All pull requests

Pull requests

• PT graph capture: captured forward steps (torch_cuda_graph forward)

  #2026·albertz opened Oct 17, 2026Member·

  ·

• RF Cache, eager dyn size value key via tobytes

  #2023·albertz opened Oct 16, 2026Member·

  ·

• Dim.get_mask, use rf.Cache

  #2021·albertz opened Oct 16, 2026Member·

  ·

• RF Cache, TensorLayout as key, output mapped in its storage layout, Dim key backend

  #2020·albertz opened Oct 16, 2026Member·

  ·

• PT while_loop: torch while_loop op under static tracing

  #2018·albertz opened Oct 16, 2026Member·

  ·

• RF mask_nested: dim sizes on the mask device under static tracing

  #2017·albertz opened Oct 16, 2026Member·

  ·

• RF RotaryPosCausalSelfAttention: bounded cache holds keys after RoPE

  #2016·albertz opened Oct 16, 2026Member·

  ·

• Packed backend lstm

  #2011·polycrit opened Oct 15, 2026Contributor·

  ·

• Dataset, reuse full seq order across partitions of an epoch

  #2008·albertz opened Oct 15, 2026Member·

  ·

• PT graph capture: compiled warm runs and capture on side stream

  #2007·albertz opened Oct 15, 2026Member·

  ·

• PT graph capture: real batch size as compiled step input

  #2006·albertz opened Oct 14, 2026Member·

  ·

• RF replace_dim_v2, fix expand under static tracing

  #2003·albertz opened Oct 14, 2026Member·

  ·

• main, finalize also when the task died, keep the original exception

  #2002·albertz opened Oct 14, 2026Member·

  ·

• PT graph capture, compiled and captured eval steps (torch_cuda_graph eval)

  #1995·albertz opened Oct 13, 2026Member·

  ·

• stop_for_resubmission_when_low_time_left, coordinated stop, SLURM-level interruption

  #1992·albertz opened Oct 12, 2026Member·

  ·

• RF packed, one helper for an operand in the packing of another

  #1989·albertz opened Oct 11, 2026Member·

  ·

• RF monotonic RNN-T loss, blank scale and label smoothing

  #1988·polycrit opened Oct 11, 2026Contributor·

  ·

• PT load_optimizer: keep the state of params moved between param groups

  #1984·polycrit opened Oct 10, 2026Contributor·

  ·

• PT slice_update in place

  #1982·polycrit opened Oct 10, 2026Contributor·

  ·
  1

• RF monotonic RNN-T loss, lattice log probs

  #1981·polycrit opened Oct 10, 2026Contributor·

  ·

• RF packed shift keeps the packing

  #1979·polycrit opened Oct 10, 2026Contributor·

  ·

• CachedDataset2, epoch without seqs with unknown num seqs

  #1973·polycrit opened Oct 8, 2026Contributor·

  ·

• MultiProcDataset, dedicated sharding yields every seq

  #1972·polycrit opened Oct 8, 2026Contributor·

  ·

• PT graph capture bounds from the batching options

  #1971·polycrit opened Oct 8, 2026Contributor·

  ·

• RF packed ctc_loss, shared target length

  #1970·polycrit opened Oct 8, 2026Contributor·

  ·