argo/qlora-training.yaml

# QLoRA Fine-tuning Workflow
# Trains QLora adapters from a reference model using data from Milvus vector database
# Triggered via NATS: ai.pipeline.trigger with pipeline="qlora-training"
---
apiVersion: argoproj.io/v1alpha1
kind: WorkflowTemplate
metadata:
  name: qlora-training
  namespace: ai-ml
  labels:
    app.kubernetes.io/name: qlora-training
    app.kubernetes.io/part-of: llm-workflows
spec:
  entrypoint: train-qlora
  serviceAccountName: argo-workflow
  
  arguments:
    parameters:
      - name: reference-model
        description: "Base model to fine-tune (HuggingFace model ID or path)"
        value: "mistralai/Mistral-7B-Instruct-v0.3"
      - name: output-name
        description: "Name for the output QLora adapter"
        value: "qlora-adapter"
      - name: milvus-collections
        description: "Comma-separated list of Milvus collections to use (empty = all available)"
        value: ""
      - name: learning-rate
        value: "2e-4"
        description: "Learning rate for training"
      - name: num-epochs
        value: "3"
        description: "Number of training epochs"
      - name: batch-size
        value: "4"
        description: "Training batch size"
      - name: max-seq-length
        value: "2048"
        description: "Maximum sequence length"
      - name: lora-r
        value: "64"
        description: "LoRA attention dimension"
      - name: lora-alpha
        value: "16"
        description: "LoRA alpha parameter"
      - name: lora-dropout
        value: "0.05"
        description: "LoRA dropout rate"

  volumeClaimTemplates:
    - metadata:
        name: model-storage
      spec:
        accessModes: ["ReadWriteMany"]
        storageClassName: nfs-slow
        resources:
          requests:
            storage: 50Gi

  templates:
    - name: train-qlora
      dag:
        tasks:
          - name: fetch-training-data
            template: fetch-data
            arguments:
              parameters:
                - name: milvus-collections
                  value: "{{workflow.parameters.milvus-collections}}"

          - name: prepare-dataset
            template: prepare-data
            dependencies: [fetch-training-data]
            arguments:
              parameters:
                - name: max-seq-length
                  value: "{{workflow.parameters.max-seq-length}}"
              artifacts:
                - name: raw-data
                  from: "{{tasks.fetch-training-data.outputs.artifacts.raw-data}}"

          - name: train-model
            template: train
            dependencies: [prepare-dataset]
            arguments:
              parameters:
                - name: reference-model
                  value: "{{workflow.parameters.reference-model}}"
                - name: output-name
                  value: "{{workflow.parameters.output-name}}"
                - name: learning-rate
                  value: "{{workflow.parameters.learning-rate}}"
                - name: num-epochs
                  value: "{{workflow.parameters.num-epochs}}"
                - name: batch-size
                  value: "{{workflow.parameters.batch-size}}"
                - name: max-seq-length
                  value: "{{workflow.parameters.max-seq-length}}"
                - name: lora-r
                  value: "{{workflow.parameters.lora-r}}"
                - name: lora-alpha
                  value: "{{workflow.parameters.lora-alpha}}"
                - name: lora-dropout
                  value: "{{workflow.parameters.lora-dropout}}"
              artifacts:
                - name: training-data
                  from: "{{tasks.prepare-dataset.outputs.artifacts.training-data}}"

    - name: fetch-data
      inputs:
        parameters:
          - name: milvus-collections
      outputs:
        artifacts:
          - name: raw-data
            path: /tmp/raw-data
      container:
        image: python:3.13-slim
        command: [python]
        args:
          - -c
          - |
            import subprocess
            subprocess.run(["pip", "install", "pymilvus", "-q"], check=True)
            
            import json
            from pathlib import Path
            from pymilvus import connections, Collection, utility

            MILVUS_HOST = "milvus.ai-ml.svc.cluster.local"
            MILVUS_PORT = 19530
            collections_param = "{{inputs.parameters.milvus-collections}}"
            
            output_dir = Path("/tmp/raw-data")
            output_dir.mkdir(parents=True, exist_ok=True)

            print(f"Connecting to Milvus at {MILVUS_HOST}:{MILVUS_PORT}")
            connections.connect(host=MILVUS_HOST, port=MILVUS_PORT)
            
            # Determine which collections to use
            if collections_param and collections_param.strip():
                collection_names = [c.strip() for c in collections_param.split(",")]
                print(f"Using specified collections: {collection_names}")
            else:
                # Get all available collections
                collection_names = utility.list_collections()
                print(f"Using all available collections: {collection_names}")
            
            all_training_data = []
            
            for collection_name in collection_names:
                if not utility.has_collection(collection_name):
                    print(f"Warning: Collection {collection_name} not found, skipping")
                    continue
                
                print(f"Fetching data from collection: {collection_name}")
                collection = Collection(collection_name)
                collection.load()
                
                # Query all data from the collection
                # Note: Adjust field names based on your schema
                try:
                    # Get collection schema to determine fields
                    schema = collection.schema
                    field_names = [field.name for field in schema.fields if field.name != "id"]
                    
                    # Query all entities (limited to reasonable batch size)
                    # For large collections, you may want to implement pagination
                    results = collection.query(
                        expr="id >= 0",
                        output_fields=field_names,
                        limit=100000
                    )
                    
                    print(f"  Retrieved {len(results)} records from {collection_name}")
                    
                    for result in results:
                        # Extract text field (adjust based on your schema)
                        text_content = result.get("text", "")
                        source = result.get("source", collection_name)
                        
                        if text_content:
                            all_training_data.append({
                                "text": text_content,
                                "source": source,
                                "collection": collection_name
                            })
                    
                except Exception as e:
                    print(f"Error querying collection {collection_name}: {e}")
                    continue
            
            # Save all training data
            output_file = output_dir / "training_data.json"
            with open(output_file, "w") as f:
                json.dump({"data": all_training_data}, f)
            
            print(f"Total training samples collected: {len(all_training_data)}")
            print(f"Saved to {output_file}")
            
            connections.disconnect("default")
        envFrom:
          - configMapRef:
              name: ai-services-config
        resources:
          requests:
            memory: 1Gi
            cpu: 500m

    - name: prepare-data
      inputs:
        parameters:
          - name: max-seq-length
        artifacts:
          - name: raw-data
            path: /tmp/raw-data
      outputs:
        artifacts:
          - name: training-data
            path: /tmp/training-data
      container:
        image: python:3.13-slim
        command: [python]
        args:
          - -c
          - |
            import json
            from pathlib import Path
            
            max_seq_length = int("{{inputs.parameters.max-seq-length}}")
            
            input_dir = Path("/tmp/raw-data")
            output_dir = Path("/tmp/training-data")
            output_dir.mkdir(parents=True, exist_ok=True)
            
            # Load raw data
            with open(input_dir / "training_data.json") as f:
                data = json.load(f)
            
            raw_samples = data["data"]
            print(f"Processing {len(raw_samples)} raw samples")
            
            # Prepare data in instruction format for fine-tuning
            # Using Alpaca-style format: instruction + response
            training_samples = []
            
            for sample in raw_samples:
                text = sample["text"]
                source = sample.get("source", "")
                
                # Create instruction-response pairs
                # You can customize this based on your use case
                training_sample = {
                    "instruction": f"Based on the following information from {source}, provide a comprehensive response:",
                    "input": text[:max_seq_length // 2],  # Truncate if needed
                    "output": text[:max_seq_length // 2],
                    "source": source
                }
                training_samples.append(training_sample)
            
            # Split into train/validation (90/10)
            split_idx = int(len(training_samples) * 0.9)
            train_data = training_samples[:split_idx]
            val_data = training_samples[split_idx:]
            
            # Save prepared datasets
            with open(output_dir / "train.json", "w") as f:
                json.dump(train_data, f)
            
            with open(output_dir / "validation.json", "w") as f:
                json.dump(val_data, f)
            
            print(f"Prepared {len(train_data)} training samples")
            print(f"Prepared {len(val_data)} validation samples")
            print("Data preparation complete")
        resources:
          requests:
            memory: 2Gi
            cpu: 500m

    - name: train
      inputs:
        parameters:
          - name: reference-model
          - name: output-name
          - name: learning-rate
          - name: num-epochs
          - name: batch-size
          - name: max-seq-length
          - name: lora-r
          - name: lora-alpha
          - name: lora-dropout
        artifacts:
          - name: training-data
            path: /tmp/training-data
      container:
        image: python:3.13-slim
        command: [bash]
        args:
          - -c
          - |
            set -e
            
            echo "Installing dependencies..."
            pip install -q torch transformers peft datasets accelerate bitsandbytes scipy
            
            echo "Starting QLoRA training..."
            
            python << 'EOF'
            import json
            import os
            from pathlib import Path
            from datasets import Dataset
            from transformers import (
                AutoModelForCausalLM,
                AutoTokenizer,
                BitsAndBytesConfig,
                TrainingArguments,
                Trainer,
                DataCollatorForLanguageModeling
            )
            from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
            import torch

            # Parameters
            reference_model = "{{inputs.parameters.reference-model}}"
            output_name = "{{inputs.parameters.output-name}}"
            learning_rate = float("{{inputs.parameters.learning-rate}}")
            num_epochs = int("{{inputs.parameters.num-epochs}}")
            batch_size = int("{{inputs.parameters.batch-size}}")
            max_seq_length = int("{{inputs.parameters.max-seq-length}}")
            lora_r = int("{{inputs.parameters.lora-r}}")
            lora_alpha = int("{{inputs.parameters.lora-alpha}}")
            lora_dropout = float("{{inputs.parameters.lora-dropout}}")
            
            data_dir = Path("/tmp/training-data")
            output_dir = Path("/mnt/model-storage") / output_name
            output_dir.mkdir(parents=True, exist_ok=True)
            
            print(f"Training configuration:")
            print(f"  Model: {reference_model}")
            print(f"  Learning rate: {learning_rate}")
            print(f"  Epochs: {num_epochs}")
            print(f"  Batch size: {batch_size}")
            print(f"  Max sequence length: {max_seq_length}")
            print(f"  LoRA r: {lora_r}, alpha: {lora_alpha}, dropout: {lora_dropout}")
            
            # Load datasets
            with open(data_dir / "train.json") as f:
                train_data = json.load(f)
            
            with open(data_dir / "validation.json") as f:
                val_data = json.load(f)
            
            print(f"Loaded {len(train_data)} training samples, {len(val_data)} validation samples")
            
            # Load tokenizer
            print(f"Loading tokenizer from {reference_model}...")
            tokenizer = AutoTokenizer.from_pretrained(reference_model, trust_remote_code=True)
            if tokenizer.pad_token is None:
                tokenizer.pad_token = tokenizer.eos_token
            
            # Prepare datasets
            def format_sample(sample):
                instruction = sample.get("instruction", "")
                input_text = sample.get("input", "")
                output = sample.get("output", "")
                
                if input_text:
                    prompt = f"### Instruction:\n{instruction}\n\n### Input:\n{input_text}\n\n### Response:\n{output}"
                else:
                    prompt = f"### Instruction:\n{instruction}\n\n### Response:\n{output}"
                
                return {"text": prompt}
            
            train_dataset = Dataset.from_list([format_sample(s) for s in train_data])
            val_dataset = Dataset.from_list([format_sample(s) for s in val_data])
            
            # Tokenize datasets
            def tokenize_function(examples):
                return tokenizer(
                    examples["text"],
                    truncation=True,
                    max_length=max_seq_length,
                    padding="max_length"
                )
            
            print("Tokenizing datasets...")
            train_dataset = train_dataset.map(tokenize_function, batched=True, remove_columns=["text"])
            val_dataset = val_dataset.map(tokenize_function, batched=True, remove_columns=["text"])
            
            # Configure quantization for QLoRA
            bnb_config = BitsAndBytesConfig(
                load_in_4bit=True,
                bnb_4bit_quant_type="nf4",
                bnb_4bit_compute_dtype=torch.float16,
                bnb_4bit_use_double_quant=True,
            )
            
            # Load model
            print(f"Loading model {reference_model} with 4-bit quantization...")
            model = AutoModelForCausalLM.from_pretrained(
                reference_model,
                quantization_config=bnb_config,
                device_map="auto",
                trust_remote_code=True
            )
            
            # Prepare model for training
            model = prepare_model_for_kbit_training(model)
            
            # Configure LoRA
            lora_config = LoraConfig(
                r=lora_r,
                lora_alpha=lora_alpha,
                target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
                lora_dropout=lora_dropout,
                bias="none",
                task_type="CAUSAL_LM"
            )
            
            # Add LoRA adapters
            print("Adding LoRA adapters...")
            model = get_peft_model(model, lora_config)
            model.print_trainable_parameters()
            
            # Training arguments
            training_args = TrainingArguments(
                output_dir=str(output_dir / "checkpoints"),
                num_train_epochs=num_epochs,
                per_device_train_batch_size=batch_size,
                per_device_eval_batch_size=batch_size,
                gradient_accumulation_steps=4,
                learning_rate=learning_rate,
                fp16=True,
                logging_steps=10,
                evaluation_strategy="steps",
                eval_steps=50,
                save_strategy="steps",
                save_steps=100,
                save_total_limit=3,
                load_best_model_at_end=True,
                report_to="none",
                remove_unused_columns=False,
            )
            
            # Data collator
            data_collator = DataCollatorForLanguageModeling(
                tokenizer=tokenizer,
                mlm=False
            )
            
            # Trainer
            trainer = Trainer(
                model=model,
                args=training_args,
                train_dataset=train_dataset,
                eval_dataset=val_dataset,
                data_collator=data_collator,
            )
            
            # Train
            print("Starting training...")
            trainer.train()
            
            # Save final model
            print(f"Saving QLora adapter to {output_dir}")
            model.save_pretrained(str(output_dir / "final"))
            tokenizer.save_pretrained(str(output_dir / "final"))
            
            # Save training metadata
            metadata = {
                "reference_model": reference_model,
                "output_name": output_name,
                "training_params": {
                    "learning_rate": learning_rate,
                    "num_epochs": num_epochs,
                    "batch_size": batch_size,
                    "max_seq_length": max_seq_length,
                    "lora_r": lora_r,
                    "lora_alpha": lora_alpha,
                    "lora_dropout": lora_dropout
                },
                "dataset_info": {
                    "train_samples": len(train_data),
                    "val_samples": len(val_data)
                }
            }
            
            with open(output_dir / "metadata.json", "w") as f:
                json.dump(metadata, f, indent=2)
            
            print("Training complete!")
            print(f"QLora adapter saved to: {output_dir}")
            EOF
        envFrom:
          - configMapRef:
              name: ai-services-config
        volumeMounts:
          - name: model-storage
            mountPath: /mnt/model-storage
        resources:
          requests:
            memory: 16Gi
            cpu: 4
            nvidia.com/gpu: 1
          limits:
            memory: 32Gi
            cpu: 8
            nvidia.com/gpu: 1