Add multi-precision training support to FSDP script#2662

Open

aagallo wants to merge 25 commits intoNVIDIA:mainfrom

aagallo:extend-precision

Commits on Feb 9, 2026

Add precision parameter support for multiple training formats

aagallo
authored and
aagallo
committed
Merge branch 'NVIDIA:main' into extend-precision
aagallo
authored
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed

Commits on Feb 10, 2026

Merge branch 'NVIDIA:main' into main
aagallo
authored
Fix FP16 dtype mapping and implement CLI flag precedence
aagallo
committed
Add logging and documentation for precision configuration
aagallo
committed
Initialize recipe variable in all precision cases
aagallo
committed
Fix dtype flag detection to support explicit override behavior
aagallo
committed
Merge remote-tracking branch 'origin/main' into extend-precision
aagallo
committed
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed
Replace sys.argv parsing with custom action and fix default case
aagallo
committed
Merge branch 'extend-precision' of https://github.com/aagallo/TransformerEngine into extend-precision
aagallo
committed
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed
Fix params_dtype to use computed dtype from precision logic
aagallo
committed
Merge branch 'extend-precision' of https://github.com/aagallo/TransformerEngine into extend-precision
aagallo
committed
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed
Merge branch 'main' into extend-precision
aagallo
authored
Fix type conversion in StoreExplicitAction for --dtype argument
aagallo
committed
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed