[DRAFT] NVFP4 / MXFP4 everything#1477

Open

jomitchellnv wants to merge 21 commits intomainfrom

jm/nvfp4-homebase-02202026

Commits on Jan 23, 2026

tries to get fp4 working
Jonathan Mitchell
committed
refactors fp8 stats logs
Jonathan Mitchell
committed
fp4 debugging stats yaml
Jonathan Mitchell
committed

Commits on Jan 24, 2026

BF16 last 6 layers
Jonathan Mitchell
committed

Commits on Jan 26, 2026

sets bf16 layers thru cli
Jonathan Mitchell
committed
donwgrade te version
Jonathan Mitchell
committed

Commits on Feb 2, 2026

layer specific autocast
Jonathan Mitchell
committed
enables layer specific fp recipes
Jonathan Mitchell
committed
adds fp32 optim weights with bf16 compute weights
Jonathan Mitchell
committed

Commits on Feb 3, 2026

enables grad reduce in fp32 for better precision
Jonathan Mitchell
committed
adds FusedAdam for fun
Jonathan Mitchell
committed
fixes up debugging yaml and adds dockerfile for te tot
Jonathan Mitchell
committed

Commits on Feb 4, 2026

inject layer regex patterns for fp4 fp8
Jonathan Mitchell
committed
enables fp4 layer with nothing
Jonathan Mitchell
committed
pins autotokenizer to previous revision
Jonathan Mitchell
committed

Commits on Feb 5, 2026

adds Dockerfile.te_testing for TE build from src but bad perf on it
Jonathan Mitchell
committed

Commits on Feb 9, 2026

enables tokenizer revision
Jonathan Mitchell
committed

Commits on Feb 10, 2026

adds nsight profiling
Jonathan Mitchell
committed
adds gemm profiling scripts
Jonathan Mitchell
committed

Commits on Feb 20, 2026

NVFP4 with Nsight code
Jonathan Mitchell
committed
adds stuff to docker ignore
Jonathan Mitchell
committed