add huggingface space mode

deepseek_vl2/serve/inference.py

@@ -19,6 +19,7 @@
 
 from threading import Thread
 from typing import List
+import spaces
 
 import torch
 import transformers
@@ -40,7 +41,9 @@ def load_model(model_path, dtype=torch.bfloat16):
     vl_gpt: DeepseekVLV2ForCausalLM = AutoModelForCausalLM.from_pretrained(
         model_path, trust_remote_code=True, torch_dtype=dtype
     )
-    vl_gpt = vl_gpt.cuda().eval()
+    if torch.cuda.is_available():
+        vl_gpt = vl_gpt.cuda()
+
     return tokenizer, vl_gpt, vl_chat_processor
 
 
@@ -87,6 +90,7 @@ class StoppingCriteriaSub(StoppingCriteria):
 
 
 @torch.inference_mode()
+@spaces.GPU(duration=120)
 def deepseek_generate(
     conversations: list,
     vl_gpt: torch.nn.Module,