changed space to use LiquidAI/LFM2-1.2

README.md

@@ -1,10 +1,10 @@
 ---
-title: SOLAR-math-2x10.7b MoE-Chat
-emoji: 🏆
-colorFrom: yellow
-colorTo: red
+title: LiquidAI - LFM2-1.2B Chat
+emoji: 🌊
+colorFrom: orange
+colorTo: pink
 sdk: gradio
-sdk_version: 4.1.2
+sdk_version: 5.7.3
 app_file: app.py
 pinned: true
 license: apache-2.0

app.py

@@ -1,82 +1,80 @@
 import spaces
 import gradio as gr
 import torch
-from transformers import AutoModelForCausalLM, AutoTokenizer
-from transformers import StoppingCriteria, StoppingCriteriaList, TextIteratorStreamer
+from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
 from threading import Thread
 
-# Lazy loading the model to meet huggingface stateless GPU requirements 
-
-# Defining a custom stopping criteria class for the model's text generation.
-class StopOnTokens(StoppingCriteria):
-    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
-        stop_ids = [50256, 50295]  # IDs of tokens where the generation should stop.
-        for stop_id in stop_ids:
-            if input_ids[0][-1] == stop_id:  # Checking if the last generated token is a stop token.
-                return True
-        return False
-
-
-# Function to generate model predictions.
 @spaces.GPU
 def predict(message, history):
     torch.set_default_device("cuda")
-
-    # Loading the tokenizer and model from Hugging Face's model hub.
-    tokenizer = AutoTokenizer.from_pretrained(
-        "macadeliccc/SOLAR-math-2x10.7b",
-        trust_remote_code=True
-    )
+    
+    # Load model and tokenizer
+    model_id = "LiquidAI/LFM2-1.2B"
+    tokenizer = AutoTokenizer.from_pretrained(model_id)
     model = AutoModelForCausalLM.from_pretrained(
-        "macadeliccc/SOLAR-math-2x10.7b",
-        torch_dtype="auto",
-        load_in_4bit=True,
-        trust_remote_code=True
+        model_id,
+        device_map="auto",
+        torch_dtype=torch.bfloat16,
+        trust_remote_code=True,
+        load_in_4bit=True,  # Keeping 4-bit quantization for efficiency
+        # attn_implementation="flash_attention_2"  # Uncomment on compatible GPU
     )
-    history_transformer_format = history + [[message, ""]]
-    stop = StopOnTokens()
-
-    # Formatting the input for the model.
-    system_prompt = "<|im_start|>system\nYou are Solar, a helpful AI assistant.<|im_end|>"
-    messages = system_prompt + "".join(["".join(["\n<|im_start|>user\n" + item[0], "<|im_end|>\n<|im_start|>assistant\n" + item[1]]) for item in history_transformer_format])
-    input_ids = tokenizer([messages], return_tensors="pt").to('cuda')
+    
+    # Format conversation history for chat template
+    messages = [{"role": "user" if i % 2 == 0 else "assistant", "content": msg} 
+                for conv in history for i, msg in enumerate(conv) if msg]
+    messages.append({"role": "user", "content": message})
+    
+    # Apply chat template
+    input_ids = tokenizer.apply_chat_template(
+        messages,
+        add_generation_prompt=True,
+        return_tensors="pt",
+        tokenize=True
+    ).to('cuda')
+    
+    # Setup streamer for real-time output
     streamer = TextIteratorStreamer(tokenizer, timeout=10., skip_prompt=True, skip_special_tokens=True)
+    
+    # Generation parameters
     generate_kwargs = dict(
-        input_ids,
+        input_ids=input_ids,
         streamer=streamer,
         max_new_tokens=256,
         do_sample=True,
-        top_p=0.95,
-        top_k=50,
-        temperature=0.7,
-        num_beams=1,
-        stopping_criteria=StoppingCriteriaList([stop])
+        temperature=0.3,
+        min_p=0.15,
+        repetition_penalty=1.05,
+        pad_token_id=tokenizer.eos_token_id
     )
+    
+    # Start generation in separate thread
     t = Thread(target=model.generate, kwargs=generate_kwargs)
-    t.start()  # Starting the generation in a separate thread.
+    t.start()
+    
+    # Stream tokens
     partial_message = ""
     for new_token in streamer:
         partial_message += new_token
-        if '<|im_end|>' in partial_message:  # Breaking the loop if the stop token is generated.
-            break
         yield partial_message
 
-
-# Setting up the Gradio chat interface.
-gr.ChatInterface(predict,
-                 description="""
-                 <center><img src="https://huggingface.co/macadeliccc/SOLAR-math-2x10.7b-v0.2/resolve/main/solar.png" width="33%"></center>\n\n
-                 Chat with [macadeliccc/SOLAR-math-2x10.7b-v0.2](https://huggingface.co/macadeliccc/SOLAR-math-2x10.7b-v0.2), the first Mixture of Experts made by merging two fine-tuned [upstage/SOLAR-10.7B-v1.0](https://huggingface.co/upstage/SOLAR-10.7B-v1.0) models.
-                 This model (19.2B param) scores top 5 on several evaluations. Output is considered experimental.\n\n
-                 ❤️ If you like this work, please follow me on [Hugging Face](https://huggingface.co/macadeliccc) and [LinkedIn](https://www.linkedin.com/in/tim-dolan-python-dev/).
-                 """,
-                 examples=[
-                     'Can you solve the equation 2x + 3 = 11 for x?',
-                     'How does Fermats last theorem impact number theory?',
-                     'What is a vector in the scope of computer science rather than physics?',
-                     'Use a list comprehension to create a list of squares for numbers from 1 to 10.',
-                     'Recommend some popular science fiction books.',
-                     'Can you write a short story about a time-traveling detective?'
-                 ],
-                 theme=gr.themes.Soft(primary_hue="purple"),
-                 ).launch()
+# Setup Gradio interface
+gr.ChatInterface(
+    predict,
+    description="""
+    <center><h2>LiquidAI LFM2-1.2B Chat</h2></center>
+    
+    Chat with [LiquidAI/LFM2-1.2B](https://huggingface.co/LiquidAI/LFM2-1.2B), a compact and efficient language model.
+    
+    This model provides high-quality responses while maintaining a small footprint, making it ideal for fast inference.
+    """,
+    examples=[
+        'Can you solve the equation 2x + 3 = 11 for x?',
+        'What is C. elegans?',
+        'Explain quantum computing in simple terms',
+        'Write a Python function to find prime numbers',
+        'What are the key differences between RNA and DNA?',
+        'Can you write a haiku about artificial intelligence?'
+    ],
+    theme=gr.themes.Soft(primary_hue="blue"),
+).launch()

requirements.txt

@@ -1,12 +1,8 @@
-git+https://github.com/huggingface/diffusers.git
 git+https://github.com/huggingface/transformers.git
 git+https://github.com/huggingface/peft.git
---extra-index-url https://download.pytorch.org/whl/cu113
+--extra-index-url https://download.pytorch.org/whl/cu126
 torch
 pydantic
-Pillow
 accelerate
-bitsandbytes
 spaces
-invisible_watermark