Repository navigation

IFU 2.20 - #756

Merged

wenchenvincent merged 107 commits into

devfrom

zain/ifu/2.20

Oct 15, 2026

Merged

IFU 2.20#756
wenchenvincent merged 107 commits into
devfrom
zain/ifu/2.20

Commits on Aug 17, 2026
- [PyTorch][torch.compile] Support for DotProductAttention on flash and unfused backends (#3286)
  
  authored
  252a474
- Change VERSION to 2.20.0.dev0 (#3386)
  fheinecke
  authored
  6f83d74
Commits on Aug 18, 2026
- [JAX] Support synchronous collectives in HLO collective bytes assert tests (#3398)
  jberchtold-nvidia
  authored
  27b12df
Commits on Aug 19, 2026
- [JAX] SBHD reorder skip uses original shape instead of swapped tensor (#3373)
  
  andrewwhitecdw
  and
  KshitijLakhani
  authored
  a8ff824
- Refactor GroupedLinear quantization dispatch (#3326)
  
  negvet
  and
  pre-commit-ci[bot]
  authored
  576b215
- [PyTorch] Fix MXFP8 master weight cast when a rank owns no shard. (#3348)
  
  rapatel
  and
  negvet
  authored
  57896b0
- fix: comment typo 'mas' -> 'mask' in two TODOs (#3372)
  andrewwhitecdw
  authored
  d4ca825
Commits on Aug 20, 2026
Commits on Aug 21, 2026
Commits on Aug 22, 2026
- Fix include after ptx move (#3416)
  mmarcinkiewicz
  authored
  849ad3f
- [PyTorch][Test-Fix] Scale test_multi_process_ep outer timeout with launcher pass count (#3415)
  
  phu0ngng
  and
  fheinecke
  authored
  d0067d0
Commits on Aug 23, 2026
- Add the support of variable dtypes of params and grads in capturable Fused Adam. (#3414)
  
  ptrendx
  and
  pre-commit-ci[bot]
  authored
  ea1a165
Commits on Aug 25, 2026
Commits on Aug 26, 2026
Commits on Aug 27, 2026
Commits on Aug 28, 2026
- [JAX] Limit max logit attr to forward (#3428)
  
  KshitijLakhani
  and
  jberchtold-nvidia
  authored
  a60f55a
- [JAX] Fix GEMM partitioning to reduce over nested contracting mesh axes (#3429)
  phu0ngng
  authored
  609584a
- [Common/PyTorch] Grouped weighted-SwiGLU MXFP8 kernel (#3315)
  
  authored
  174caef
- Add a backwards linear function to be used with the fused mla q up-proj (#3330)
  
  authored
  7695bbd
- [Common] Improved performance of Group MXFP8 kernels (#3337)
  
  Oleg-Goncharov
  and
  pre-commit-ci[bot]
  authored
  8930052
- Fix linter error (#3435)
  fheinecke
  authored
  ace1873
Commits on Aug 31, 2026
- [Pytorch][Attention][CP][CI] Make 3221 changes robust (#3438)
  
  sudhakarsingh27
  and
  pre-commit-ci[bot]
  authored
  aa58996
- [Common] Gate FP8 THD sink backward on cuDNN 9.26 (#3441)
  sudhakarsingh27
  authored
  0ae8996
- Minimize the memory usage of the fused cross entropy kernel (#3273)
  
  authored
  4fba94a
- [PyTorch] Fix fake kernel metadata mismatch for empty-input MoE permute ops (#3444)
  pggPL
  authored
  fbeb316
- fix: guard backward_dw() bias-grad backfill against a legitimately-empty bgrad (#3400)
  
  nvegesna-netizen
  and
  ksivaman
  authored
  af999f9
- Publishing updates (#3399)
  
  authored
  e5e46da
- Fix GH action generated docs (#3449)
  ksivaman
  authored
  82853f1
- Use `dlopen` to load NCCL EP (#3434)
  fheinecke
  authored
  1ff9c37
Commits on Sep 1, 2026
- [PyTorch] Fix EP zero-copy MXFP8 scale offset to be window-relative (#3447)
  phu0ngng
  authored
  e0f6a15
- [CI] Improve build time dependency resolution (#3251)
  
  fheinecke
  and
  pre-commit-ci[bot]
  authored
  6398bbb
- Bump NCCL-EP submodule commit (#3451)
  phu0ngng
  authored
  7080239
- [Common] Reduce default cfg's SMEM usage for mhc triton kernel (#3442)
  
  kainzhong
  and
  ptrendx
  authored
  f307af8
- [Core] Fix L40 cublas heuristics (#3440)
  
  jberchtold-nvidia
  and
  pre-commit-ci[bot]
  authored
  8adc494
- Misc docs rendering fixes (#3450)
  
  authored
  abf0ebb
- [Common] Split grouped activation build (#3430)
  
  harryzhou2000
  and
  ptrendx
  authored
  30f6008
- [PyTorch] Deprecate unused is_cg_capturable in parallel_cross_entropy (#3455)
  pggPL
  authored
  918a61e
- fix: simplify redundant softmax version condition (#3363)
  andrewwhitecdw
  authored
  1ee3047
- [Common] Use wide instructions in SR to reduce issue-bound bottleneck (#3357)
  
  janekb04
  and
  Oleg-Goncharov
  authored
  8589b15
Commits on Sep 2, 2026
Commits on Sep 3, 2026
- [PyT] Fix PyTorch BF16 reduction numerics tests (#3461)
  jberchtold-nvidia
  authored
  c829717
- Allow newer NLTK versions (#3464)
  fheinecke
  authored
  9c2fcd6
- [PyTorch] Fused GDN attention (#3351)
  
  ksivaman
  and
  pre-commit-ci[bot]
  authored
  5c8c6c0
- [PyTorch] Add ScaledTanhSReLU and wire it into the fused grouped MLP (#3463)
  wanyingw
  authored
  20f4cec
- [PyT] Add per-sequence attention policies to packed THD attention (#3274)
  
  authored
  2a435a2
- [PyTorch] Pair delayed-scaling FP8 recompute metadata per module (#3394)
  
  nvegesna-netizen
  and
  pggPL
  authored
  d16c754
- Compress CUDA binaries (#3465)
  fheinecke
  authored
  888da30
- fix: make_mask uses deprecated jnp.bool instead of jnp.bool_ (#3371)
  andrewwhitecdw
  authored
  362d77a
- [JAX] Add sqrtsoftplus router score function (#3448)
  
  jberchtold-nvidia
  and
  pre-commit-ci[bot]
  authored
  53d8981
Commits on Sep 4, 2026
- WAR for the cuBLAS grouped gemm issue (#3475)
  ptrendx
  authored
  ee253aa
- [PyT] Fix lint failures (#3476)
  KshitijLakhani
  authored
  45808f1
- [PyTorch] Decline fused grouped MLP when the backward format is not E4M3 (#3352)
  
  wilyan09007
  and
  ptrendx
  authored
  b86d8f9
- [PyTorch] Fix: Resolve EP symm-mem window offset for both old and new torch version (#3466)
  phu0ngng
  authored
  e61b66d
- [NCCL][PyTorch] Fix test-fusible-ops-file-rendezvous-bug (#3478)
  
  authored
  1f25d6a
- Reduce Grouped MLP Fuser CPU Overhead (#3410)
  
  zhongbozhu
  and
  tingyangk
  authored
  80a89ad
- Use cuDNN's deterministic dprob in the fused grouped MLP (#3407)
  
  ZhiyuLi-Nvidia
  and
  vthumbe1503
  authored
  d1e9c33
- [PyTorch] Reduce CUDA graph memory retention (#3427)
  buptzyb
  authored
  29229b5
- Relax runtime checks for activation recompute into Warnings (#3436)
  
  ghadiaravi13
  and
  timmoon10
  authored
  846379d
Commits on Sep 5, 2026
- fix: Unreachable backend check after earlier backend skip (#3368)
  
  andrewwhitecdw
  and
  ptrendx
  authored
  a30aee5
- [Common] Preserve shared-memory pointer provenance in TMA kernels (#3482)
  
  Oleg-Goncharov
  and
  pre-commit-ci[bot]
  authored
  5f6105b
Commits on Sep 8, 2026
- Update Linear docstring referring torch.Linear (#3491)
  negvet
  authored
  dc8909d
- [PyTorch] Document fine-grained quantization recipes (#3336)
  
  authored
  f2bec23
- Fix hybrid fp8 test guard (#3492)
  negvet
  authored
  cd24504
- [PyTorch] Use torch's register_custom_class API for opaque quantizers when available (#3495)
  
  mmarcinkiewicz
  and
  claude
  authored
  b97f5ab
Commits on Sep 9, 2026
Commits on Sep 10, 2026
Commits on Sep 11, 2026
- Support paged stashing for GroupedLinear activations (#3423)
  lhb8125
  authored
  02f8e75
- [Pytorch] Replace TorchScript fallbacks with eager execution (#3502)
  sudhakarsingh27
  authored
  224f6ec
- Insert assert header (#3511)
  
  aybchan
  and
  pre-commit-ci[bot]
  authored
  3efd23a
- Cutlass Upgrade to 4.8 (#3443)
  
  vthumbe1503
  and
  pre-commit-ci[bot]
  authored
  9c735bf
Commits on Sep 14, 2026
- [PyTorch] Split FusedAttnFunc into single-argument forward/backward helpers (#3480)
  pggPL
  authored
  e26066b
- [CI] Remove failing GitHub Build workflow (#3514)
  pggPL
  authored
  51317b4
Commits on Sep 15, 2026
- [JAX] Add EP borrowed-comm path via XLA FFI collectives extension (#3452)
  phu0ngng
  authored
  73b5970
- Add MXFP8 support with cuBLASMp (#3145)
  almogsegal
  authored
  110791f
- [PyTorch] Fix mutable QB bounds in CUDA graphs (#3426)
  
  harryzhou2000
  and
  ksivaman
  authored
  796346c
Commits on Oct 2, 2026
- Merge commit '796346c0e0497b1f56a8d36ec76e02db5a6fed47' into zain/ifu/2.20
  Micky774
  committed
  3b9c0a0
Commits on Oct 15, 2026
- [ROCm] IFU v2.20: resolve merge conflicts and adapt for ROCm
  
  Micky774
  and
  claude
  committed
  e942d0d

Commits

Commits on Aug 17, 2026

Commits on Aug 18, 2026

Commits on Aug 19, 2026

Commits on Aug 20, 2026

Commits on Aug 21, 2026

Commits on Aug 22, 2026

Commits on Aug 23, 2026

Commits on Aug 25, 2026

Commits on Aug 26, 2026

Commits on Aug 27, 2026

Commits on Aug 28, 2026

Commits on Aug 31, 2026

Commits on Sep 1, 2026

Commits on Sep 2, 2026

Commits on Sep 3, 2026

Commits on Sep 4, 2026

Commits on Sep 5, 2026

Commits on Sep 8, 2026

Commits on Sep 9, 2026

Commits on Sep 10, 2026

Commits on Sep 11, 2026

Commits on Sep 14, 2026

Commits on Sep 15, 2026

Commits on Oct 2, 2026

Commits on Oct 15, 2026