1️⃣ Log In

Create your account or sign in. Your session is remembered so you stay logged in.

2️⃣ Add API Keys

Add your HuggingFace token (required), plus Vast.ai or RunPod key for cloud GPU access.

3️⃣ Pick a Model & GPU

Choose your model. GPUs with a ⭐ star have enough memory for full fine-tuning.

4️⃣ Configure & Generate

Review the Training Config tab. Click Create Package — EzEpoch resolves all conflicts automatically.

5️⃣ Deploy to Cloud

Use the Deployment Wizard to rent a GPU, upload your package, and start training — all from the browser.

The 4 HuggingFace token permission levels

🤗 How to create a HuggingFace token (required)

1. Go to huggingface.co/settings/tokens
2. Click + Create new token
3. Give it a name (e.g. "EzEpoch")
4. Choose a permission level:
   • Read — download public & gated models (recommended for most users)
   • Write — upload fine-tuned models back to HF Hub
   • Fine-grained — custom per-repo permissions
   • Legacy — older token format (avoid for new tokens)
5. Click Generate token and copy it (starts with hf_)
6. Paste it into the HuggingFace Token field in EzEpoch and click Test & Save

Other API keys

• Vast.ai — get your key from cloud.vast.ai/account
• RunPod — get your key from runpod.io → Settings
• GitHub — optional, for repo access. Use a token with repo scope.
• All keys are encrypted before storage. Use 👁 to reveal, or Clear to remove.

Default view — clean and simple

All settings expanded for full control

⭐ GPU Star System

Stars appear on GPUs with enough VRAM for full fine-tuning with AdamW. No star = use LoRA/QLoRA.

🤖 Auto-Optimization

Incompatible settings are greyed out automatically. The system picks the best defaults for your model + GPU combo.

📁 Dataset Upload

Upload a JSONL file or paste a HuggingFace dataset path. The system validates format automatically.

🔢 Batch Size

Set manually or use Auto — the AI probes your GPU and finds the largest safe batch size.

Advanced tab — precision & optimizer settings

Expert tab — scheduler, warmup, weight decay

🎯 Precision

bf16 (recommended for Ampere+), fp16, or fp32. BF16 uses less memory with no accuracy loss.

⚡ Optimizer

AdamW (default), 8-bit Adam (saves memory), Adafactor (very low memory), or SGD.

📉 Scheduler

Cosine, linear, constant, or warmup variants. Cosine with warmup is recommended for most fine-tuning.

💾 Gradient Checkpointing

Reduces VRAM by recomputing activations during backward pass. Slightly slower but essential for large models.

Step 1 — Select GPU & provider

Step 2 — Review & confirm

Step 3 — Launch & monitor

Step 1 — Choose GPU

Pick from Vast.ai or RunPod. Filter by VRAM required, price per hour, and CUDA version.

Step 2 — Review & Confirm

See the full cost estimate, instance specs, and package contents before committing. Estimated training time shown.

Step 3 — Auto-Launch

EzEpoch rents the GPU, uploads your package, and starts training. Live progress streams to your Dashboard.

Embedded dashboard — inside EzEpoch

Standalone dashboard — dashboard.ezepoch.com

📈 Live Loss Curve

Training and validation loss plotted in real time. Divergence is flagged automatically.

🖥️ GPU Stats

VRAM usage, GPU utilisation %, temperature, and power draw — all streaming live.

📜 Live Logs

Full training stdout/stderr streamed to the browser. Click any line to jump to that epoch.

🌐 Standalone

Visit dashboard.ezepoch.com to monitor from any device.

Core Settings

Epochs

Full passes through your training dataset

Batch size

Samples processed per gradient update step

Learning rate

Step size for weight updates (1e-4 to 1e-5 typical for fine-tuning)

Max seq len

Maximum token length per sample — longer = more VRAM

Grad accum

Accumulate N gradients before updating — simulates larger batch sizes

Training Methods

Full Fine-tune

All model weights updated. Best quality, needs most VRAM.

LoRA

Trains small adapters alongside frozen base model. ~10-30% VRAM vs full.

QLoRA

LoRA on a 4-bit quantised base. Run 70B models on a single 24GB GPU.

DoRA

Direction + magnitude LoRA variant. Often better than standard LoRA.

Multi-GPU Strategies

DDP

Data Parallel. Each GPU holds a full model copy. Simplest, best throughput.

ZeRO-2

Shards optimizer states + gradients. Cuts VRAM ~50% vs DDP.

ZeRO-3

Also shards model weights. Required for very large models across GPUs.

FSDP

PyTorch native sharding. Good for models too large for a single GPU.

LoRA Parameters

Rank (r)

Adapter size. Higher = more capacity but more VRAM. 8-64 typical.

Alpha

Scales the LoRA update. Usually set equal to rank or 2x rank.

Dropout

Regularisation for adapter layers. 0.05-0.1 typical.

Target modules

Which layers get adapters. EzEpoch auto-selects the optimal set.