Fix device configuration and hardware requirements

- Update README.md to request a10g-small GPU hardware instead of t4-medium
- Fix inference.yaml to use auto device detection instead of hardcoded cuda
- Disable xformers and flash_attention for CPU compatibility
- Add device auto-detection to inference script
- This should fix the CPU/GPU mismatch causing generation failures

README.md

@@ -6,7 +6,7 @@ colorTo: pink
 sdk: docker
 pinned: false
 license: apache-2.0
-suggested_hardware: t4-medium
+suggested_hardware: a10g-small
 suggested_storage: large
 ---
 
@@ -72,3 +72,4 @@ Apache 2.0 - See LICENSE file for details
 *Powered by OmniAvatar-14B and ElevenLabs TTS*
 
 **Note**: This space requires large storage capacity due to the 14B parameter models. The models are downloaded on first startup and cached for subsequent uses.
+

configs/inference.yaml

@@ -15,16 +15,16 @@ inference:
   duration: 5.0
 
 hardware:
-  device: "cuda"
+  device: "auto"  # Auto-detect GPU/CPU
   mixed_precision: "fp16"
-  enable_xformers: true
-  enable_flash_attention: true
+  enable_xformers: false  # Disable for CPU
+  enable_flash_attention: false  # Disable for CPU
 
 output:
   output_dir: "./outputs"
   format: "mp4"
   codec: "h264"
-  bitrate: "5M"
+  bitrate: "2M"
 
 tea_cache:
   enabled: false

scripts/inference.py

@@ -6,8 +6,22 @@ import sys
 from pathlib import Path
 import logging
 
-logging.basicConfig(level=logging.INFO)
-logger = logging.getLogger(__name__)
+
+
+def get_device(config_device):
+    """Auto-detect available device"""
+    if config_device == "auto":
+        if torch.cuda.is_available():
+            device = "cuda"
+            logger.info("CUDA available, using GPU")
+        else:
+            device = "cpu"
+            logger.info("CUDA not available, using CPU")
+    else:
+        device = config_device
+        logger.info(f"Using configured device: {device}")
+    
+    return device
 
 def parse_args():
     parser = argparse.ArgumentParser(description="OmniAvatar-14B Inference")
@@ -75,3 +89,4 @@ def main():
 
 if __name__ == "__main__":
     main()
+