from transformers import TextStreamer
from unsloth import FastLanguageModel
import torch
max_seq_length = 2048 # Choose any! We auto support RoPE Scaling internally!
dtype = 'Bfloat16' # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
load_in_4bit = True
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name ="alibidaran/Qwen3-instructive_reasoning",
    max_seq_length = max_seq_length,
    #dtype = dtype,
    load_in_4bit = load_in_4bit,
    #fast_inference = True, # Enable vLLM fast inference
    max_lora_rank = 128,
    gpu_memory_utilization = 0.6, # Reduce if out of memory
    # token = "hf_...", # use one if using gated models like meta-llama/Llama-2-7b-hf
)
FastLanguageModel.for_inference(model) # Enable native 2x faster inference
system_prompt="""
    You are a reasonable expert who thinks and answer the users question.
    Before respond first think and create a chain of thoughts in your mind.
    Then respond to the client.
    Your chain of thought and reflection must be in <thinking>..</thinking> format and your respond
    should be in the <output>..</output> format.
    """

messages = [
    {'role':'system','content':system_prompt},
    {"role": "user", "content":'How many r has the word of strawberry?' },

]
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize = True,
    add_generation_prompt = True, # Must add for generation
    return_tensors = "pt",
).to("cuda")

text_streamer = TextStreamer(tokenizer, skip_prompt = True)
_ = model.generate(input_ids = inputs, streamer = text_streamer, max_new_tokens =2048,
                   use_cache = True, temperature = 0.7, min_p = 0.9)