AMD x GPU MODE — E2E Model Speedrun

Team submission for the AMD x GPU MODE E2E Model Speedrun competition.

• Phase 1 deadline: April 6, 2026
• Hardware: AMD Instinct MI355X (CDNA3)
• Prize pool: $1.1M total — top 10 Phase 1 teams advance to finals
• Official challenge folder: gpu-mode/reference-kernels/problems/amd_202602

Phase 1 Challenges

Folder	Challenge	Model	Owner
mixed-mla/	MLA Decode (mixed FP8/MXFP4 KV)	DeepSeek R1	Person A
mxfp4-mm/	MXFP4 Matrix Multiply	—	Person B
moe-mxfp4/	MXFP4 MoE Fused Kernel	DeepSeek R1	Person B

Each challenge: implement custom_kernel(data) in submission.py to beat the AITER reference.

Quick Start

# 1. Clone this repo on your GPU machine
git clone git@github.com:dorhuri123/AMD_optimization_challenge.git
cd AMD_optimization_challenge

# 2. Run the setup script (installs ROCm deps, AITER, popcorn-cli, copies reference files)
bash setup.sh

# 3. Verify the environment
python scripts/verify_env.py

# 4. Work on a challenge
cd mixed-mla
python reference.py    # run the AITER baseline
python benchmark.py    # compare reference vs your submission.py

Submission

popcorn login   # once

# Submit from inside a challenge folder
cd mixed-mla
popcorn submit submission.py --problem mixed-mla

cd mxfp4-mm
popcorn submit submission.py --problem mxfp4-mm

cd moe-mxfp4
popcorn submit submission.py --problem moe-mxfp4

Repository Structure

AMD_optimization_challenge/
├── setup.sh                   # one-shot environment setup + copies reference files
├── eval.py                    # fetched by setup.sh — official eval harness
├── utils.py                   # fetched by setup.sh — shared test utilities
├── scripts/
│   ├── verify_env.py          # check ROCm, AITER, FP8, popcorn-cli
│   └── benchmark_all.py       # run all 3 challenge benchmarks
├── mixed-mla/
│   ├── README.md              # architecture details + optimization checklist
│   ├── reference.py           # fetched by setup.sh — AITER a8w8 baseline (do not modify)
│   ├── task.py                # fetched by setup.sh — input_t / output_t / TestSpec
│   ├── task.yml               # fetched by setup.sh — benchmark configs
│   ├── submission.py          # ← EDIT THIS — custom_kernel() implementation
│   └── benchmark.py           # ref vs submission latency table
├── mxfp4-mm/
│   ├── README.md
│   ├── reference.py           # fetched by setup.sh
│   ├── task.py                # fetched by setup.sh
│   ├── task.yml               # fetched by setup.sh
│   ├── submission.py          # ← EDIT THIS
│   └── benchmark.py
└── moe-mxfp4/
    ├── README.md
    ├── reference.py           # fetched by setup.sh
    ├── task.py                # fetched by setup.sh
    ├── task.yml               # fetched by setup.sh
    ├── submission.py          # ← EDIT THIS
    └── benchmark.py

reference.py, task.py, task.yml, eval.py, utils.py are gitignored — fetched from gpu-mode/reference-kernels by setup.sh.

Key Resources

• Official challenge folder
• ROCm/aiter — AMD's optimized AI operators (AITER)
• AITER MLA decode tutorial
• vLLM ROCm attention backends
• MXFP4 quantization on AMD
• vLLM PR #36297 — fused BMM+FP8 for MLA (5.2–6.6x on MI300X)
• GPU MODE Discord: #amd-competition channel

Profiling Cheatsheet

# Quick kernel timing
rocprof --stats python submission.py

# Deep hardware counter analysis
omniperf profile --name run1 -- python submission.py
omniperf analyze -p workloads/run1/

# Memory bandwidth baseline
rocm-bandwidth-test