DeepSpeed Fine-tuning Skill#

This skill enables efficient model fine-tuning using DeepSpeed with various optimization strategies.

Prerequisites#

• Python 3.8+
• GPU(s) or accelerator(s) with DeepSpeed-supported backend (CUDA, ROCm, Intel XPU, etc.)
• DeepSpeed: pip install deepspeed
• Transformers, Datasets, PEFT (for LoRA support)
• sshpass: sudo apt-get install sshpass (for remote training)

Plan Selection Workflow#

Never auto-select a plan. List viable options based on user hardware and requirements, and let the user decide.

Step 1: Gather Information#

Confirm the following with the user:

• Target model: Model name and parameter count (e.g., Qwen2.5-7B)

• Hardware environment:

  • GPU VRAM x count (e.g., "single 24GB GPU")
  • CPU core count
  • RAM size
  • Free disk space
  • NVMe SSD availability (affects ZeRO NVMe offload)

• Training goal: Full fine-tuning or parameter-efficient? Dataset size? Expected quality?

• Budget/time constraints: Acceptable training duration?

If the user only provides an SSH or remote machine address, connect first and auto-detect hardware ( nvidia-smi, free -h, df -h, nproc).

Step 2: Evaluate Feasibility#

Estimate VRAM requirements based on model size (bf16):

ParamsModel Weights (bf16)+ Adam Optimizer + Gradients0.5B1 GB5 GB1.5B3 GB15 GB3B6 GB30 GB7B14 GB70 GB14B28 GB140 GB32B64 GB320 GB72B144 GB720 GB

Breakdown: Adam optimizer stores 2 fp32 state tensors (momentum + variance) = 8 bytes/param. Gradients = 2 bytes/param (bf16). Total approx. 10 bytes/param (5x model weight size).

Activation memory: Depends on sequence length and batch size, not model params alone.

• Formula: activation approx. 34 x seq_len x hidden_size x batch_size x bytes_per_element
• Example: 7B model (hidden=4096), seq_len=2048, batch_size=4, bf16 -> ~1.5 GB per layer; ~60 GB total (can dominate VRAM)
• Gradient checkpointing reduces this by ~80% (recomputes instead of storing), but adds ~20% compute overhead
• Rule of thumb: if seq_len x batch_size > 8192, activation memory likely exceeds model weights

LoRA/QLoRA: VRAM depends on rank, target modules, and layer dimensions — not directly proportional to total model params. See references/lora_guide.md for LoRA-specific memory estimation.

Step 2.5: Activation Checkpointing#

If VRAM is tight, activation checkpointing is the most impactful knob — it can reduce activation memory by ~80%.

How it works: Instead of storing all intermediate activations for backprop, only save checkpoints at select layers. Remaining activations are recomputed during backward pass. Trades compute for memory.

Two ways to enable:

1. HF Trainer flag (simplest, works out of the box):

bash

python scripts/ds_train.py --gradient_checkpointing ...

1. DeepSpeed config (fine-grained control):

json

{
  "activation_checkpointing": {
    "partition_activations": true,
    "cpu_checkpointing": true,
    "contiguous_memory_optimization": true,
    "number_checkpoints": 4
  }
}

OptionEffectWhen to usepartition_activationsShard checkpoints across model-parallel GPUsMulti-GPU with model parallelismcpu_checkpointingStore checkpoints in CPU RAM instead of GPUGPU memory very tightcontiguous_memory_optimizationReduce memory fragmentationLarge models, many checkpointsnumber_checkpointsControl checkpoint frequency (fewer = less VRAM, more compute)Tune based on VRAM budget

Step 3: List Options#

Based on the VRAM assessment, list all viable approaches. Example:

text

Based on your hardware (single 24GB GPU, 64GB RAM, 500GB disk),
Qwen2.5-7B has these training options:

Option A: LoRA Fine-tuning (Recommended)
  - VRAM needed: ~22 GB
  - Speed: Fast
  - Quality: Good for instruction alignment, style adaptation
  - Trainable params: ~20M (0.4% of total)

Option B: QLoRA Fine-tuning (Saves VRAM)
  - VRAM needed: ~12 GB
  - Speed: Medium (quantization/dequantization overhead)
  - Quality: Slightly below LoRA, but gap is small

Option C: Full Fine-tuning (Not feasible)
  - VRAM needed: ~56 GB (exceeds 24GB)
  - Requires ZeRO-2 + CPU offload, or larger GPU

Which option do you prefer?

Step 4: Hardware Insufficient? Make Recommendations#

If no plan is viable on current hardware, recommend specs using generic hardware metrics (no brand names):

text

You want to fully fine-tune a 7B model, but current hardware (single 24GB GPU) is insufficient.
Recommended hardware specs:

Minimum:
  - GPU: single 80GB VRAM
  - CPU: 16+ cores
  - RAM: 128 GB+
  - Disk: 200 GB+ free space

Recommended:
  - GPU: 2x 80GB VRAM (ZeRO-2 doubles training speed)
  - CPU: 32+ cores
  - RAM: 256 GB+
  - Disk: 500 GB+ free space

Alternatively, use LoRA — 24GB VRAM is sufficient for 7B models.

Key Principles#

• Never auto-select and start training — always list options and wait for user confirmation
• Recommend but don't decide — say "I recommend Option A because..." but let the user choose
• Use generic hardware metrics — VRAM in GB, GPU count, CPU cores, RAM in GB, disk in GB. No brand names.
• Leave VRAM headroom — recommend at least 20% buffer to avoid OOM
• If user picks an infeasible option, warn them clearly rather than silently switching

Core Capabilities#

1. Training Configuration#

Generate DeepSpeed ZeRO configurations:

python

from scripts.generate_ds_config import generate_zero_config

# ZeRO Stage 2 with optimizer offloading
config = generate_zero_config(
    zero_stage=2,
    offload_optimizer=True,
    offload_device="nvme",
    nvme_path="/local_nvme"
)

2. Training Launch#

Use the training launcher script:

bash

python scripts/ds_train.py \
  --model_name_or_path meta-llama/Llama-2-7b-hf \
  --dataset_path data/my_dataset \
  --output_dir ./outputs \
  --deepspeed assets/ds_config_zero2.json \
  --num_train_epochs 3 \
  --per_device_train_batch_size 4 \
  --learning_rate 2e-5 \
  --lora_r 16 \
  --lora_alpha 32

3. LoRA/QLoRA Integration#

For parameter-efficient fine-tuning:

python

# LoRA config is auto-generated based on arguments
peft_config = {
    "peft_type": "LORA",
    "r": 16,
    "lora_alpha": 32,
    "target_modules": ["q_proj", "v_proj", "k_proj", "o_proj"],
    "lora_dropout": 0.05,
    "bias": "none",
    "task_type": "CAUSAL_LM"
}

4. Multi-GPU Training#

Use the deepspeed launcher for multi-GPU training (recommended over torchrun):

bash

# Multi-GPU on single node
deepspeed --num_gpus=4 scripts/ds_train.py \
  --model_name_or_path meta-llama/Llama-2-7b-hf \
  --deepspeed assets/ds_config_zero3.json \
  ...

# Multi-node
deepspeed --hostfile hosts.txt scripts/ds_train.py \
  --model_name_or_path meta-llama/Llama-2-7b-hf \
  --deepspeed assets/ds_config_zero3.json \
  ...

5. Training Monitoring#

Monitor training progress:

python

from scripts.monitor_training import TrainingMonitor

monitor = TrainingMonitor(log_dir="./outputs")
monitor.plot_loss()
monitor.get_latest_checkpoint()

6. Early Stopping#

Automatically monitors eval loss and stops training early when there's no improvement across consecutive evaluations, then loads the best checkpoint.

Parameters:

• --early_stopping_patience — How many consecutive evals without improvement to tolerate. Set to 0 to disable (default). Recommended: 3-10.
• --early_stopping_threshold — Minimum eval loss improvement to count as an improvement. Default 0.0 (any decrease counts).

Example:

bash

python scripts/ds_train.py \
  --model_name_or_path Qwen/Qwen2.5-0.5B \
  --dataset_path tatsu-lab/alpaca \
  --use_peft True \
  --early_stopping_patience 5 \
  --early_stopping_threshold 0.001 \
  --eval_strategy steps \
  --eval_steps 100 \
  --num_train_epochs 3 \
  ...

Auto-configuration: When early_stopping_patience > 0, the script automatically:

1. Enables load_best_model_at_end=True
2. Sets metric_for_best_model=eval_loss, greater_is_better=False
3. Aligns save_strategy with eval_strategy (synced saving is needed to restore best checkpoint)

Notes:

• Must also set eval_strategy (e.g., steps + eval_steps), otherwise early stopping won't work
• Don't set patience too low (<3) — early training fluctuations may cause premature stopping
• For LoRA fine-tuning, patience=5 with eval_steps=100 typically works well

Remote Training#

When training needs to run on a remote GPU server, see references/remote_training.md for the complete guide including agent guidelines, security model, and command reference.

Troubleshooting#

OOM Errors#

• Reduce batch size or increase gradient accumulation steps
• Enable gradient checkpointing: --gradient_checkpointing
• Use ZeRO-3 with CPU/NVMe offloading
• Reduce LoRA rank: --lora_r 8
• See references/troubleshooting.md for detailed solutions

Slow Training#

• Ensure bf16/fp16 is enabled
• Check GPU utilization with nvidia-smi
• Use FlashAttention if available
• Optimize data loading with --dataloader_num_workers
• See references/troubleshooting.md for detailed solutions

Checkpoint Issues#

• Use --save_strategy steps with --save_steps
• Enable --save_total_limit to cap checkpoint count
• For ZeRO-3, use --zero3_save_16bit_model to save FP16 weights
• See references/troubleshooting.md for detailed solutions

MPI Errors (multi-GPU only)#

• Single-GPU training does not need MPI
• If you see MPI errors on single GPU, use python3 directly instead of deepspeed launcher
• See references/troubleshooting.md for full MPI debugging guide

Single-GPU Strategy#

• See references/single_gpu_strategy.md for strategy selection, CPU/NVMe offload examples, and decision principles

References#

• Quick Start Guide — Common training patterns and full examples
• DeepSpeed Guide — DeepSpeed documentation and configuration reference
• LoRA/PEFT Best Practices — LoRA/QLoRA parameter tuning guide
• ZeRO Optimization Guide — ZeRO stage comparison and optimization tips
• Single-GPU Strategy — Strategy selection for single-GPU training
• Remote Training Guide — Remote training via SSH, agent guidelines, and security model
• Troubleshooting — Common errors and solutions (OOM, NaN loss, MPI, NCCL, etc.)