Common Training Patterns

This guide provides common training patterns and use cases for TRL on Hugging Face Jobs.

Multi-GPU Training

Automatic distributed training across multiple GPUs. TRL/Accelerate handles distribution automatically:

hf_jobs("uv", {
    "script": """
# Your training script here (same as single GPU)
# No changes needed - Accelerate detects multiple GPUs
""",
    "flavor": "a10g-largex2",  # 2x A10G GPUs
    "timeout": "4h",
    "secrets": {"HF_TOKEN": "$HF_TOKEN"}
})

Tips for multi-GPU:

No code changes needed
Use per_device_train_batch_size (per GPU, not total)
Effective batch size = per_device_train_batch_size × num_gpus × gradient_accumulation_steps
Monitor GPU utilization to ensure both GPUs are being used

DPO Training (Preference Learning)

Train with preference data for alignment:

hf_jobs("uv", {
    "script": """
# /// script
# dependencies = ["trl>=0.12.0", "trackio"]
# ///

from datasets import load_dataset
from trl import DPOTrainer, DPOConfig
import trackio

trackio.init(project="dpo-training", space_id="username/trackio")

dataset = load_dataset("trl-lib/ultrafeedback_binarized", split="train")

# Create train/eval split
dataset_split = dataset.train_test_split(test_size=0.1, seed=42)

config = DPOConfig(
    output_dir="dpo-model",
    push_to_hub=True,
    hub_model_id="username/dpo-model",
    num_train_epochs=1,
    beta=0.1,  # KL penalty coefficient
    eval_strategy="steps",
    eval_steps=50,
    report_to="trackio",
)

trainer = DPOTrainer(
    model="Qwen/Qwen2.5-0.5B-Instruct",  # Use instruct model as base
    train_dataset=dataset_split["train"],
    eval_dataset=dataset_split["test"],  # IMPORTANT: Provide eval_dataset when eval_strategy is enabled
    args=config,
)

trainer.train()
trainer.push_to_hub()
trackio.finish()
""",
    "flavor": "a10g-large",
    "timeout": "3h",
    "secrets": {"HF_TOKEN": "$HF_TOKEN"}
})

For DPO documentation: Use hf_doc_fetch("https://huggingface.co/docs/trl/dpo_trainer")

GRPO Training (Online RL)

Group Relative Policy Optimization for online reinforcement learning:

hf_jobs("uv", {
    "script": "https://raw.githubusercontent.com/huggingface/trl/main/examples/scripts/grpo.py",
    "script_args": [
        "--model_name_or_path", "Qwen/Qwen2.5-0.5B-Instruct",
        "--dataset_name", "trl-lib/math_shepherd",
        "--output_dir", "grpo-model",
        "--push_to_hub",
        "--hub_model_id", "username/grpo-model"
    ],
    "flavor": "a10g-large",
    "timeout": "4h",
    "secrets": {"HF_TOKEN": "$HF_TOKEN"}
})

For GRPO documentation: Use hf_doc_fetch("https://huggingface.co/docs/trl/grpo_trainer")

Pattern Selection Guide

Use Case	Pattern	Hardware	Time
SFT training	`scripts/train_sft_example.py`	a10g-large	2-6 hours
Large dataset (>10K)	Multi-GPU	a10g-largex2	4-12 hours
Preference learning	DPO Training	a10g-large	2-4 hours
Online RL	GRPO Training	a10g-large	3-6 hours

Critical: Evaluation Dataset Requirements

⚠️ IMPORTANT: If you set eval_strategy="steps" or eval_strategy="epoch", you MUST provide an eval_dataset to the trainer, or the training will hang.

✅ CORRECT - With eval dataset:

dataset_split = dataset.train_test_split(test_size=0.1, seed=42)

trainer = SFTTrainer(
    model="Qwen/Qwen2.5-0.5B",
    train_dataset=dataset_split["train"],
    eval_dataset=dataset_split["test"],  # ← MUST provide when eval_strategy is enabled
    args=SFTConfig(eval_strategy="steps", ...),
)

❌ WRONG - Will hang:

trainer = SFTTrainer(
    model="Qwen/Qwen2.5-0.5B",
    train_dataset=dataset,
    # NO eval_dataset but eval_strategy="steps" ← WILL HANG
    args=SFTConfig(eval_strategy="steps", ...),
)

Option: Disable evaluation if no eval dataset

config = SFTConfig(
    eval_strategy="no",  # ← Explicitly disable evaluation
    # ... other config
)

trainer = SFTTrainer(
    model="Qwen/Qwen2.5-0.5B",
    train_dataset=dataset,
    # No eval_dataset needed
    args=config,
)

Best Practices

Use train/eval splits - Create evaluation split for monitoring progress
Enable Trackio - Monitor progress in real-time
Add 20-30% buffer to timeout - Account for loading/saving overhead
Test with TRL official scripts first - Use maintained examples before custom code
Always provide eval_dataset - When using eval_strategy, or set to "no"
Use multi-GPU for large models - 7B+ models benefit significantly

Datasets:

mcp-tools
/

skills_go_to_github