Preparation Checklist
Training: AWS Neuron Deep Dive (Day 1 & Day 2)
Instance: trn2.3xlarge (1 device, 4 logical cores, 96GB HBM)
1. System & Permission Requirements
Common (Day 1 & Day 2)
| Category |
Item |
Detail |
| EC2 Quota |
trn2.3xlarge available |
Service Quota: Running On-Demand Trn Instances ≥ 12 vCPU |
| EBS |
Minimum 200GB gp3 |
For model cache + NEFF compile cache |
| Security Group |
SSH (22) |
Participant access |
|
Port 8000 |
vLLM API testing |
|
Port 3001 |
Neuron Explorer Web UI |
|
Port 3002 |
Neuron Explorer API Backend |
|
Port 8888 |
Jupyter Notebook (optional) |
| Outbound |
443 (HTTPS) |
HuggingFace Hub, PyPI, ECR access |
| HuggingFace |
HF account + Access Token |
Pre-approve gated models (Llama 3.1, etc.) |
| IAM — ECR |
① Public ECR |
public.ecr.aws/neuron/... — no IAM required |
|
② Private ECR (customer account) |
ecr:GetDownloadUrlForLayer, ecr:BatchGetImage, ecr:GetAuthorizationToken |
Day 1 Only
| Category |
Item |
Detail |
| OS/AMI |
Neuron DLAMI (Ubuntu 24.04) |
ImageID:ami-098a48f87d9c99e69
Neuron driver + SDK + vLLM pre-installed |
| Model Access |
meta-llama/Llama-3.2-1B-Instruct |
Pre-approve Meta license on HuggingFace |
Day 2 Only
| Category |
Item |
Detail |
| OS/AMI |
Ubuntu 24.04 + Docker |
DLC container-based labs |
| Docker |
Installed + participant in docker group |
usermod -aG docker <user> |
| DLC Image |
PyTorch Native DLC (Private Beta) |
Image URI shared separately. Pre-pull recommended (large size) |
| Additional Disk |
EBS total 300GB+ recommended |
NKI artifacts, NEFF cache, profile data accumulation |
2. Participant Prerequisites
Before Day 1
| Item |
Required/Recommended |
Description |
| Docker basics |
Required |
pull, run, exec, volume mount |
| Linux CLI |
Required |
SSH connection, basic commands (ls, cd, cat, grep) |
| ML/DL fundamentals |
Required |
Model, inference, parameters, tokens |
| AWS EC2 basics |
Required |
Instance types, SSH access, Security Groups |
| LLM concepts |
Recommended |
Transformer, attention, tokenizer |
| GPU-based inference experience |
Recommended |
vLLM or TGI experience is helpful |
| HuggingFace usage |
Recommended |
Model hub, config.json, tokenizer |
Before Day 2
| Item |
Required/Recommended |
Description |
| Day 1 completion |
Required |
Or equivalent Neuron fundamentals + vLLM serving experience |
| PyTorch experience |
Required |
model.to(), forward/backward, DataLoader |
| torch.compile concepts |
Recommended |
Graph tracing, FX Graph basics |
| GPU kernel concepts |
Recommended |
CUDA/Triton experience helps with NKI understanding |
| Profiler experience |
Recommended |
torch.profiler or Nsight experience |
Learning Path
For self-study before Day 1 based on official documentation, see the Learning Path page.
3. Summary Comparison
|
Day 1 |
Day 2 |
| Theme |
Neuron Fundamentals & Inference Operations |
PyTorch Native & NKI Optimization |
| OS |
Neuron DLAMI |
Ubuntu + Docker |
| Environment |
DLAMI bare metal (vLLM built-in) |
PyTorch Native DLC (Private Beta) |
| Docker |
Not required |
Required |
| Lab |
vLLM deployment → monitoring → tuning |
NKI kernel writing → benchmark → profiling |