Feature(MInference): add local mode

app.py

@@ -1,53 +1,47 @@
 import subprocess
 import os
-# Install flash attention, skipping CUDA build if necessary
-
-# os.environ["CPATH"] = "$CPATH:/usr/local/cuda/include"
-# os.environ["CUDA_HOME"] = "/usr/local/cuda"
-# os.environ["PATH"] = "$PATH:$CUDA_HOME/bin"
-# os.environ["LIBRARY_PATH"] = "$LIBRARY_PATH:/usr/local/cuda/lib64"
-
-subprocess.run(
-    "/home/user/.pyenv/shims/pip install flash-attn --no-build-isolation",
-    env={"FLASH_ATTENTION_SKIP_CUDA_BUILD": "TRUE"},
-    shell=True,
-)
-# subprocess.run(
-#     "/home/user/.pyenv/shims/pip install pycuda==2023.1",
-#     env={"CPATH": "$CPATH:/usr/local/cuda/include", "LIBRARY_PATH": "$LIBRARY_PATH:/usr/local/cuda/lib64"},
-#     shell=True,
-# )
+import torch
 
 import gradio as gr
 import os
 import spaces
-from transformers import AutoModelForCausalLM
 from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
 from threading import Thread
+from transformers.utils.import_utils import _is_package_available
 
 # Set an environment variable
 HF_TOKEN = os.environ.get("HF_TOKEN", None)
 
 
-DESCRIPTION = '''
-<div>
-<h1 style="text-align: center;">Meta Llama3 8B</h1>
-<p>This Space demonstrates the instruction-tuned model <a href="https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct"><b>Meta Llama3 8b Chat</b></a>. Meta Llama3 is the new open LLM and comes in two sizes: 8b and 70b. Feel free to play with it, or duplicate to run privately!</p>
-<p>🔎 For more details about the Llama3 release and how to use the model with <code>transformers</code>, take a look <a href="https://huggingface.co/blog/llama3">at our blog post</a>.</p>
-<p>🦕 Looking for an even more powerful model? Check out the <a href="https://huggingface.co/chat/"><b>Hugging Chat</b></a> integration for Meta Llama 3 70b</p>
-</div>
-'''
+DESCRIPTION = """
+# MInference 1.0: Accelerating Pre-filling for Long-Context LLMs via Dynamic Sparse Attention (Under Review) [[paper](https://arxiv.org/abs/2406.05736)]
+_Huiqiang Jiang†, Yucheng Li†, Chengruidong Zhang†, Qianhui Wu, Xufang Luo, Surin Ahn, Zhenhua Han, Amir H. Abdi, Dongsheng Li, Chin-Yew Lin, Yuqing Yang and Lili Qiu_
+
+<h2 style="text-align: center;"><a href="https://github.com/microsoft/MInference" target="blank"> [Code]</a> 
+<a href="https://hqjiang.com/minference.html" target="blank"> [Project Page]</a>
+<a href="https://arxiv.org/abs/2406.05736" target="blank"> [Paper]</a></h2>
+
+<font color="brown"><b>This is only a deployment demo. Due to limited GPU resources, we do not provide an online demo. You will need to follow the code below to try MInference locally.</b></font>
+
+```bash
+git clone https://huggingface.co/spaces/microsoft/MInference
+cd MInference
+pip install -r requirments.txt
+pip install flash_attn pycuda==2023.1
+python app.py
+```
+<br/>
+"""
 
 LICENSE = """
-<p/>
----
-Built with Meta Llama 3
+<div style="text-align: center;">
+    <p>© 2024 Microsoft</p>
+</div>
 """
 
 PLACEHOLDER = """
 <div style="padding: 30px; text-align: center; display: flex; flex-direction: column; align-items: center;">
-   <img src="https://ysharma-dummy-chat-app.hf.space/file=/tmp/gradio/8e75e61cc9bab22b7ce3dec85ab0e6db1da5d107/Meta_lockup_positive%20primary_RGB.jpg" style="width: 80%; max-width: 550px; height: auto; opacity: 0.55;  "> 
-   <h1 style="font-size: 28px; margin-bottom: 2px; opacity: 0.55;">Meta llama3</h1>
+   <h1 style="font-size: 28px; margin-bottom: 2px; opacity: 0.55;">LLaMA-3-8B-Gradient-1M w/ MInference</h1>
    <p style="font-size: 18px; margin-bottom: 2px; opacity: 0.65;">Ask me anything...</p>
 </div>
 """
@@ -58,34 +52,28 @@ h1 {
   text-align: center;
   display: block;
 }
-#duplicate-button {
-  margin: auto;
-  color: white;
-  background: #1565c0;
-  border-radius: 100vh;
-}
 """
 
 # Load the tokenizer and model
-model_name = "gradientai/Llama-3-8B-Instruct-262k"
+model_name = "gradientai/Llama-3-8B-Instruct-Gradient-1048k"
 tokenizer = AutoTokenizer.from_pretrained(model_name)
-model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", device_map="auto")  # to("cuda:0") 
+model = AutoModelForCausalLM.from_pretrained(
+    model_name, torch_dtype="auto", device_map="auto"
+)  # to("cuda:0")
+
+if torch.cuda.is_available() and _is_package_available("pycuda"):
+    from minference import MInference
+
+    minference_patch = MInference("minference", model_name)
+    model = minference_patch(model)
 
-from minference import MInference
-minference_patch = MInference("minference", model_name)
-model = minference_patch(model)
+terminators = [tokenizer.eos_token_id, tokenizer.convert_tokens_to_ids("<|eot_id|>")]
 
-terminators = [
-    tokenizer.eos_token_id,
-    tokenizer.convert_tokens_to_ids("<|eot_id|>")
-]
 
 @spaces.GPU(duration=120)
-def chat_llama3_8b(message: str, 
-              history: list, 
-              temperature: float, 
-              max_new_tokens: int
-             ) -> str:
+def chat_llama3_8b(
+    message: str, history: list, temperature: float, max_new_tokens: int
+) -> str:
     """
     Generate a streaming response using the llama3-8b model.
     Args:
@@ -99,81 +87,86 @@ def chat_llama3_8b(message: str,
     # global model
     conversation = []
     for user, assistant in history:
-        conversation.extend([{"role": "user", "content": user}, {"role": "assistant", "content": assistant}])
+        conversation.extend(
+            [
+                {"role": "user", "content": user},
+                {"role": "assistant", "content": assistant},
+            ]
+        )
     conversation.append({"role": "user", "content": message})
 
-    input_ids = tokenizer.apply_chat_template(conversation, return_tensors="pt").to(model.device)
-    print(model.device)
-    # subprocess.run(
-    #     "pip install pycuda==2023.1",
-    #     shell=True,
-    # )
-    # if "has_patch" not in model.__dict__:
-    #     from minference import MInference
-    #     minference_patch = MInference("minference", model_name)
-    #     model = minference_patch(model)
-    
-    streamer = TextIteratorStreamer(tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True)
+    input_ids = tokenizer.apply_chat_template(conversation, return_tensors="pt").to(
+        model.device
+    )
+
+    streamer = TextIteratorStreamer(
+        tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True
+    )
 
     generate_kwargs = dict(
-        input_ids= input_ids,
+        input_ids=input_ids,
         streamer=streamer,
         max_new_tokens=max_new_tokens,
         do_sample=True,
         temperature=temperature,
         eos_token_id=terminators,
     )
-    # This will enforce greedy generation (do_sample=False) when the temperature is passed 0, avoiding the crash.             
+    # This will enforce greedy generation (do_sample=False) when the temperature is passed 0, avoiding the crash.
     if temperature == 0:
-        generate_kwargs['do_sample'] = False
-        
+        generate_kwargs["do_sample"] = False
+
     t = Thread(target=model.generate, kwargs=generate_kwargs)
     t.start()
 
     outputs = []
     for text in streamer:
         outputs.append(text)
-        #print(outputs)
+        # print(outputs)
         yield "".join(outputs)
-        
+
 
 # Gradio block
-chatbot=gr.Chatbot(height=450, placeholder=PLACEHOLDER, label='Gradio ChatInterface')
+chatbot = gr.Chatbot(height=450, placeholder=PLACEHOLDER, label="Gradio ChatInterface")
 
 with gr.Blocks(fill_height=True, css=css) as demo:
-    
+
     gr.Markdown(DESCRIPTION)
-    gr.DuplicateButton(value="Duplicate Space for private use", elem_id="duplicate-button")
     gr.ChatInterface(
         fn=chat_llama3_8b,
         chatbot=chatbot,
         fill_height=True,
-        additional_inputs_accordion=gr.Accordion(label="⚙️ Parameters", open=False, render=False),
+        additional_inputs_accordion=gr.Accordion(
+            label="⚙️ Parameters", open=False, render=False
+        ),
         additional_inputs=[
-            gr.Slider(minimum=0,
-                      maximum=1, 
-                      step=0.1,
-                      value=0.95, 
-                      label="Temperature", 
-                      render=False),
-            gr.Slider(minimum=128, 
-                      maximum=4096,
-                      step=1,
-                      value=512, 
-                      label="Max new tokens", 
-                      render=False ),
-            ],
+            gr.Slider(
+                minimum=0,
+                maximum=1,
+                step=0.1,
+                value=0.95,
+                label="Temperature",
+                render=False,
+            ),
+            gr.Slider(
+                minimum=128,
+                maximum=4096,
+                step=1,
+                value=512,
+                label="Max new tokens",
+                render=False,
+            ),
+        ],
         examples=[
-            ['How to setup a human base on Mars? Give short answer.'],
-            ['Explain theory of relativity to me like I’m 8 years old.'],
-            ['What is 9,000 * 9,000?'],
-            ['Write a pun-filled happy birthday message to my friend Alex.'],
-            ['Justify why a penguin might make a good king of the jungle.']
-            ],
+            ["How to setup a human base on Mars? Give short answer."],
+            ["Explain theory of relativity to me like I’m 8 years old."],
+            ["What is 9,000 * 9,000?"],
+            ["Write a pun-filled happy birthday message to my friend Alex."],
+            ["Justify why a penguin might make a good king of the jungle."],
+        ],
         cache_examples=False,
-                     )
-    
+    )
+
     gr.Markdown(LICENSE)
-    
+
 if __name__ == "__main__":
-    demo.launch()
+    demo.launch(share=False)

requirements.txt

@@ -2,5 +2,4 @@ triton==2.1.0
 accelerate
 transformers
 wheel
-setuptools
-pycuda==2023.1
+setuptools