trl/references/hardware_guide.md

Hardware Selection Guide

Choosing the right hardware (flavor) is critical for cost-effective training.

Available Hardware

CPU

• cpu-basic - Basic CPU, testing only
• cpu-upgrade - Enhanced CPU

Use cases: Dataset validation, preprocessing, testing scripts Not recommended for training: Too slow for any meaningful training

GPU Options

Flavor	GPU	Memory	Use Case	Cost/hour
t4-small	NVIDIA T4	16GB	<1B models, demos	~$0.50-1
t4-medium	NVIDIA T4	16GB	1-3B models, development	~$1-2
l4x1	NVIDIA L4	24GB	3-7B models, efficient training	~$2-3
l4x4	4x NVIDIA L4	96GB	Multi-GPU training	~$8-12
a10g-small	NVIDIA A10G	24GB	3-7B models, production	~$3-4
a10g-large	NVIDIA A10G	24GB	7-13B models	~$4-6
a10g-largex2	2x NVIDIA A10G	48GB	Multi-GPU, large models	~$8-12
a10g-largex4	4x NVIDIA A10G	96GB	Multi-GPU, very large models	~$16-24
a100-large	NVIDIA A100	40GB	13B+ models, fast training	~$8-12

TPU Options

Flavor	Type	Use Case
v5e-1x1	TPU v5e	Small TPU workloads
v5e-2x2	4x TPU v5e	Medium TPU workloads
v5e-2x4	8x TPU v5e	Large TPU workloads

Note: TPUs require TPU-optimized code. Most TRL training uses GPUs.

Selection Guidelines

By Model Size

Tiny Models (<1B parameters)

• Recommended: t4-small
• Example: Qwen2.5-0.5B, TinyLlama
• Batch size: 4-8
• Training time: 1-2 hours for 1K examples

Small Models (1-3B parameters)

• Recommended: t4-medium or a10g-small
• Example: Qwen2.5-1.5B, Phi-2
• Batch size: 2-4
• Training time: 2-4 hours for 10K examples

Medium Models (3-7B parameters)

• Recommended: a10g-small or a10g-large
• Example: Qwen2.5-7B, Mistral-7B
• Batch size: 1-2 (or LoRA with 4-8)
• Training time: 4-8 hours for 10K examples

Large Models (7-13B parameters)

• Recommended: a10g-large or a100-large
• Example: Llama-3-8B, Mixtral-8x7B (with LoRA)
• Batch size: 1 (full fine-tuning) or 2-4 (LoRA)
• Training time: 6-12 hours for 10K examples
• Note: Always use LoRA/PEFT

Very Large Models (13B+ parameters)

• Recommended: a100-large with LoRA
• Example: Llama-3-13B, Llama-3-70B (LoRA only)
• Batch size: 1-2 with LoRA
• Training time: 8-24 hours for 10K examples
• Note: Full fine-tuning not feasible, use LoRA/PEFT

By Budget

Minimal Budget (<$5 total)

• Use t4-small
• Train on subset of data (100-500 examples)
• Limit to 1-2 epochs
• Use small model (<1B)

Small Budget ($5-20)

• Use t4-medium or a10g-small
• Train on 1K-5K examples
• 2-3 epochs
• Model up to 3B parameters

Medium Budget ($20-50)

• Use a10g-small or a10g-large
• Train on 5K-20K examples
• 3-5 epochs
• Model up to 7B parameters

Large Budget ($50-200)

• Use a10g-large or a100-large
• Full dataset training
• Multiple epochs
• Model up to 13B parameters with LoRA

By Training Type

Quick Demo/Experiment

• t4-small
• 50-100 examples
• 5-10 steps
• ~10-15 minutes

Development/Iteration

• t4-medium or a10g-small
• 1K examples
• 1 epoch
• ~30-60 minutes

Production Training

• a10g-large or a100-large
• Full dataset
• 3-5 epochs
• 4-12 hours

Research/Experimentation

• a100-large
• Multiple runs
• Various hyperparameters
• Budget for 20-50 hours

Memory Considerations

Estimating Memory Requirements

Full fine-tuning:

Memory (GB) ≈ (Model params in billions) × 20

LoRA fine-tuning:

Memory (GB) ≈ (Model params in billions) × 4

Examples:

• Qwen2.5-0.5B full: ~10GB ✅ fits t4-small
• Qwen2.5-1.5B full: ~30GB ❌ exceeds most GPUs
• Qwen2.5-1.5B LoRA: ~6GB ✅ fits t4-small
• Qwen2.5-7B full: ~140GB ❌ not feasible
• Qwen2.5-7B LoRA: ~28GB ✅ fits a10g-large

Memory Optimization

If hitting memory limits:

1. Use LoRA/PEFT

   peft_config=LoraConfig(r=16, lora_alpha=32)
   

2. Reduce batch size

   per_device_train_batch_size=1
   

3. Increase gradient accumulation

   gradient_accumulation_steps=8  # Effective batch size = 1×8
   

4. Enable gradient checkpointing

   gradient_checkpointing=True
   

5. Use mixed precision

   bf16=True  # or fp16=True
   

6. Upgrade to larger GPU

   • t4 → a10g → a100

Cost Estimation

Formula

Total Cost = (Hours of training) × (Cost per hour)

Example Calculations

Quick demo:

• Hardware: t4-small ($0.75/hour)
• Time: 15 minutes (0.25 hours)
• Cost: $0.19

Development training:

• Hardware: a10g-small ($3.50/hour)
• Time: 2 hours
• Cost: $7.00

Production training:

• Hardware: a10g-large ($5/hour)
• Time: 6 hours
• Cost: $30.00

Large model with LoRA:

• Hardware: a100-large ($10/hour)
• Time: 8 hours
• Cost: $80.00

Cost Optimization Tips

1. Start small: Test on t4-small with subset
2. Use LoRA: 4-5x cheaper than full fine-tuning
3. Optimize hyperparameters: Fewer epochs if possible
4. Set appropriate timeout: Don't waste compute on stalled jobs
5. Use checkpointing: Resume if job fails
6. Monitor costs: Check running jobs regularly

Multi-GPU Training

TRL automatically handles multi-GPU training with Accelerate when using multi-GPU flavors.

Multi-GPU flavors:

• l4x4 - 4x L4 GPUs
• a10g-largex2 - 2x A10G GPUs
• a10g-largex4 - 4x A10G GPUs

When to use:

• Models >13B parameters
• Need faster training (linear speedup)
• Large datasets (>50K examples)

Example:

hf_jobs("uv", {
    "script": "train.py",
    "flavor": "a10g-largex2",  # 2 GPUs
    "timeout": "4h",
    "secrets": {"HF_TOKEN": "$HF_TOKEN"}
})

No code changes needed—TRL/Accelerate handles distribution automatically.

Choosing Between Options

a10g vs a100

Choose a10g when:

• Model <13B parameters
• Budget conscious
• Training time not critical

Choose a100 when:

• Model 13B+ parameters
• Need fastest training
• Memory requirements high
• Budget allows

Single vs Multi-GPU

Choose single GPU when:

• Model <7B parameters
• Budget constrained
• Simpler debugging

Choose multi-GPU when:

• Model >13B parameters
• Need faster training
• Large batch sizes required
• Cost-effective for large jobs

Quick Reference

# Model size → Hardware selection
HARDWARE_MAP = {
    "<1B":     "t4-small",
    "1-3B":    "a10g-small",
    "3-7B":    "a10g-large",
    "7-13B":   "a10g-large (LoRA) or a100-large",
    ">13B":    "a100-large (LoRA required)"
}