Update readme

README.md

@@ -46,6 +46,7 @@ widget:
   - role: user
     content: Can you provide ways to eat combinations of bananas and dragonfruits?
 library_name: transformers
+paper: arxiv.org/abs/2503.01743
 ---
 
 ## Model Summary
@@ -407,8 +408,9 @@ model = AutoModelForCausalLM.from_pretrained(
     model_path, 
     device_map="cuda", 
     torch_dtype="auto", 
-    trust_remote_code=True, 
-    attn_implementation='flash_attention_2',
+    trust_remote_code=True,
+    # if you do not Ampere or later GPUs, change attention to "eager"
+    _attn_implementation='flash_attention_2',
 ).cuda()
 
 # Load generation config
@@ -466,6 +468,8 @@ response = processor.batch_decode(
 print(f'>>> Response\n{response}')
 ```
 
+**Notes**: 
+
 ## Responsible AI Considerations
 
 Like other language models, the Phi family of models can potentially behave in ways that are unfair, unreliable, or offensive. Some of the limiting behaviors to be aware of include:   
@@ -561,7 +565,7 @@ Note that by default, the Phi-4-multimodal-instruct model uses flash attention,
 * NVIDIA H100
 
 If you want to run the model on:
-* NVIDIA V100 or earlier generation GPUs: call AutoModelForCausalLM.from_pretrained() with attn_implementation="eager"
+* NVIDIA V100 or earlier generation GPUs: call AutoModelForCausalLM.from_pretrained() with _attn_implementation="eager"
   
 ## License
 The model is licensed under the [MIT license](./LICENSE).