torchtitan

[rl] Add batched RL training with varlen sequence packing

[RL] Fix RL h100 workflow by installing FA3 and add explicit enable_gqa flag (#2891)

fix(data-ckpting-resume): fix broken shuffling (#2865)

[GraphTrainer][AutoDev] Enable cudagraph in aot_fx_trace mode (#2868)

[compile] Share SimpleFSDP wrapper class across same-type module instances (#2754)