GitHub · Where software is built

Labels Milestones New issue

SimpleFSDP Status Tracking

#1980

· ruisizhang123 opened

on Nov 3, 2025

question of PP x aux_loss for MoE

#1979

· rakkit opened

on Nov 3, 2025

Why is the ep mesh derived from a factoring of the dp mesh, instead of its own dimension?

#1977

· man2machine opened

on Nov 1, 2025

AC + compile interaction bug with MXFP8 MoE Training

#1971

· danielvegamyhre opened

on Oct 31, 2025

Avoiding device-to-host sync for input/output split sizes in expert parallel

#1968

· man2machine opened

on Oct 30, 2025

NVFP4 MoE Training Status

#1962

· syed-ahmed opened

on Oct 29, 2025

B200 Llama4 Compile dp2ep=4, pp=2 (interleave 1f1b), full AC fails when compiling token dispatch

module: torch.compile

#1940

· xmfan opened

on Oct 27, 2025

Is it possible to train Vision-Language Model with different parallelism plan for vision and language parts of the model?

#1936

· airlsyn opened

on Oct 27, 2025

deepseek compile fail to run

#1935

· BoyuanFeng opened

on Oct 25, 2025

grouped_mm illegal memory access when the input tensor size > 1640968*4096 for llama4

module: torch.compile

#1932

· yoonniverse opened

on Oct 24, 2025

why not using set_requires_all_reduce for hsdp

#1929

· samsja opened

on Oct 22, 2025

Benchmark SymmMem's all_to_all_vdev_2d on NVL72

#1914

· syed-ahmed opened

on Oct 17, 2025