Repository Issue Activity (beta)

#2541 no attention backend support for softmax = learnable and cp_comm_type = p2p

#2986 Expose Batch Invariant Kernels

#1761 Create option to control data type for tensor parallel all-reduce

#2995 Reduce CPU overheads in te Sequential GroupedLinear Op

#2405 CUDA extension loading should be modular

#3437 Tied weights do not accumulate with delayed wgrad

#3550 Preserve NaN through half-precision MXFP8 amax reductions

#789 Request for Adaptive Layer Norm MLP

#3528 FA4 returns an all-zero output for causal attention on SM100 after flash-attention #2490

#3518 MXFP8 GEMM rejects compact unswizzled scales that quantize_ accepts

#3520 Graph-Safe FP8 BlockScaling Support on Blackwell for GroupedLinear

#3487 [PyTorch][Attention] THD P2P context-parallel regression when padded cu_seqlens are value-equal but not object-identical

#2807 Adaptive Compression

#3312 thd + attention_dropout routes to the composite cuDNN SDPA engine and costs 5x on Blackwell

#1688 fp8_model_init does nothing when used with FSDP2

#1436 If no Windows support is planned for 5090 as it wasn't for 4090, pass along to corporate to mention this in advertisement

#1532 How can we use te.Linear with weight parallel?

#2703 Performance bottleneck in TransformerLayer and Attention when using attention mask

#3484 [Bug] NaN expert weight gradients at num_groups == 1 with a padded token buffer for SReLU

#3028 [PyTorch] Integrate cuDNN GQA + DSA backend into DotProductAttention

#3481 [Bug] Backend selection picks FA3 for training with head_dim_qk=192 / v_head_dim=128, but FA3 backward cannot run it

#3249 [Bug] FusedAttention THD + learnable softmax backward IMA (cuDNN err 700) on Hopper

#3248 FP8 DPA future token leakage

#3218 cuDNN to accept actual seqlens in addition to cumulative seqlens

#3188 Plumb dbias request in isolation into the common backend selector