frames: {prompt[:100]}...", flush=True)
+        print(f"[video] Generating {num_frames} frames: {prompt[:100]}", flush=True)
+        print(f"[video] Params: {width}x{height}, steps={steps}, fps={fps}", flush=True)
+
         frames = pipe(
             prompt=prompt,
             negative_prompt=negative,
@@ -68,7 +80,7 @@ def generate():
         export_to_video(frames, buf, fps=fps)
         video_b64 = base64.b64encode(buf.getvalue()).decode("utf-8")
         duration = len(frames) / fps
-        print(f"[video] Done, {len(frames)} frames, {duration:.1f}s", flush=True)
+        print(f"[video] Done: {len(frames)} frames, {duration:.1f}s, {len(buf.getvalue())} bytes", flush=True)
 
         return jsonify({
             "video": video_b64,
@@ -77,8 +89,9 @@ def generate():
             "duration": duration,
         })
     except Exception as e:
-        print(f"[video] Error: {e}", flush=True)
-        return jsonify({"error": str(e)}), 500
+        tb = traceback.format_exc()
+        print(f"[video] ERROR: {e}\n{tb}", flush=True)
+        return jsonify({"error": str(e), "traceback": tb}), 500
 
 if __name__ == "__main__":
     load_model()