Upload folder using huggingface_hub

app.py

@@ -54,17 +54,18 @@ def apply_delta(base_model_path, target_model_path, delta_path):
 
     print(f"Saving the target model to {target_model_path}")
     base.load_state_dict(target_weights)
-    base.save_pretrained(target_model_path)
-    delta_tokenizer.save_pretrained(target_model_path)
+    # base.save_pretrained(target_model_path)
+    # delta_tokenizer.save_pretrained(target_model_path)
+    return base, delta_tokenizer
 
 
 base_weights = 'decapoda-research/llama-7b-hf'
 target_weights = 'expertllama' # local path
 delta_weights = 'OFA-Sys/expertllama-7b-delta'
-apply_delta(base_weights, target_weights, delta_weights)
+model, tokenizer = apply_delta(base_weights, target_weights, delta_weights)
 
-tokenizer = transformers.LlamaTokenizer.from_pretrained(expertllama_path)
-model = transformers.LlamaForCausalLM.from_pretrained(expertllama_path, torch_dtype=torch.float16, low_cpu_mem_usage=True)
+# tokenizer = transformers.LlamaTokenizer.from_pretrained(expertllama_path)
+# model = transformers.LlamaForCausalLM.from_pretrained(expertllama_path, torch_dtype=torch.float16, low_cpu_mem_usage=True)
 # model.cuda()
 
 with gr.Blocks() as demo:

requirements.txt

@@ -1,2 +1,3 @@
 torch
-transformers
+transformers
+SentencePiece