V0.12.0 Release Plan

[guoshzhao]

Release Manager

@guoshzhao

Endgame

• Code freeze: July. 02, 2025
• Bug Bash date: July. 17, 2025
• Release date: July. 18, 2025

Main Features

SuperBench Improvement

1. • Update Image Build Pipeline (CI/CD - Update Image Build Pipeline #659)
2. • Add support for arm64 build (Dockerfile - Add support for arm64 build #660)
3. • Upgrade dependency versions in pipeline (CI/CD - Upgrade dependency versions in pipeline #671)
4. • Fix installation and lint issues (Setup - Fix installation and lint issues #684)
5. • Update Flake8 repo (Flake8 repo from github #683)
6. • Init latest python support. (Init latest python support. #687)
7. • Add image build on arm64 arch (CI/CD - Add image build on arm64 arch #690)
8. • Enhancement of ignoring errors for import pkg_resources (Lint: Enhancement of ignoring errors for import pkg_resources #692)
9. • Update label in the ROCm image build (CI/CD - Update label in the ROCm image build #693)
10. • Support cuda12.8 for Blackwell arch (Dockerfile - Support cuda12.8 for Blackwell arch #682)
11. • Merge multi-arch image (CI/CD - Merge multi-arch image #696)
12. • Update OS of runner to the latest. (CI/CD - Update OS of runner to the latest. #702)
13. • cuda arch flag for cublaslt (cuda arch flag for cublaslt #701)

Micro-benchmark Improvement

1. • Bug Fix - Fix numa error on grace cpu in gpu-copy (Bug Fix - Fix numa error on grace cpu in gpu-copy #658)
2. • Dependency - Bump onnxruntime-gpu version from 1.10.0 to 1.12.0 (Dependency - Bump onnxruntime-gpu version from 1.10.0 to 1.12.0 #663)
3. • Benchmarks: micro benchmarks - add general CPU bandwidth and latency benchmark (Benchmarks: micro benchmarks - add general CPU bandwidth and latency benchmark #662)
4. • Benchmarks: micro benchmarks - add nvbandwidth build and benchmark (Benchmarks: micro benchmarks - add nvbandwidth build #665 and Benchmarks: micro benchmarks - add nvbandwidth benchmark #669)
5. • Fix stderr message in gpu-copy benchmark (Bug Fix - Fix stderr message in gpu-copy benchmark #673)
6. • Add arch support for 10.0 in gemm-flops (Microbenchmark: gemm-flops - Add arch support for 10.0 #680)
7. • Fix tensorrt-inference parsing (Bug - Fix tensorrt-inference parsing #674)
8. • nvbandwidth benchmark need to handle N/A value (Bugfix - nvbandwidth benchmark need to handle N/A value #675)
9. • Avoid Unintended nvbandwidth Function Calls in All Benchmarks (Bugfix: Avoid Unintended nvbandwidth Function Calls in All Benchmarks #685)
10. • Add GPU Stream Micro Benchmark (Benchmarks - Add GPU Stream Micro Benchmark #697)
11. • Cuda arch flag for cublaslt (cuda arch flag for cublaslt #701)
12. • Support autotuning in cublaslt gemm (Benchmark - Support autotuning in cublaslt gemm #706)
13. • Add FP4 GEMM FLOPS support for cublaslt_gemm benchmark (Benchmarks - Add FP4 GEMM FLOPS support for cublaslt_gemm benchmark #711)
14. • CPU Stream Benchmark Revise (microbenchmark - CPU Stream Benchmark Revise #712)
15. • Add cuda12.9 docker image (Dockerfile - Add cuda12.9 docker image #716)
16. • Add Grace CPU support for CPU Stream (Benchmark - Add Grace CPU support for CPU Stream. #719)

Model Benchmark Improvement

1. • Add LLaMA-2 Models (Benchmarks - Add LLaMA-2 Models #668)
2. • Fix typos in documentation and code files (Docs - Fix typos in documentation and code files #686)
3. • Add Mixture of Experts Model (Benchmarks: Add Mixture of Experts Model  #679)
4. • Add DeepSeek Training Benchmark
5. • Add DeepSeek Inference Benchmark (AMD GPU) (Benchmarks - Add deepseek megatron-lm benchmark #713)

Documentation

1. • Update CODEOWNERS (Docs - Update CODEOWNERS #670)
2. • Update CODEOWNERS (Update CODEOWNERS #718)

Result Analysis

1. • Enhance logging information for diagnosis rule op baseline errors. (Refactor: Enhance logging information for diagnosis rule op baseline errors. #689)