Delete setup_runpod.sh

setup_runpod.sh

@@ -1,208 +0,0 @@
-#!/bin/bash
-# =============================================================================
-# Blender Documentation Fine-tuning Setup for RunPod
-# Model: Qwen 3.5 27B | Training: FP16
-# =============================================================================
-
-set -e
-
-echo "=========================================="
-echo "Blender Qwen 3.5-27B Fine-tuning Setup"
-echo "=========================================="
-
-# =============================================================================
-# Step 1: Install Dependencies (exact versions from Unsloth Colab)
-# =============================================================================
-echo "[1/5] Installing dependencies..."
-
-pip install --upgrade uv
-
-# Install PyTorch and core packages
-uv pip install -qqq \
-    "torch==2.8.0" \
-    "triton>=3.3.0" \
-    "numpy" \
-    "pillow" \
-    "torchvision" \
-    "bitsandbytes" \
-    "unsloth_zoo[base] @ git+https://github.com/unslothai/unsloth-zoo" \
-    "unsloth[base] @ git+https://github.com/unslothai/unsloth"
-
-# Install Unsloth and extras
-uv pip install --upgrade --no-deps tokenizers
-uv pip install -qqq trl==0.22.2
-uv pip install -qqq unsloth unsloth_zoo
-uv pip install -qqq transformers==5.2.0
-
-# Install flash-linear-attention and xformers
-uv pip install --no-build-isolation flash-linear-attention
-uv pip install -qqq causal_conv1d==1.6.0
-uv pip install -qqq xformers==0.0.32.post2
-
-echo "✓ Dependencies installed"
-
-# =============================================================================
-# Step 2: Verify GPU
-# =============================================================================
-echo "[2/5] Verifying GPU..."
-nvidia-smi
-python3 -c "import torch; print(f'CUDA: {torch.cuda.is_available()}'); print(f'GPUs: {torch.cuda.device_count()}'); print(f'PyTorch: {torch.__version__}')"
-
-# =============================================================================
-# Step 3: Create Training Script
-# =============================================================================
-echo "[3/5] Creating training script..."
-
-cat > train_blender.py << 'TRAINEOF'
-#!/usr/bin/env python3
-"""
-Blender Documentation Fine-tuning
-Model: Qwen 3.5 27B (unsloth version - optimized)
-Training: FP16 (full precision)
-"""
-
-import os
-import json
-from datasets import load_dataset
-from unsloth import FastLanguageModel
-from trl import SFTTrainer
-from transformers import TrainingArguments
-
-# =============================================================================
-# Configuration
-# =============================================================================
-MODEL_NAME = "unsloth/Qwen3.5-27B"
-MAX_SEQ_LENGTH = 2048
-LEARNING_RATE = 2e-4
-NUM_EPOCHS = 2
-BATCH_SIZE = 2
-GRADIENT_ACCUMULATION = 4
-OUTPUT_DIR = "./blender_qwen_output"
-
-# =============================================================================
-# Load Model (FP16 - no quantization for best quality)
-# =============================================================================
-print("Loading model in FP16...")
-model, tokenizer = FastLanguageModel.from_pretrained(
-    model_name=MODEL_NAME,
-    max_seq_length=MAX_SEQ_LENGTH,
-    load_in_4bit=False,  # FP16 - best quality!
-    dtype="float16",     # Explicit FP16
-)
-
-# =============================================================================
-# Add LoRA Adapters
-# =============================================================================
-print("Adding LoRA adapters...")
-model = FastLanguageModel.get_peft_model(
-    model,
-    r=64,
-    lora_alpha=64,
-    lora_dropout=0,
-    target_modules=[
-        "q_proj", "k_proj", "v_proj", "o_proj",
-        "gate_proj", "up_proj", "down_proj"
-    ],
-    bias="none",
-    task_type="CAUSAL_LM",
-)
-
-# =============================================================================
-# Load Dataset
-# =============================================================================
-print("Loading dataset...")
-dataset = load_dataset("json", data_files={
-    "train": "train.jsonl",
-    "valid": "valid.jsonl"
-})
-
-print(f"Train samples: {len(dataset['train'])}")
-print(f"Valid samples: {len(dataset['valid'])}")
-
-# =============================================================================
-# Training Arguments (FP16)
-# =============================================================================
-training_args = TrainingArguments(
-    output_dir=OUTPUT_DIR,
-    num_train_epochs=NUM_EPOCHS,
-    per_device_train_batch_size=BATCH_SIZE,
-    per_device_eval_batch_size=1,
-    gradient_accumulation_steps=GRADIENT_ACCUMULATION,
-    learning_rate=LEARNING_RATE,
-    fp16=True,              # FP16 training
-    bf16=False,
-    save_strategy="epoch",
-    eval_strategy="epoch",
-    save_total_limit=2,
-    logging_steps=10,
-    warmup_steps=100,
-    weight_decay=0.01,
-    lr_scheduler_type="cosine",
-    seed=42,
-    dataloader_num_workers=4,
-    report_to="tensorboard",
-    load_best_model_at_end=True,
-)
-
-# =============================================================================
-# Initialize Trainer
-# =============================================================================
-trainer = SFTTrainer(
-    model=model,
-    tokenizer=tokenizer,
-    train_dataset=dataset["train"],
-    eval_dataset=dataset["valid"],
-    dataset_text_field="conversations",
-    max_seq_length=MAX_SEQ_LENGTH,
-    args=training_args,
-)
-
-# =============================================================================
-# Train!
-# =============================================================================
-print("\n" + "="*50)
-print("STARTING TRAINING - FP16 Mode")
-print("="*50)
-print(f"Model: {MODEL_NAME}")
-print(f"Trainable params: {model.num_parameters(only_trainable=True):,}")
-print(f"Epochs: {NUM_EPOCHS}")
-print(f"Batch size: {BATCH_SIZE} x {GRADIENT_ACCUMULATION} = {BATCH_SIZE * GRADIENT_ACCUMULATION}")
-print("="*50 + "\n")
-
-trainer.train()
-
-# =============================================================================
-# Save Final Model
-# =============================================================================
-print("\nSaving model...")
-trainer.save_model(f"{OUTPUT_DIR}/final")
-tokenizer.save_pretrained(f"{OUTPUT_DIR}/final")
-
-print("\n" + "="*50)
-print("TRAINING COMPLETE!")
-print("="*50)
-print(f"Model saved to: {OUTPUT_DIR}/final")
-TRAINEOF
-
-chmod +x train_blender.py
-echo "✓ Training script created"
-
-# =============================================================================
-# Step 4: Instructions
-# =============================================================================
-echo "[4/5] Setup complete!"
-echo ""
-echo "=========================================="
-echo "NEXT STEPS:"
-echo "=========================================="
-echo ""
-echo "1. Upload train.jsonl and valid.jsonl to RunPod (via JupyterLab drag & drop)"
-echo "2. Run training with:"
-echo "   python train_blender.py"
-echo ""
-echo "3. Estimated time: 3-4 hours"
-echo "   Estimated cost: ~\$15-25"
-echo ""
-echo "Monitor with: tensorboard --logdir ./blender_qwen_output"
-echo ""
-echo "=========================================="