Add the trainer script and corresponding slurm

fine-tuning/dataset_standardization.py

+from datasets import load_dataset
+from unsloth import FastLanguageModel
+def formatting_prompts_func(examples):
+    convos = examples["conversations"]
+    texts = [tokenizer.apply_chat_template(convo, tokenize = False, add_generation_prompt = False) for convo in convos]
+    return { "text" : texts, }
+pass
+
+def format_to_conversations(examples):
+  conversations = []
+  codes = examples["code"]
+  refined_codes = examples["refined code"]
+  summaries = examples["summary"]
+  for i in range(len(refined_codes)):
+      user_content = f'''Refine the C# code enclosed within tags [C#] and [/C#]. Return the refined code should be enclosed with tags [refined_C#] and [/refined_C#].
+      Summary of changes should be enclosed with [code_changes] and [/code_changes].
+      You do not do anything more than user asks you do it.
+      You do not generate any additional text.
+
+      [C#]
+      {codes[i]}
+      [/C#]
+      '''
+      assistant_content = f'''
+      [refined_C#]
+      {refined_codes[i]}
+      [/refined_C#]
+      [code_changes]
+      {summaries[i]}
+      [/code_changes]
+      '''
+      conversation = []
+
+      user_dict = {'content': user_content, 'role': 'user'}
+      assistant_dict = {'content': assistant_content, 'role': 'assistant'}
+      conversation.append(user_dict)
+      conversation.append(assistant_dict)
+
+      conversations.append(conversation)
+
+  return { "conversations" : conversations }
+pass
+
+max_seq_length = 32768 # Choose any! We auto support RoPE Scaling internally!
+dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
+load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.
+
+model, tokenizer = FastLanguageModel.from_pretrained(
+    model_name = "unsloth/Qwen2.5-Coder-7B-Instruct",
+    max_seq_length = max_seq_length,
+    dtype = dtype,
+    load_in_4bit = load_in_4bit,
+)
+
+dataset = load_dataset("atharva2721/qwen-refined-code", split = "train")
+
+dataset = dataset.map(format_to_conversations, batched = True,)
+dataset = dataset.map(formatting_prompts_func, batched = True,)
\ No newline at end of file

fine-tuning/main_fine_tuning.py

+from unsloth import FastLanguageModel, is_bfloat16_supported
+from unsloth.chat_templates import get_chat_template, train_on_responses_only
+from datasets import load_dataset
+from trl import SFTTrainer
+from transformers import TrainingArguments, DataCollatorForSeq2Seq
+
+
+def formatting_prompts_func(examples):
+    convos = examples["conversations"]
+    texts = [tokenizer.apply_chat_template(convo, tokenize = False, add_generation_prompt = False) for convo in convos]
+    return { "text" : texts, }
+pass
+
+def format_to_conversations(examples):
+  conversations = []
+  codes = examples["code"]
+  refined_codes = examples["refined code"]
+  summaries = examples["summary"]
+  for i in range(len(refined_codes)):
+      user_content = f'''Refine the C# code enclosed within tags [C#] and [/C#]. Return the refined code should be enclosed with tags [refined_C#] and [/refined_C#].
+      Summary of changes should be enclosed with [code_changes] and [/code_changes].
+      You do not do anything more than user asks you do it.
+      You do not generate any additional text.
+
+      [C#]
+      {codes[i]}
+      [/C#]
+      '''
+      assistant_content = f'''
+      [refined_C#]
+      {refined_codes[i]}
+      [/refined_C#]
+      [code_changes]
+      {summaries[i]}
+      [/code_changes]
+      '''
+      conversation = []
+
+      user_dict = {'content': user_content, 'role': 'user'}
+      assistant_dict = {'content': assistant_content, 'role': 'assistant'}
+      conversation.append(user_dict)
+      conversation.append(assistant_dict)
+
+      conversations.append(conversation)
+
+  return { "conversations" : conversations }
+pass
+
+max_seq_length = 32768 # Choose any! We auto support RoPE Scaling internally!
+dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
+load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.
+
+model, tokenizer = FastLanguageModel.from_pretrained(
+    model_name = "unsloth/Qwen2.5-Coder-7B-Instruct",
+    max_seq_length = max_seq_length,
+    dtype = dtype,
+    load_in_4bit = load_in_4bit,
+)
+
+model = FastLanguageModel.get_peft_model(
+    model,
+    r = 128, # Choose any number > 0 ! Suggested 8, 16, 32, 64, 128
+    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
+                      "gate_proj", "up_proj", "down_proj",],
+    lora_alpha = 16,
+    lora_dropout = 0, # Supports any, but = 0 is optimized
+    bias = "none",    # Supports any, but = "none" is optimized
+    # [NEW] "unsloth" uses 30% less VRAM, fits 2x larger batch sizes!
+    use_gradient_checkpointing = "unsloth", # True or "unsloth" for very long context
+    random_state = 3407,
+    use_rslora = False,  # We support rank stabilized LoRA
+    loftq_config = None, # And LoftQ
+)
+
+tokenizer = get_chat_template(
+    tokenizer,
+    chat_template = "qwen-2.5",
+)
+
+
+
+dataset = load_dataset("atharva2721/qwen-refined-code", split = "train")
+
+dataset = dataset.map(format_to_conversations, batched = True,)
+dataset = dataset.map(formatting_prompts_func, batched = True,)
+
+trainer = SFTTrainer(
+    model = model,
+    tokenizer = tokenizer,
+    train_dataset = dataset,
+    dataset_text_field = "text",
+    max_seq_length = max_seq_length,
+    data_collator = DataCollatorForSeq2Seq(tokenizer = tokenizer),
+    dataset_num_proc = 4,
+    packing = False, # Can make training 5x faster for short sequences.
+    args = TrainingArguments(
+        per_device_train_batch_size = 1,
+        gradient_accumulation_steps = 4, # Fixed major bug in latest Unsloth
+        warmup_steps = 5,
+        num_train_epochs = 1, # Set this for 1 full training run.
+        #max_steps = 10,
+        learning_rate = 2e-4,
+        fp16 = not is_bfloat16_supported(),
+        bf16 = is_bfloat16_supported(),
+        logging_steps = 1,
+        optim = "paged_adamw_8bit", # Save more memory
+        weight_decay = 0.01,
+        lr_scheduler_type = "linear",
+        seed = 3407,
+        output_dir = "outputs",
+        report_to = "none", # Use this for WandB etc
+    ),
+)
+
+trainer = train_on_responses_only(
+    trainer,
+    instruction_part = "<|im_start|>user\n",
+    response_part = "<|im_start|>assistant\n",
+)
+
+
+tokenizer.decode(trainer.train_dataset[0]["input_ids"])
+
+space = tokenizer(" ", add_special_tokens = False).input_ids[0]
+tokenizer.decode([space if x == -100 else x for x in trainer.train_dataset[0]["labels"]])
+
+gpu_stats = torch.cuda.get_device_properties(0)
+start_gpu_memory = round(torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3)
+max_memory = round(gpu_stats.total_memory / 1024 / 1024 / 1024, 3)
+print(f"GPU = {gpu_stats.name}. Max memory = {max_memory} GB.")
+print(f"{start_gpu_memory} GB of memory reserved.")
+
+trainer_stats = trainer.train()
+
+used_memory = round(torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3)
+used_memory_for_lora = round(used_memory - start_gpu_memory, 3)
+used_percentage = round(used_memory / max_memory * 100, 3)
+lora_percentage = round(used_memory_for_lora / max_memory * 100, 3)
+print(f"{trainer_stats.metrics['train_runtime']} seconds used for training.")
+print(
+    f"{round(trainer_stats.metrics['train_runtime']/60, 2)} minutes used for training."
+)
+print(f"Peak reserved memory = {used_memory} GB.")
+print(f"Peak reserved memory for training = {used_memory_for_lora} GB.")
+print(f"Peak reserved memory % of max memory = {used_percentage} %.")
+print(f"Peak reserved memory for training % of max memory = {lora_percentage} %.")
+
+model.save_pretrained("models/first_finetuned_model")  # Local saving
+tokenizer.save_pretrained("models/first_finetuned_model")
+# model.push_to_hub("your_name/lora_model", token = "...") # Online saving
+# tokenizer.push_to_hub("your_name/lora_model", token = "...") # Online saving
\ No newline at end of file

fine-tuning/slurm_qwen.sh

+#!/usr/bin/zsh
+
+### Add basic configuration for job
+
+#SBATCH --job-name=fine_tuning
+#SBATCH --output=logs/fine_tuning_%j.log
+#SBATCH --error=logs/fine_tuning_error_%j.log
+#SBATCH --nodes=1
+#SBATCH --ntasks=1
+#SBATCH --cpus-per-task=8
+#SBATCH --gres=gpu:1
+#SBATCH --time=2-15:00:00
+
+
+###------------------------------------------------------------------------------------------------------------------------------
+
+### Run the project in work directory of the cluster (configure based on need!! 
+### RWTH File System : https://help.itc.rwth-aachen.de/en/service/rhr4fjjutttf/article/da307ec2c60940b29bd42ac483fc3ea7/
+cd $HPCWORK
+cd codebud/fine-tuning
+###------------------------------------------------------------------------------------------------------------------------------
+
+### JOB SCRIPT RUN
+module load GCCcore/.13.2.0
+module load Python/3.11.5
+module load CUDA
+
+source ../../venvs/codebud/bin/activate
+echo $VIRTUAL_ENV
+
+python --version
+python main_fine_tuning.py
+
+module unload CUDA
+module unload Python/3.11.5
+
+deactivate
+echo "Script ran successfully"
\ No newline at end of file