mmaemu documentation¶

Using mmaemu¶

Page	What it covers
User guide	install and check, a first GEMM, profiles, split-K, backends, using mmaemu inside a model, testing
API reference	every public function and class, the exceptions and warnings, the command line
Evo 2 on other GPUs	why Evo 2 needs FP8 numerics, how to route its FP8 projections through mmaemu, measured results
Evo 2 on Blackwell notebook	the runnable version of the above, with its outputs on an RTX 5070 Ti
FAQ	short answers to the questions people ask first

Page	What it covers
Semantics	the numerics contract: the instruction, the accumulation plan, the epilogue, what raises
Presets and profiles	tensor-core models and reference platforms, their maturity, split-K resolution, environment variables
Split-K tables	the recorded cuBLASLt heuristic tables (RTX 4060 and H100 PCIe), their coverage and how they are used

Page	What it covers
Backends	the CUDA kernels, how they are loaded (fatbin, NVRTC, reference), their gates, platforms
Gated fast kernels	v9 and tc3: how they use the tensor cores, their gates and their measured evidence

Page	What it covers
Validation	the hardware recordings behind every claim, the scope table and the release results, with references
Limitations	what v0.1 does not do, and corrections of statements made about the prototype
Performance	how speed is measured and the expected kernel throughput
Testing	the test tiers (T0-T7), markers and the bitwise policy
Validation logs	the committed logs of every validation run

Page	What it covers
Releasing	the release checklist: freeze, distributions, publishing
sm_120 bring-up runbook	how to collect hardware evidence for a new GPU's FP8 arithmetic