All pull requests

Pull requests

• [PyTorch] Fix full-graph compilation of linear attention

  #66·pggPL opened Oct 14, 2026Owner·5 of 6 tasks·

  ·

• [PyTorch] Add torch.compile support for fused MoE routing

  #65·pggPL opened Oct 14, 2026Owner·8 tasks done·

  ·

• Add BulkTensor with native tensor unpacking

  #64·pggPL opened Oct 13, 2026Owner·

  ·

• [PyTorch] Skip torch.compile tests when required PyTorch APIs are unavailable

  #63·pggPL opened Oct 13, 2026Owner·4 of 13 tasks·

  ·

• Add torch.compile support for context-parallel fused attention

  #62·pggPL opened Oct 9, 2026Owner·5 of 6 tasks·

  ·

• Refactor context-parallel attention autograd state

  #61·pggPL opened Oct 9, 2026Owner·5 tasks done·

  ·

• Prefer Python NCCL communicator getters with a C++ fallback

  #59·pggPL opened Oct 8, 2026Owner·7 of 8 tasks·

  ·

• Add GroupedTensor alignment metadata and guard grouped MLP fusion

  #58·pggPL opened Oct 7, 2026Owner·6 of 7 tasks·

  ·
  1

• [PyTorch] Compile activation operations and fused bias gradients

  #57·pggPL opened Oct 6, 2026Owner·7 of 9 tasks·

  ·

• [PyTorch] Preserve bulk allocation during CPU offload

  #56·pggPL opened Oct 2, 2026Owner·8 of 13 tasks·

  ·

• [PyTorch] Avoid concatenation copies for compiled split Linear parameters

  #55·pggPL opened Oct 1, 2026Owner·3 of 4 tasks·

  ·

• [PyTorch] Add torch.compile support for MultiheadAttention

  #54·pggPL opened Oct 1, 2026Owner·5 of 6 tasks·

  ·

• [PyTorch] Benchmark Sequential CPU overhead on larger workloads

  #53·pggPL opened Sep 17, 2026Owner·

  ·

• [PyTorch] Compile Linear and Bias operations with forward fusion

  #52·pggPL opened Sep 17, 2026Owner·5 of 6 tasks·

  ·

• [PyTorch] Support FP8 weight caching in compiled Linear

  #51·pggPL opened Sep 10, 2026Owner·6 of 8 tasks·

  ·

• [PyTorch] Refactor GroupedLinear eager forward and backward

  #50·pggPL opened Sep 10, 2026Owner·5 tasks done·

  ·

• [CI] Make GitHub Build workflow fit the hosted runner (test PR)

  #49·pggPL opened Sep 8, 2026Owner·

  ·

• [PyTorch] Docs: fine-grained quantization recipes (rework of NVIDIA#3336)

  #48·pggPL opened Sep 7, 2026Owner·5 of 13 tasks·

  ·

• [PyTorch][torch.compile] Differentiable fp8_output and fp8_grad for Linear

  #47·pggPL opened Sep 4, 2026Owner·8 of 13 tasks·

  ·

• [PyTorch] torch.compile support for FusedAttention (diff vs #3480 only)

  #46·pggPL opened Sep 4, 2026Owner·

  ·

• [PyTorch] torch.compile for an OperationFuser group holding one operation

  #45·pggPL opened Sep 4, 2026Owner·7 of 13 tasks·

  ·

• [PyTorch] [torch.compile] Prepare the custom-op framework and LayerNorm modules for compile wiring

  #44·pggPL opened Sep 3, 2026Owner·6 of 8 tasks·

  ·

• [PyTorch] [torch.compile] torch.compile support for LayerNormLinear and LayerNormMLP

  #43·pggPL opened Sep 3, 2026Owner·

  ·

• [PyTorch] [torch.compile] Prepare LayerNormLinear and LayerNormMLP for torch.compile

  #42·pggPL opened Sep 3, 2026Owner·6 of 8 tasks·

  ·

• [PyTorch] Schedule delayed-scaling updates after backward

  #39·pggPL opened Aug 20, 2026Owner·9 of 13 tasks·

  ·