update

app.py

@@ -20,11 +20,10 @@ import tempfile
 import os
 import gc
 from openai import OpenAI
+import devicetorch
 
 # Load Hugging Face token if needed
-hf_token = os.getenv("HF_TOKEN")
-openai_api_key = os.getenv("OPENAI_API_KEY")
-client = OpenAI(api_key=openai_api_key)
+#hf_token = os.getenv("HF_TOKEN")
 system_prompt_t2v_path = "assets/system_prompt_t2v.txt"
 system_prompt_i2v_path = "assets/system_prompt_i2v.txt"
 with open(system_prompt_t2v_path, "r") as f:
@@ -37,7 +36,8 @@ with open(system_prompt_i2v_path, "r") as f:
 model_path = "asset"
 if not os.path.exists(model_path):
     snapshot_download(
-        "Lightricks/LTX-Video", local_dir=model_path, repo_type="model", token=hf_token
+        #"Lightricks/LTX-Video", local_dir=model_path, repo_type="model", token=hf_token
+        "Lightricks/LTX-Video", local_dir=model_path, repo_type="model",
     )
 
 # Global variables to load components
@@ -45,7 +45,8 @@ vae_dir = Path(model_path) / "vae"
 unet_dir = Path(model_path) / "unet"
 scheduler_dir = Path(model_path) / "scheduler"
 
-device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+#device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+device = devictorch.get(torch)
 
 
 def load_vae(vae_dir):
@@ -113,6 +114,8 @@ def enhance_prompt_if_enabled(prompt, enhance_toggle, type="t2v"):
     ]
 
     try:
+        openai_api_key = os.getenv("OPENAI_API_KEY")
+        client = OpenAI(api_key=openai_api_key)
         response = client.chat.completions.create(
             model="gpt-4o-mini",
             messages=messages,
@@ -261,7 +264,8 @@ def generate_video_from_text(
             duration=5,
         )
     finally:
-        torch.cuda.empty_cache()
+        devicetorch.empty_cache(torch)
+#        torch.cuda.empty_cache()
         gc.collect()
 
     output_path = tempfile.mktemp(suffix=".mp4")
@@ -278,7 +282,8 @@ def generate_video_from_text(
     # Explicitly delete tensors and clear cache
     del images
     del video_np
-    torch.cuda.empty_cache()
+    #torch.cuda.empty_cache()
+    devicetorch.empty_cache(torch)
     return output_path
 
 
@@ -366,7 +371,8 @@ def generate_video_from_image(
         )
 
     finally:
-        torch.cuda.empty_cache()
+        #torch.cuda.empty_cache()
+        devicetorch.empty_cache(torch)
         gc.collect()
 
     return output_path

requirements.txt

@@ -12,3 +12,4 @@ ftfy
 gradio
 openai
 gradio_toggle
+numpy==2.1.3