V0.13.0 Release Plan

Release Manager

@cp5555

Endgame

• Code freeze: TBD
• Bug Bash date: TBD
• Release date: TBD

Main Features

SuperBench Improvement

1. • Add cuda13.0.dockerfile support (#739)
2. • Add nsys and pytorch profiler debug trace support (#744)

Micro-benchmark Improvement

1. • Collect per-snapshot per-GPU flops/temp in gpu burn (#735)
2. • Add simultanneously all-to-host / host-to-all bandwidth testcases to nvbandwidth (#736)
3. • Add ncu profile support in cublaslt-gemm (#740)
4. • Support verification and parallel run for disk performance benchmark (#741)
5. • Add numa support for nvbandwidth (#742)
6. • Change cublasLtMatmulDescCreate scaleType from CUDA_R_32F to CUDA_R_16F in FP16 dist inference (#732)
7. • Support gemm correctness check in cublaslt-gemm
8. • Multi node nccl validation enhancement
9. • mscclpp support
10. • Add new busbw metrics for NCCL/MSCCL testing with specific algorithm
11. • Fix NVBandwidth benchmark results parsing bug
12. • Support FP4 kernels for cutlass benchmark

Model Benchmark Improvement

1. • Add option to exclude data copy time in model benchmarks (#734)
2. • Support state-of-art LLM model training perf including Deepseek, qwen
3. • Support state-of-art LLM model inference perf including Deepseek, qwen
4. • Support state-of-art LLM module and model correctness benchmark
5. • Deterministic training support (#731)

Bug fix

1. • dist-inference raise cublaslt error
2. • Add --set_ib_devices option to auto-select IB device by MPI local rank in ib validation (#733)
3. • NVBandwidth benchmark results parsing bug (#748)
4. • CI/CD - Fix image merge in GitHub Action (#749)
5. • Fix pipelines - Update mlc version in dockerfiles from v3.11 to v3.12 (#752)
6. • CI/CD - Fix python3.10 pipeline (#753)
7. • CI/CD - Fix Azure test pipeline (#754)

Tools

1. • System info enhancement