MoE training optimization

[ptrendx]

This is a tracking issue for the multiple efforts needed for the performance optimization of MoE training, with the focus on D2H Sync-Free MoE. All the problem sizes should be supplied from device buffers.

TE/common:

• GroupedTensor type [Common] NVTEGroupedTensor class and helpers #2388
• Quantization support for the GroupedTensor
  • FP8 per-tensor
  • MXFP8 [Common] MXFP8 kernel for grouped tensors #2586
  • NVFP4
    • Grouped amax kernel (Unofficial API) [PyTorch][NVFP4][MOE] NVFP4 Grouped Hadamard Amax Kernel #2351
    • Grouped quantization kernel & grouped RHT quantization kernel separately ((Unofficial API)) [PyTorch][NVFP4][MOE] NVFP4 Grouped Quantize with Hadamard Transform #2411
    • Grouped_amax and Grouped_quantize APIs with NVTEGroupedTensor
    • Grouped rowwise quantization fused with RHT+quantization fusion
• Grouped swizzle kernel
• GroupedGemm with GroupedTensor support
  • FP8 per-tensor
    • via cuBLAS [common] Add support for cuBLASLt GEMM for GroupedTensor #2502
    • via CUTLASS/cuDNN kernel
  • MXFP8
    • via cuBLAS (needs cuBLAS support)
    • via CUTLASS/cuDNN kernel Add device-Initiated Grouped GEMM supporting m_splits on device #2360
  • NVFP4
    • via cuBLAS (needs cuBLAS support)
    • via CUTLASS/cuDNN kernel
• Fused router optimization
• Activation support for the GroupedTensor
  • BF16 -> BF16 version
  • BF16 -> quantized format should be mostly covered by the effort needed to bring quantization
  • Ability to return amax together with the result for NVFP4 and FP8 current scaling quantization
• Padding in permutation [PyTorch] Fuse permute+pad and unpermute+unpad ops for FP8 optimization #1921
• Fusion of swizzle with quantization Add logic for block-scaled tensors with GEMM swizzled scales #2486
  • Changes to the NVTETensor/GroupedTensor
  • MXFP8 quantization support
  • NVFP4 quantization support
  • unswizzle kernel + grouped version (for checkpointing/debugging)

TE/pyTorch:

• Expose the grouped tensor type internally in PyTorch modules
• Expose the grouped tensor type externally [@timmoon10 is doubtful of feasibility]
  • Expose the grouped tensor type as pyTorch tensor
  • Enable grouped tensor input to GroupedLinear
• Enable single grouped tensor weight option in GroupedLinear
• Utilize preswizzled inputs in the gemm
• Changes to te.Sequential to enable grouped tensors
• End to end MoE support in TransformerLayer

TE/JAX:

• Triton binding [JAX] Triton binding #2437
• Router / TopK custom call [Draft] TopK Fusion to JAX #2385
• Permutation custom call [JAX] primitives for permutation on single GPU #2473
• Custom partitioning for grouped_quantize and grouped_gemm
• Code refactor:
  • GroupedScaledTensor/GroupedGEMM with first_dims and last_dims instead of group_sizes
  • Remove D2H in GroupedQuantizeFFI/GroupedGemmFFI
• Utilize preswizzled inputs in the gemm
• E2E
  • GroupedMLP
  • MaxText integration