app.py

import gc

import gradio as gr
# import torch
# from transformers import AutoTokenizer, AutoModelForCausalLM #, HqqConfig

# # # quant_config = HqqConfig(nbits=8, group_size=64)

# MODEL_ID = "HuggingFaceTB/SmolLM3-3B"
# DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

# print("Loading tokenizer & model…")
# tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
# # # model = AutoModelForCausalLM.from_pretrained(MODEL_ID, torch_dtype=torch.bfloat16).to(DEVICE)

# model =\
#         AutoModelForCausalLM\
#                 .from_pretrained(
#                         MODEL_ID, 
#                         torch_dtype=torch.float16, 
#                         # device_map="cuda", 
#                         # quantization_config=quant_config
#         ).to(DEVICE)

# gc.collect()

#########

import torch
from transformers import TorchAoConfig, AutoModelForCausalLM, AutoTokenizer
from torchao.quantization import Float8DynamicActivationFloat8WeightConfig, Float8WeightOnlyConfig, Float4DynamicActivationFloat4WeightConfig
# quant_config = Float8WeightOnlyConfig()
quant_config = Float4DynamicActivationFloat4WeightConfig() # Float8DynamicActivationFloat8WeightConfig()
quantization_config = TorchAoConfig(quant_type=quant_config)

MODEL_ID = "HuggingFaceTB/SmolLM3-3B"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype="auto",
    device_map="auto",
    quantization_config=quantization_config)

gc.collect()

pt2e.move_exported_model_to_eval(model)

gc.collect()

#########

# from unsloth import FastLanguageModel

# model, tokenizer = FastLanguageModel.from_pretrained(
#     "unsloth/Llama-3.2-3B-Instruct-bnb-4bit",
#     max_seq_length=128_000,
#     load_in_4bit=True
# )

#########

# import gc

# import gradio as gr
# from transformers import AutoTokenizer
# from optimum.onnxruntime import ORTModelForCausalLM, ORTQuantizer
# from optimum.onnxruntime.configuration import AutoQuantizationConfig

# MODEL_NAME = "HuggingFaceTB/SmolLM3-3B"



# tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
# model = ORTModelForCausalLM.from_pretrained(MODEL_NAME, export=True)

# print("Creating quant config")
# qconfig = AutoQuantizationConfig.avx512_vnni(is_static=False, per_channel=True)
# print("Creating quant config successful")

# print("Creating quantizer")
# quantizer = ORTQuantizer.from_pretrained(model)
# print("Creating quantizer successful")
# # Step 4: Perform quantization saving output in a new directory
# quantized_model_dir = "./quantized_model"
# print("Starting quantization...")
# quantizer.quantize(save_dir=quantized_model_dir, quantization_config=qconfig)
# print("Quantization was successful. Garbage collecting...")

# del(quantizer)
# del(qconfig)
# del(model)

# Run garbage collection again to release memory from quantizer objects
gc.collect()

# # Step 5: Load the quantized ONNX model for inference
# print("Loading quantized ONNX model for inference...")
# model = ORTModelForCausalLM.from_pretrained(quantized_model_dir)
# print("Loading model was succcessful. Garbage collecting.")

# Garbage collection again after final loading
gc.collect()

#########

# print("Loading tokenizer & model…")
# import gradio as gr
# from transformers import AutoTokenizer
# from optimum.onnxruntime import ORTModelForCausalLM

# MODEL_ID = "HuggingFaceTB/SmolLM3-3B"
# tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
# model = ORTModelForCausalLM.from_pretrained(MODEL_ID, export=True, quantize=True)

#########


# -------------------------------------------------
# Optional tool(s)
# -------------------------------------------------
# TOOLS = [{
#     "name": "get_weather",
#     "description": "Get the current weather in a given city",
#     "parameters": {
#         "type": "object",
#         "properties": {
#             "city": {"type": "string", "description": "City name"}
#         },
#         "required": ["city"]
#     }
# }]

# -------------------------------------------------
# Helpers
# -------------------------------------------------

def build_messages(history, enable_thinking: bool):
    """Convert Gradio history to the chat template."""
    messages = []
    for h in history:
        messages.append({"role": h["role"], "content": h["content"]})
    # Add system instruction for mode
    system_flag = "/think" if enable_thinking else "/no_think"
    messages.insert(0, {"role": "system", "content": system_flag})
    return messages

def chat_fn(history, enable_thinking, temperature, top_p, top_k, repetition_penalty, max_new_tokens):
    """Generate a streaming response."""
    messages = build_messages(history, enable_thinking)
    text = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True,
        # xml_tools=TOOLS
    )
    inputs = tokenizer(text, return_tensors="pt")
    gc.collect()
    with torch.inference_mode():
        streamer = model.generate(
            **inputs,
            max_new_tokens=max_new_tokens,
            do_sample=True,
            temperature=temperature,
            top_p=top_p,
            top_k=top_k,
            repetition_penalty=repetition_penalty,
            pad_token_id=tokenizer.eos_token_id,
            streamer=None          # we'll yield manually
        )
    gc.collect()
    output_ids = streamer[0][len(inputs.input_ids[0]):]
    response = tokenizer.decode(output_ids, skip_special_tokens=True)
    if isinstance(response, str): 
        response = response.replace('<think>',"# &lt;think&gt;").replace('</think>',"&lt;/think&gt;")
    elif isinstance(response,list):
        response = [paper.replace('<think>',"# &lt;think&gt;").replace('</think>',"&lt;/think&gt;") for paper in response]
    else:
        raise ValueError("Tokenizer response seems malformed; Not a string, nor a list?!?!")

    # streaming char-by-char
    history.append({"role": "assistant", "content": ""})
    for ch in response:
        history[-1]["content"] += ch
        yield history

# -------------------------------------------------
# Blocks UI
# -------------------------------------------------
with gr.Blocks(title="SmolLM3-3B Chat") as demo:
    gr.Markdown("## 🤖 SmolLM3-3B Chatbot (Streaming)")
    with gr.Row():
        enable_think = gr.Checkbox(label="Enable Extended Thinking (/think)", value=True)
        temperature = gr.Slider(0.0, 1.0, value=0.6, label="Temperature")
        top_p = gr.Slider(0.0, 1.0, value=0.95, label="Top-p")
        top_k = gr.Slider(1,40,value=20,label="Top_k")
        repetition_penalty = gr.Slider(1.0,1.4,value=1.1,label="Repetition_Penalty")
        max_new_tokens = gr.Slider(1000,32768,value=32768,label="Max_New_Tokens")
    chatbot = gr.Chatbot(type="messages")
    msg = gr.Textbox(placeholder="Type your message here…", lines=1)
    clear = gr.Button("Clear")

    def user_fn(user_msg, history):
        return "", history + [{"role": "user", "content": user_msg}]

    msg.submit(
        user_fn, [msg, chatbot], [msg, chatbot], queue=False
    ).then(
        chat_fn, [chatbot, enable_think, temperature, top_p, top_k, repetition_penalty, max_new_tokens], chatbot
    )
    clear.click(lambda: None, None, chatbot, queue=False)

demo.queue().launch()