All issues

[BUG] Fused MoE auxiliary loss accepts shapes that overflow int offsets

#3568

· arcusbuilds opened

on Sep 24, 2026

[Feature Request] Support dynamic token counts in TP communication/GEMM overlap

#3564

· ewan0x79 opened

on Sep 23, 2026

[Bug] Non-FP8 `general_gemm` silently misreads a bias whose dtype differs from the output dtype and reads out of bounds

#3562

· MarkXCloud opened

on Sep 23, 2026

Preserve NaN through half-precision MXFP8 amax reductions

#3550

· sylvesterkaczmarek opened

on Sep 19, 2026

FA4 returns an all-zero output for causal attention on SM100 after flash-attention #2490

#3528

· nvegesna-netizen opened

on Sep 16, 2026

Graph-Safe FP8 BlockScaling Support on Blackwell for GroupedLinear

#3520

· vthumbe1503 opened

on Sep 15, 2026

MXFP8 GEMM rejects compact unswizzled scales that quantize_ accepts

#3518

· wujingyue opened

on Sep 15, 2026

[PyTorch][Attention] THD P2P context-parallel regression when padded cu_seqlens are value-equal but not object-identical

#3487

· cuichenx opened

on Sep 5, 2026

[Bug] Backend selection picks FA3 for training with head_dim_qk=192 / v_head_dim=128, but FA3 backward cannot run it

#3481

· yuweih205 opened

on Sep 4, 2026

Tied weights do not accumulate with delayed wgrad

#3437

· wujingyue opened

on Aug 28, 2026

NCCL EP hard link makes `import transformer_engine` require `libcuda.so.1`

#3381

· cuichenx opened

on Aug 13, 2026

[Bug] Triton mask-map moe_permute ignores num_out_tokens

#3346

· truong-v opened

on Aug 11, 2026