Issues

• [PyTorch] Avoid selecting FA4 for deterministic training on SM120

  bug
  NVIDIA/TransformerEngine#3594 · Filter by author janbernloehr opened on Sep 30, 2026

• [Bug] Float8CurrentScaling / Float8BlockScaling: fp8_quant_* QParams ignore use_power_2_scales / use_f32_scales passed to the constructor

  bug
  NVIDIA/TransformerEngine#3593 · Filter by author Varun-51 opened on Sep 30, 2026
  2

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

  NVIDIA/TransformerEngine#3568 · Filter by author arcusbuilds opened on Sep 24, 2026

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

  NVIDIA/TransformerEngine#3564 · Filter by author ewan0x79 opened on Sep 23, 2026
  2

• Preserve NaN through half-precision MXFP8 amax reductions

  NVIDIA/TransformerEngine#3550 · Filter by author sylvesterkaczmarek opened on Sep 19, 2026

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

  NVIDIA/TransformerEngine#3528 · Filter by author nvegesna-netizen opened on Sep 16, 2026
  3

• Graph-Safe FP8 BlockScaling Support on Blackwell for GroupedLinear

  NVIDIA/TransformerEngine#3520 · Filter by author vthumbe1503 opened on Sep 15, 2026
  1

  

• MXFP8 GEMM rejects compact unswizzled scales that quantize_ accepts

  NVIDIA/TransformerEngine#3518 · Filter by author wujingyue opened on Sep 15, 2026
  2

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

  attention
  NVIDIA/TransformerEngine#3487 · Filter by author 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

  attention
  NVIDIA/TransformerEngine#3481 · Filter by author yuweih205 opened on Sep 4, 2026
  4

• Tied weights do not accumulate with delayed wgrad

  NVIDIA/TransformerEngine#3437 · Filter by author wujingyue opened on Aug 28, 2026
  5

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

  NVIDIA/TransformerEngine#3381 · Filter by author cuichenx opened on Aug 13, 2026
  3

  

• [Bug] Triton mask-map moe_permute ignores num_out_tokens

  NVIDIA/TransformerEngine#3346 · Filter by author truong-v opened on Aug 11, 2026
  2

• [Bug] Deferred init issue in te.pytorch.ops

  bug
  NVIDIA/TransformerEngine#3322 · Filter by author yaox12 opened on Aug 6, 2026
  2

  

• [Bug] CUDA graph backward replay illegal memory access after validation with mixed NVFP4/MXFP8 Mamba

  NVIDIA/TransformerEngine#3316 · Filter by author Phlip79 opened on Aug 4, 2026

• [Build] Publish PyTorch 2.11 CUDA 13 aarch64 wheels for GB200

  NVIDIA/TransformerEngine#3310 · Filter by author bvolpato opened on Aug 4, 2026

• [Build] Restore CUDA 13 Linux aarch64 core wheels

  NVIDIA/TransformerEngine#3309 · Filter by author bvolpato opened on Aug 4, 2026

• [Bug] Transformer Engine SM120 (Blackwell) Compatibility Issue

  bug
  NVIDIA/TransformerEngine#3299 · Filter by author lmcl90 opened on Aug 3, 2026
  2

• [Bug] Using QuantizedTensor on CPU causes a RecursionError

  bug
  NVIDIA/TransformerEngine#3294 · Filter by author janekb04 opened on Jul 30, 2026
  4

• [Bug] FusedAttention THD + learnable softmax backward IMA (cuDNN err 700) on Hopper

  attentionbug
  NVIDIA/TransformerEngine#3249 · Filter by author fy1214 opened on Jul 24, 2026
  6

  

• FP8 DPA future token leakage

  attentionbug
  NVIDIA/TransformerEngine#3248 · Filter by author jeromeku opened on Jul 24, 2026
  1

• cuDNN to accept actual seqlens in addition to cumulative seqlens

  attention
  NVIDIA/TransformerEngine#3218 · Filter by author sudhakarsingh27 opened on Jul 16, 2026
  1

  

• [BUG] NVFP4 fails on RTX 5070 Laptop (SM120): CUDA Error: no kernel image is available for execution on the device

  NVIDIA/TransformerEngine#3217 · Filter by author Yasho4867 opened on Jul 16, 2026
  2

• [PyTorch] FusedAdam/multi_tensor_apply: OOB metadata writes when a zero-numel tensor lands in the last TensorListMetadata slot (CUDA IMA / misaligned address; silent corruption below threshold)

  NVIDIA/TransformerEngine#3202 · Filter by author yuhezhang-ai opened on Jul 10, 2026

• [PyTorch] Request a migration path for downstream _GroupedLinear integrations after the 2.17 signature change

  NVIDIA/TransformerEngine#3191 · Filter by author cuichenx opened on Jul 8, 2026
  2