Skip to content

release_v1.2.md

Latest commit

 

History

History
64 lines (46 loc) · 2.38 KB

release_v1.2.md

File metadata and controls

64 lines (46 loc) · 2.38 KB
Raw

SimuMax v1.2 Release Notes

Release status: released

Release date: 2026-05-11

Summary

SimuMax v1.2 refreshes the public B200 benchmark surface, adds formal dense CP A2A result coverage for B200, improves first-user system-config generation, and keeps sync-VPP available as Preview.

Supported

  • B200 perf-vs-real benchmark summary and reproduction workflow:
  • Dense Llama3 B200 CP A2A rows for 32K and 128K sequence lengths.
  • Megatron-LM 0.14 selective recompute modeling with discard_output semantics through megatron_recompute.
  • Existing public A100-PCIe benchmark summary:
  • Simulator trace export through simulate(), including trace and optional memory artifacts as documented in tutorial.md.

Preview

  • sync-VPP is included as Preview while target-machine validation and support boundaries continue to tighten.

Representative B200 sync-VPP preview check:

case real ms perf ms simulator trace end status
llama3_70b_l8_tp1_pp4_vp2_dp2_mbc8_sync 680.70 661.21 663.29 Preview

This row is included only as a current behavior check for sync-VPP. It is not part of the formal B200 benchmark table.

Tooling

  • The public one-click compute-efficiency path supports fresh cache namespaces through --compute-cache-mode rebuild and --compute-cache-tag.
  • Strategy configs can opt into Megatron-LM 0.14 selective recompute with megatron_recompute=true and a non-empty megatron_recompute_modules list.
  • GEMM and grouped-GEMM efficiency scripts are aligned with the current TransformerEngine workspace interfaces.
  • The shared system-config generation path now verifies that generated configs can be loaded by SystemConfig.init_from_config_file.

Results

B200 CP A2A perf vs real

The retained B200 CP rows are dense Llama3 A2A cases. DeepSeek CP rows remain outside the formal public table.

Known Boundaries

  • sync-VPP is Preview, not a broad VPP support claim.
  • async-VPP perf timing is not part of the public support surface.
  • New or materially different machines should measure their own operator efficiency and communication data before treating timing estimates as benchmark-grade.