Add NVIDIA Model Optimizer FP8 quantization recipe (single-GPU EC2)

[mvinci12]

Purpose

Adds a new test case demonstrating NVIDIA Model Optimizer (ModelOpt) FP8 post-training quantization on a single GPU EC2 instance, served with vLLM. Unlike the distributed-training test cases, ModelOpt PTQ is a single-process, single-GPU workload, so the recipe provisions one instance with Terraform and connects via SSM Session Manager (no SSH).

Changes

• New test case: 3.test_cases/pytorch/model-optimizer/
  • terraform/ — single g6e.xlarge (L40S) instance: SSM IAM role + instance profile, egress-only security group, GPU DLAMI resolved from SSM Parameter Store, gp3 root volume, IMDSv2 enforced. Connects via SSM Session Manager (no inbound SSH, no key pair).
  • src/ — pinned setup.sh (nvidia-modelopt==0.44.0), quantize_fp8.sh (FP8 PTQ via examples/llm_ptq/hf_ptq.py, weights-only, cnn_dailymail calibration), smoke_test_vllm.py, pinned requirements.txt.
  • README.md — EC2-vs-EKS rationale, format↔GPU-architecture matrix, full walkthrough, reference results, and known issues.

Test Plan

Environment:

• AWS Service: EC2 (single instance, provisioned via the recipe's Terraform)
• Instance type: g6e.xlarge (1× NVIDIA L40S, 48 GB, Ada)
• Number of nodes: 1
• Region: us-west-2

Test commands:

# Infra (verified apply -> SSM -> nvidia-smi -> destroy)
cd 3.test_cases/pytorch/model-optimizer/terraform
terraform init && terraform validate && terraform plan
terraform apply -auto-approve
aws ssm start-session --target <instance-id>     # then: nvidia-smi
terraform destroy -auto-approve

# Workload (verified on the L40S)
cd ../src
bash setup.sh
bash quantize_fp8.sh
python smoke_test_vllm.py --model ~/qwen2.5-7b-fp8

Test Results

• terraform init/validate/plan/apply/destroy all succeeded against a real account; AMI resolved to ami-03ee6edb81d0951b9; SSM came online ~30 s after launch; nvidia-smi reported NVIDIA L40S, 46068 MiB, 595.71.05. terraform destroy removed all 5 resources with zero left billable.
• ModelOpt 0.44.0 installed in ~90 s; FP8 PTQ of Qwen/Qwen2.5-7B-Instruct (weights-only, cnn_dailymail) calibrated + exported in ~45 s; peak GPU memory ~16.5 GB.
• Checkpoint size: 15 GB BF16 → 8.2 GB FP8 (~1.8×).
• vLLM 0.23.0 loaded the checkpoint (CutlassFP8ScaledMMLinearKernel) and produced coherent output:
  • What is the capital of France? → The capital of France is Paris.

Documented known issue surfaced during testing: FP8 KV-cache produces garbled output on Ada (L40S) → recipe defaults to weights-only (--kv_cache_qformat none).

Directory Structure

3.test_cases/pytorch/model-optimizer/
├── README.md
├── terraform/
│   ├── versions.tf
│   ├── providers.tf
│   ├── variables.tf
│   ├── main.tf
│   ├── outputs.tf
│   └── README.md
└── src/
    ├── requirements.txt
    ├── setup.sh
    ├── quantize_fp8.sh
    └── smoke_test_vllm.py

Checklist

• I have read the contributing guidelines.
• I am working against the latest main branch.
• I have searched existing open and recently merged PRs to confirm this is not a duplicate.
• The contribution is self-contained with documentation and scripts.
• External dependencies are pinned to a specific version or tag (no latest).
• A README is included with prerequisites, instructions, and known issues.
• New test cases follow the expected directory structure.

---

Draft: opening for early review. The EC2 path is live-tested; the EKS/HyperPod graduation path is documented but intentionally out of scope for this PR.