Release Manager

@cp5555

Endgame

• Code freeze: March 28th, 2023
• Bug Bash date: March 29th, 2023
• Release date: April 7th, 2023

Main Features

SuperBench Improvement

1. • Support SuperBench Executor running on Windows (Executor - Support SuperBench Executor running on Windows #475)
2. • Remove fixed rccl version in rocm5.1.x docker file (Dockerfile: Remove fixed rccl version in rocm5.1.x docker file #476)
3. • Upgrade networkx version to fix installation compatibility issue (CI/CD - Upgrade networkx version to fix installation compatibility issue #478)
4. • Pin setuptools version to v65.7.0 (Pin setuptools version to v65.7.0 #483)
5. • Limit ansible_runner version for Python3.6 (Limit ansible_runner version for Python3.6 #485)
6. • Support cgroup V2 when read system metrics in Monitor (Monitor - Support cgroup V2 when read system metrics. #491, Monitor - Fix the cgroup version checking logic. #502)
7. • Fix analyzer bug in python3.8 due to pandas api change (Analyzer: Fix bug in python3.8 due to pandas api change #504)
8. • Collect real-time GPU power in Monitor (Monitor - Collect realtime GPU power when benchmarking. #507)
9. • Remove unreachable condition when write host list (Remove unreachable condition when write host list #512)
10. • Update to cuda12.1, nccl 2.17.1, hpcx 2.14, and mlc 3.10 (Update to cuda12.1, nccl 2.17.1, hpcx 2.14, and mlc 3.10 #513)
11. • Fix wrong unit of cpu-memory-bw-latency in doc (Doc - Fix wrong unit of cpu-memory-bw-latency in doc #515)

Micro-benchmark Improvement

1. • Add STREAM benchmark for sustainable memory bandwidth and the corresponding computation rate. (Adding Stream Benchmark #473)
2. • Add HPL Benchmark for HPC Linpack Benchmark. (Adding HPL benchmark #482)
3. • Support flexible warmup and non-random data initialization in cublas-benchmark (Benchmarks: Revision - Support flexible warmup and non-random data initialization in cublas-benchmark  #479)
4. • Support error tolerance in micro-benchmark for CuDNN function (Benchmarks: Support error tolerance in micro-benchmark for CuDNN function #490, Benchmarks - Fix bug to get metric from cmd when error happens in cudnn #506)
5. • Add distributed inference benchmark (Benchmarks - Add distributed inference benchmark #493 and Revise Code - Fix wrong torch usage in communication wrapper for Distributed Inference Benchmark #505)
6. • Support tensor core precisions (e.g., FP8) and batch/shape range in cublaslt gemm (Benchmarks - Support tensor core precisions in cublaslt gemm #492, Benchmark - Support batch/shape range in cublaslt gemm #494, and Benchmark - Fix matrix size overflow issue in cuBLASLt GEMM #503)

Model Benchmark Improvement

1. • Fix torch.dist init issue with multiple models (Benchmark - Fix torch.dist init issue with multiple models #495)
2. • Support TE FP8 in BERT/GPT2 models (Benchmarks - Support TE FP8 in BERT/GPT2 models #496, Benchmark - Update TE FP8 model conversion #499)
3. • Add num_workers configurable in model benchmark (Add num_workers argument in model benchmark #511)