Upload proton_code_llama_fine_tuning_using_qlora_inferencing.py

proton_code_llama_fine_tuning_using_qlora_inferencing.py

@@ -0,0 +1,143 @@
+# -*- coding: utf-8 -*-
+"""Proton_Code_Llama_Fine_Tuning_using_QLora_Inferencing.ipynb
+
+Automatically generated by Colaboratory.
+
+Original file is located at
+    https://colab.research.google.com/drive/1snQ0jgdbZDXNuVpvryjfENfCjk8K2SE2
+"""
+
+!pip install -q -U trl transformers accelerate git+https://github.com/huggingface/peft.git
+!pip install -q datasets bitsandbytes einops wandb
+!pip install peft
+!pip install --upgrade gradio
+!pip install sentence_transformers
+!pip install accelerate
+!pip install bitsandbytes
+
+import pandas as pd
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, AutoTokenizer
+from peft import PeftModel, PeftConfig
+from sentence_transformers.util import semantic_search
+import requests
+
+model_id = "sentence-transformers/all-MiniLM-L6-v2"
+hf_token = "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"
+
+data = pd.read_excel('/content/sample_data/Proton Gen AI.xlsx')
+
+text_col = []
+
+for _, row in data.iterrows():
+
+  prompt = "Below is an instruction that describes a task to generate a domain specific code called a Proton UI. It is almost identical to YAML with some additions. Write a response that appropriately completes the request and generates the code. \n\n"
+
+  instruction = str(row['Description'])
+
+  response = str(row['UI'])
+
+  text = prompt + "### Instruction:\n" + instruction + "\n### Response:\n" + response
+
+  text_col.append(text)
+
+data.loc[:, "text"] = text_col
+
+api_url = f"https://api-inference.huggingface.co/pipeline/feature-extraction/{model_id}"
+headers = {"Authorization": f"Bearer {hf_token}"}
+
+def create_embeddings(texts):
+    response = requests.post(api_url, headers=headers, json={"inputs": texts, "options":{"wait_for_model":True}})
+    return response.json()
+
+output = create_embeddings(data["Description"].to_list())
+embeddings = pd.DataFrame(output)
+dataset_embeddings = torch.from_numpy(embeddings.to_numpy()).to(torch.float)
+
+bnb_config = BitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_quant_type="nf4",
+    bnb_4bit_compute_dtype=torch.bfloat16,
+)
+config = PeftConfig.from_pretrained("Siva1995/codellama_finetuned_proton_ui")
+model = AutoModelForCausalLM.from_pretrained("codellama/CodeLlama-7b-Instruct-hf",quantization_config=bnb_config,
+    trust_remote_code=True)
+model = PeftModel.from_pretrained(model, "Siva1995/codellama_finetuned_proton_ui")
+
+model_tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path,trust_remote_code = True)
+model_tokenizer.pad_token = model_tokenizer.eos_token
+
+generation_config = model.generation_config
+generation_config.max_new_token = 1000
+generation_config.temperature = 0.1
+generation_config.top_p = 0.5
+generation_config.num_return_sequence = 1
+generation_config.pad_token_id = model_tokenizer.pad_token_id
+generation_config.eos_token_id = model_tokenizer.eos_token_id
+
+device = 'cuda:0'
+
+priming_text = "You are an expert at writing a complete Proton UI code. Here are the examples of Proton description and corresponding UI responses."
+decorator = "Your response should follow the Proper Proton UI format as mentioned in the above examples"
+
+def create_fewshot_examples(question):
+  # question = ["Create a dropdown type UI with Sports?"]
+  # print(question)
+  output = create_embeddings(question)
+  query_embeddings = torch.FloatTensor(output)
+  hits = semantic_search(query_embeddings, dataset_embeddings, top_k=3)
+  similar_description = [data["Description"].to_list()[hits[0][i]['corpus_id']] for i in range(len(hits[0]))]
+  similar_data = data[data['Description'].isin(similar_description)]
+  dyn_fewshot_examples = similar_data[['Description','UI']].to_dict(orient='records')
+  return dyn_fewshot_examples
+
+def getExamples(question):
+    dyn_fewshot_examples = create_fewshot_examples(question)
+    r = ''
+    for example in dyn_fewshot_examples:
+        r = '\r\n'.join([r,'Description:',example['Description'],'UI:',example['UI']])
+    return r
+
+def prepare_prompt(query):
+  prompt_prep = ''.join([priming_text,getExamples(query),'\n',decorator,'\n','Similar to the above examples, generate only one Proton UI for the following description:',query])
+  return prompt_prep
+
+def generate(query):
+  prompt_prep = prepare_prompt(query)
+  encodings = model_tokenizer(prompt_prep, return_tensors = 'pt').to(device)
+
+  with torch.inference_mode():
+    outputs = model.generate(input_ids=encodings.input_ids,
+                            attention_mask = encodings.attention_mask,
+                            generation_config = generation_config,
+                          max_new_tokens=1000,do_sample=True)
+
+  outputs = model_tokenizer.decode(outputs[0],skip_special_tokens=True,best_of=1)
+
+  return outputs[outputs.find(query): ]#outputs.find('Explanation')
+
+output = generate('Create a choice type of Proton UI for choosing vegetables')
+print(output)
+
+import gradio as gr
+with gr.Blocks() as demo:
+    gr.Markdown("****This application is built on fine-tuned on specific data on codellama 7b on 4 bit fine tuned****")
+    gr.Markdown("**No Paid service has been leveraged**")
+    gr.Markdown("**Type your Query Run to see the output**")
+    with gr.Row():
+        out = gr.Textbox()
+    with gr.Row():
+        inp = gr.Textbox(placeholder="Type your Question")
+    btn = gr.Button("Run")
+    btn.click(fn=output, inputs=inp, outputs=out)
+demo.launch(auth=("user", "password"),share=True)
+
+import random
+import gradio as gr
+
+def alternatingly_agree(message, history):
+  output = generate(str(message))
+  return output
+
+gr.ChatInterface(alternatingly_agree).launch()
+