All pull requests

New pull request

Grouped Linear CPU Overhead Reductions By Removing Unnecessary Python Logic
#3613zhongbozhu opened Oct 2, 2026Collaborator13 tasks
[JAX] Initialize softmax custom calls before execution
#3612KshitijLakhani opened Oct 1, 2026Collaborator5 of 13 tasks
4
Fix JAX distributed unit test failure
#3611fheinecke opened Oct 1, 2026Collaborator1 of 13 tasks
4
[JAX] Add Triton dependency to JAX extension
#3610jberchtold-nvidia opened Oct 1, 2026Collaborator8 of 13 tasks
5
[PyTorch][CP][CI] Reset all CP process pools before retrying a failed test
#3608sudhakarsingh27 opened Oct 1, 2026Member5 of 13 tasks
3
[PyTorch] Add torch.compile support for GroupedLinear
#3607pggPL opened Oct 1, 2026Collaborator
[PyTorch] Refactor GroupedLinear forward and backward
#3606pggPL opened Oct 1, 2026Collaborator6 of 13 tasks
3
Build FP8 recipe quantizer params per instance
#3605Varun-51 opened Oct 1, 20266 of 8 tasks
1
Fix FP8 recipe quantizer options
#3603efegokdemir opened Oct 1, 20265 tasks done
[Common] Fix CUDA 12.8 build: use an L2 policy in the 128-bit evict_first load fallback
#3602ravimajeti opened Oct 1, 20266 of 13 tasks
4
[Common] Fall back from gated TMA kernels when shared memory does not fit
#3600ravimajeti opened Oct 1, 20266 of 13 tasks
1
[Common] Fix grouped MXFP8 quantization with varying last dim & swizzled scale
#3599kainzhong opened Sep 30, 2026Collaborator8 of 13 tasks
2
TE MoE Return Expert Counts and Custom Routing for DSv4 Support
#3596dinodeep opened Sep 30, 2026Contributor1 of 13 tasks
11
Skip FA4 for deterministic backward on SM12x
#3595janbernloehr opened Sep 30, 2026Contributor6 of 13 tasks
9
Fix MXFP8 ONNX block quantization attributes
#3592pggPL opened Sep 30, 2026Collaborator4 tasks done
3
[Common, PyTorch] Add grouped NVFP4 dequantization
#3591harshithkantamneni opened Sep 30, 20268 of 13 tasks
2
[JAX] Disable async all-reduce for distributed softmax CI
#3590KshitijLakhani opened Sep 29, 2026Collaborator13 tasks
1
[JAX] Add MoE checkpoint names
#3589jberchtold-nvidia opened Sep 29, 2026Collaborator8 of 13 tasks
4
[Common] Compute BF16 GeLU/dGeLU on packed FP32 pairs in the MXFP8 quantize kernel
#3588tdophung opened Sep 29, 2026Collaborator13 tasks
4
Deterministic MoE Router Kernel By Removing Global Atomic Ops
#3587zhongbozhu opened Sep 29, 2026Collaborator13 tasks
2
[Common] Add NVTE_GROUPED_QUANTIZE_SM_MARGIN to reserve SMs from grouped quantize/swizzle
#3585phu0ngng opened Sep 29, 2026Collaborator8 of 13 tasks
17
[PyTorch] Deprecate grouped linear and MLP environment gates
#3584zhongbozhu opened Sep 29, 2026Collaborator13 tasks
22
perf: avoid quadratic saved-tensor restoration
#3583bradhilton opened Sep 29, 20264 of 6 tasks
Add context-parallel attention benchmark results
#3582sudhakarsingh27 opened Sep 29, 2026Member4 of 5 tasks
[JAX] Use target-based loss in collective GEMM gradient tests
#3580jberchtold-nvidia opened Sep 28, 2026Collaborator13 tasks
4