NVIDIA/TransformerEngine issues and pull requests

#7119 - Changed the way pyright translates `tuple[()]` into a specialized `Se…

Pull Request - State: closed - Opened by erictraut almost 2 years ago - 1 comment

#3558 - [Common] Enable EP with FP32/FP16 payload dtypes

Pull Request - State: open - Opened by phu0ngng 1 day ago - 4 comments

#3557 - Treat total_num_pages as a physical KV cache budget, not the page-table size

Pull Request - State: open - Opened by 0z5a 2 days ago - 1 comment
Labels: community-contribution, attention

#3556 - Add an opt-in batch-invariant BF16 GEMM

Pull Request - State: open - Opened by 0z5a 2 days ago - 1 comment
Labels: community-contribution

#3555 - Add an opt-in float32 reduction dtype for the row-parallel all-reduce

Pull Request - State: open - Opened by 0z5a 2 days ago - 1 comment
Labels: community-contribution

#3554 - Reuse the host split sizes GroupedLinear's forward already computed

Pull Request - State: open - Opened by 0z5a 2 days ago - 1 comment
Labels: community-contribution

#3553 - Isolate FlashAttention 2/3 imports so a broken install cannot break TE

Pull Request - State: open - Opened by 0z5a 2 days ago - 1 comment
Labels: community-contribution, attention

#3552 - Accumulate delayed wgrads into shared (tied) parameters

Pull Request - State: open - Opened by 0z5a 2 days ago - 1 comment
Labels: community-contribution

#3551 - Support learnable softmax (attention sink) with cp_comm_type="p2p"

Pull Request - State: open - Opened by 0z5a 2 days ago - 1 comment
Labels: community-contribution, attention

#3550 - Preserve NaN through half-precision MXFP8 amax reductions

Issue - State: open - Opened by sylvesterkaczmarek 4 days ago

#3549 - Fix FusedSGD zero_grad argument forwarding

Pull Request - State: open - Opened by wujingyue 4 days ago - 1 comment
Labels: community-contribution

#3548 - KDA attention

Pull Request - State: open - Opened by ksivaman 4 days ago - 2 comments
Labels: attention, 2.21

#3547 - [PyTorch] Compile Linear and Bias operations with forward fusion

Pull Request - State: open - Opened by pggPL 4 days ago

#3546 - [PyTorch] Enable LayerNormLinear UB AG overlap under no_grad

Pull Request - State: open - Opened by ravimajeti 5 days ago - 1 comment
Labels: community-contribution

#3545 - [PyTorch] Keep DistributedWeight objects on ctx across saved-tensor hooks

Pull Request - State: open - Opened by xrennvidia 5 days ago - 1 comment
Labels: community-contribution

#3544 - feat: vmm slot for local cuda graph offload

Pull Request - State: open - Opened by Xdydy 5 days ago - 1 comment
Labels: community-contribution

#3543 - vmm support for cuda graph offload

Pull Request - State: closed - Opened by Xdydy 5 days ago - 1 comment
Labels: community-contribution

#3542 - Use PyTorch CUTLASS for BF16 grouped GEMM on SM100

Pull Request - State: open - Opened by wujingyue 5 days ago - 2 comments
Labels: community-contribution

#3541 - Fix distributed Newton-Schulz correctness

Pull Request - State: closed - Opened by vcherepanov-nv 5 days ago - 1 comment
Labels: 2.20

#3540 - [PyTorch][CP] Support compact metadata with FA4

Pull Request - State: open - Opened by sudhakarsingh27 5 days ago - 3 comments

#3539 - GDP attention support

Pull Request - State: closed - Opened by ksivaman 5 days ago - 3 comments
Labels: 2.20

#3538 - fix: add missing NVTE_BHSD labels to fused_attn to_string

Pull Request - State: closed - Opened by andrewwhitecdw 5 days ago - 1 comment
Labels: community-contribution, attention

#3537 - Prototype of green context + VMM localization

Pull Request - State: open - Opened by WanZzzzzz 5 days ago
Labels: community-contribution

#3536 - test(attention): anchor the context-parallel suite to an independent reference

Pull Request - State: open - Opened by nvegesna-netizen 5 days ago - 2 comments
Labels: community-contribution, attention

#3535 - [Docs] Refresh README

Pull Request - State: open - Opened by sbhavani 5 days ago - 1 comment

#3534 - [PyTorch] [torch.compile] torch.compile support for LayerNormLinear and LayerNormMLP

Pull Request - State: open - Opened by pggPL 6 days ago - 1 comment

#3533 - [PyTorch] Retain NVFP4 RNG tensors through quantization

Pull Request - State: open - Opened by Connor-XY 6 days ago - 1 comment
Labels: community-contribution

#3532 - fix(attention): stop handing FlashAttention 4 the -1 window sentinel

Pull Request - State: open - Opened by nvegesna-netizen 6 days ago - 1 comment
Labels: community-contribution, attention

#3531 - [PyTorch] Remove stale xfail for selective attention recompute

Pull Request - State: closed - Opened by negvet 6 days ago - 2 comments
Labels: 2.20

#3530 - [PyTorch] Extend no-load-balance CP to the a2a comm type

Pull Request - State: open - Opened by Rudin6 6 days ago - 1 comment
Labels: community-contribution, attention

#3529 - [PyTorch] Allow F16 GDN inside FP8 autocast

Pull Request - State: closed - Opened by layalir 6 days ago - 2 comments
Labels: community-contribution

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

Issue - State: open - Opened by nvegesna-netizen 7 days ago - 3 comments

#3527 - feat(attention): cuDNN FROST attention backend for head_dim in (256, 512], with context parallelism

Pull Request - State: open - Opened by nvegesna-netizen 7 days ago - 3 comments
Labels: community-contribution, attention, 2.21

#3525 - [JAX] WAR for XLA associative scan rewriter crash

Pull Request - State: open - Opened by KshitijLakhani 7 days ago - 4 comments

#3524 - [JAX] Remove duplicate score-mod test run

Pull Request - State: closed - Opened by KshitijLakhani 7 days ago - 4 comments
Labels: 2.20

#3523 - Remove the view() calls from Linear/LNLinear/LNMLP modules

Pull Request - State: open - Opened by ptrendx 7 days ago - 5 comments

#3522 - [JAX] Support single-process-multi-device EP via XLA-borrowed communicator

Pull Request - State: closed - Opened by phu0ngng 7 days ago - 4 comments
Labels: 2.20

#3521 - [PyTorch] GDN2 support and linear attention refactor

Pull Request - State: closed - Opened by ksivaman 7 days ago - 4 comments
Labels: 2.20

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

Issue - State: open - Opened by vthumbe1503 7 days ago - 1 comment

#3517 - [PyTorch] Support distributed weights in GroupedLinear's grouped-tensor path

Pull Request - State: open - Opened by fanshiqing 8 days ago - 2 comments
Labels: org-contribution

#3516 - [JAX] Align fused attn backward output gradient sharding and other test strengthening

Pull Request - State: open - Opened by KshitijLakhani 8 days ago - 3 comments
Labels: bug, attention, 2.20

#3513 - Abstract CUDA hardcodes into configurable te_device_type / te_platform

Pull Request - State: open - Opened by lxd-cumt 9 days ago - 1 comment
Labels: community-contribution

#3512 - Abstract CUDA hardcodes into configurable te_device_type / te_platform

Pull Request - State: closed - Opened by lxd-cumt 9 days ago - 1 comment
Labels: community-contribution

#3511 - Insert assert header

Pull Request - State: closed - Opened by aybchan 11 days ago - 2 comments
Labels: org-contribution

#3510 - [PyTorch] Add head-parallel FA4 backward for cuDNN CP attention

Pull Request - State: open - Opened by bzantium 11 days ago - 2 comments
Labels: community-contribution, attention

#3508 - [Draft] Port cuDNN frontend attention to Python API

Pull Request - State: closed - Opened by vcherepanov-nv 12 days ago - 2 comments

#3506 - Fix THD P2P pad detection and tail-zero gating

Pull Request - State: open - Opened by RPalmr 12 days ago - 3 comments
Labels: community-contribution, attention

#3503 - MOE Sequential Block with Dispatch and Combine as Basic Ops

Pull Request - State: open - Opened by vthumbe1503 13 days ago - 7 comments
Labels: 2.20

#3502 - [Pytorch] Replace TorchScript fallbacks with eager execution

Pull Request - State: closed - Opened by sudhakarsingh27 13 days ago - 4 comments
Labels: 3.0

#3499 - Fix hybrid extra state size mismatch

Pull Request - State: open - Opened by negvet 13 days ago - 2 comments

#3498 - [JAX] Fix undefined sr_rng_state that breaks --dry-run in two encoder examples

Pull Request - State: open - Opened by Anai-Guo 14 days ago - 1 comment
Labels: community-contribution

#3497 - Replace RuntimeError with warnings in ScaledSReLU

Pull Request - State: closed - Opened by ghadiaravi13 14 days ago - 1 comment
Labels: community-contribution

#3496 - [PyTorch] torch.compile for an OperationFuser group holding one operation

Pull Request - State: closed - Opened by pggPL 14 days ago - 3 comments

#3495 - [PyTorch] Use torch's register_custom_class API for opaque quantizers when available

Pull Request - State: closed - Opened by mmarcinkiewicz 15 days ago - 2 comments
Labels: community-contribution

#3494 - [Docs] Add Mixture of Experts guide

Pull Request - State: open - Opened by pggPL 15 days ago - 1 comment
Labels: documentation

#3493 - [PyTorch] [torch.compile] Prepare LayerNormLinear and LayerNormMLP for torch.compile

Pull Request - State: closed - Opened by pggPL 15 days ago - 5 comments

#3492 - Fix hybrid fp8 test guard

Pull Request - State: closed - Opened by negvet 15 days ago - 2 comments

#3491 - Update Linear docstring referring torch.Linear

Pull Request - State: closed - Opened by negvet 15 days ago - 1 comment
Labels: documentation

#3490 - [PyT] Disable FA3 for training when head_dim_qk != head_dim_v

Pull Request - State: open - Opened by yuweih205 16 days ago - 1 comment
Labels: community-contribution, attention

#3489 - Add native transport for dynamic context parallelism

Pull Request - State: closed - Opened by xiaoyao0115 16 days ago - 1 comment
Labels: community-contribution

#3488 - Support row-only MXFP8 distributed master-weight casts

Pull Request - State: open - Opened by xiuhu17 17 days ago - 1 comment
Labels: community-contribution

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

Issue - State: open - Opened by cuichenx 17 days ago
Labels: attention

#3486 - [JAX] Interpolate the tensor-sequence-parallelism warning

Pull Request - State: open - Opened by Anai-Guo 18 days ago - 3 comments
Labels: community-contribution

#3485 - [PyTorch] Fix NaN expert weight gradients at num_groups == 1 with SReLU

Pull Request - State: closed - Opened by GarlGuo 18 days ago - 3 comments
Labels: community-contribution

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

Issue - State: closed - Opened by GarlGuo 18 days ago - 1 comment
Labels: bug

#3483 - [Common] Fix Grouped MXFP8 work mapping and TMA synchronization

Pull Request - State: closed - Opened by Oleg-Goncharov 18 days ago - 3 comments

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

Issue - State: open - Opened by yuweih205 19 days ago - 4 comments
Labels: attention

#3480 - [PyTorch] Split FusedAttnFunc into single-argument forward/backward helpers

Pull Request - State: open - Opened by pggPL 19 days ago - 2 comments
Labels: 2.20

#3479 - Drop the unsupported window_size argument from attention backend queries

Pull Request - State: open - Opened by Anai-Guo 19 days ago - 1 comment
Labels: community-contribution, attention

#3477 - [PyT] Linear Attention API

Pull Request - State: open - Opened by KshitijLakhani 19 days ago

#3474 - [Bug] Illegal memory access in grouped MXFP8 quantize on GB200 (sm_100) introduced by #3337; TMA store completion handoff missing in group_quantize_mxfp8.cuh

Issue - State: open - Opened by janbernloehr 19 days ago - 6 comments

GitHub / NVIDIA/TransformerEngine issues and pull requests