48 Pull requests merged by 12 people

• [mxfp8 moe training] add per group blocked scale kernels

  #2886 merged Aug 28, 2025

• Rename to_float8 to from_hp

  #2893 merged Aug 28, 2025

• [mxfp8 moe training] add grouped gemm benchmark script

  #2882 merged Aug 28, 2025

• [mxfp8 moe training] refactor all var names with suffix _mx to _data for clarity

  #2879 merged Aug 28, 2025

• [fp8 blockwise] wrap triton quantization kernels in custom ops for torch.compile compatibility

  #2829 merged Aug 28, 2025

• exclude libcudart.so.13 from auditwheel repair to fix CUDA 13.0 wheel build

  #2892 merged Aug 28, 2025

• [CPU][float8] Add scaled_embedding_bag kernel

  #2686 merged Aug 28, 2025

• Support QAT int4 v1 path for BC

  #2888 merged Aug 28, 2025

• [fp8 blockwise] load 2d chunks for groupwise quant to enable coalesced gmem accesses

  #2827 merged Aug 27, 2025

• use shared bench + profile utils in blockwise fwd bwd bench script

  #2826 merged Aug 27, 2025

• integrate torch._scaled_mm into Float8BlockwiseLinear and add bench script

  #2785 merged Aug 27, 2025

• [moe fp8 training] fused reduction kernel along dim1 for 3d expert weights in backward

  #2865 merged Aug 27, 2025

• Update AWQ implementation to not use extra wrapper tensor subclass

  #2753 merged Aug 27, 2025

• [mxfp8 moe] add support for fbgemm 2d-3d mx8mx8bf16 grouped gemm

  #2848 merged Aug 27, 2025

• Introduce IntxOpaqueTensor to replace PackedInt8DynamicActivationIntxWeightLayout in AQT

  #2742 merged Aug 27, 2025

• [moe fp8 training] use transpose method when quantizing to avoid uncoalesced gmem accesses

  #2864 merged Aug 27, 2025

• [moe fp8 training] test and bench new faster method for per group rowwise scaling

  #2863 merged Aug 27, 2025

• [CPU] Introduce Int4OpaqueTensor to replace Int4CPULayout in AQT

  #2798 merged Aug 27, 2025

• Enable quantizing local checkpoints in model release script

  #2859 merged Aug 26, 2025

• Conditional ROCm kernel build

  #2839 merged Aug 26, 2025

• fix ci import error

  #2876 merged Aug 26, 2025

• TorchAOBaseTensor __tensor_flatten__ and __tensor_unflatten__ use…

  #2874 merged Aug 26, 2025

• Update IntxUnpackedTensor to support dynamic activation

  #2861 merged Aug 26, 2025

• release notes script: keep not user facing rows

  #2875 merged Aug 26, 2025

• Fix UT assertion error for int8 sdpa fusion

  #2816 merged Aug 26, 2025

• Add OPAQUE packing format

  #2878 merged Aug 26, 2025

• add mxfp8 to test_tp

  #2870 merged Aug 25, 2025

• Fix test tolerance

  #2871 merged Aug 25, 2025

• [mxfp8 moe training] Add mxfp8 to FSDP tests

  #2849 merged Aug 25, 2025

• bump version to 0.14.0

  #2872 merged Aug 25, 2025

• Add NVFP4 QAT

  #2666 merged Aug 25, 2025

• [reland] Refactor TorchAOBaseTensor for better BC (#2793)

  #2855 merged Aug 23, 2025

• Add test for lut based embedding quantization.

  #2825 merged Aug 23, 2025

• Add lut quantized embedding.

  #2824 merged Aug 23, 2025

• Fix autoquant after version util changes

  #2858 merged Aug 22, 2025

• Fix test_nvfp4_tensor.py merge conflict

  #2857 merged Aug 22, 2025

• Fix NVFP4 to_copy

  #2812 merged Aug 22, 2025

• Remove TORCH_VERSION_AT_LEAST* warnings when importing torch

  #2852 merged Aug 22, 2025

• Fix float8 + int4 QAT

  #2851 merged Aug 22, 2025

• Revert "Refactor TorchAOBaseTensor for better BC support"

  #2854 merged Aug 22, 2025

• fix incorrect torch version test

  #2786 merged Aug 22, 2025

• Refactor TorchAOBaseTensor for better BC support

  #2793 merged Aug 22, 2025

• Revert "Add the ops for groupwise lut quantization for embeding"

  #2850 merged Aug 22, 2025

• float8 kernel test: make more robust

  #2847 merged Aug 22, 2025

• [mxfp8 moe] replace per group scaling with conventional scaling

  #2841 merged Aug 22, 2025

• [mxfp8 moe] add compile test; add mxfp8 to bench script

  #2835 merged Aug 22, 2025

• mx: delete use_fp4_custom_triton_dequant_kernel option

  #2831 merged Aug 22, 2025

• mx: delete triton_f4_to_bf16 kernel

  #2830 merged Aug 22, 2025

17 Pull requests opened by 10 people

• Add Int4XPUTensorIntZP

  #2845 opened Aug 22, 2025

• Fix Llama4 example

  #2846 opened Aug 22, 2025

• use gcnArchName to get gpu_arch

  #2853 opened Aug 22, 2025

• [moe training] add test case for shared expert in distributed tests

  #2856 opened Aug 22, 2025

• Port metadata from the linear node onto the reference custom op for int4

  #2860 opened Aug 22, 2025

• Split implements and implements_torch_function

  #2866 opened Aug 24, 2025

• Move CPU kernels out of experimental

  #2868 opened Aug 25, 2025

• safetensors support

  #2881 opened Aug 26, 2025

• Fix Float8Tensor quantize op kernrel preference dispatch

  #2883 opened Aug 26, 2025

• Float8Tensor per row quantization pass bias to fbgemm kernel

  #2884 opened Aug 26, 2025

• [not for land] guard against fbgemm_gpu core dump

  #2887 opened Aug 27, 2025

• [mxfp8 moe training] add triton kernel for blocked swizzled 3d weight scales

  #2894 opened Aug 28, 2025

• Add tracking for new tensors, AQT and layouts

  #2895 opened Aug 28, 2025

• [mxfp8 moe training] use dim1 cast cuda kernel in bwd

  #2897 opened Aug 28, 2025

• hf integration doc page

  #2899 opened Aug 28, 2025

• [mxfp8 moe training] integrate triton kernels for converting scales to blocked format

  #2902 opened Aug 28, 2025

• [moe training] update tests + benchmarks with conditional runs based on SM arch; make test cases more comprehensive and consistent

  #2905 opened Aug 28, 2025

5 Issues closed by 2 people

• Using triton_op + wrap_triton introduces kernel performance regression

  #2898 closed Aug 28, 2025

• CI wheel build for CUDA 13.0 faliing with "ValueError: Cannot repair wheel, because required library "libcudart.so.13" could not be located"

  #2891 closed Aug 28, 2025

• fbgemm mxfp8 grouped gemm issues

  #2877 closed Aug 26, 2025

• NVFP4Tensor to_copy is wrong?

  #2811 closed Aug 22, 2025

• `torch_version_at_least` semantics are incorrect

  #2722 closed Aug 22, 2025

12 Issues opened by 8 people

• float8 rowwise scaled grouped mm doesn't support B200

  #2904 opened Aug 28, 2025

• Loading fp8-int4 model got unexpected keyword argument 'requires_grad'

  #2903 opened Aug 28, 2025

• Aborted (core dumped) when importing v0.13.0 RC

  #2901 opened Aug 28, 2025

• [CPU][FP8][Inductor] How to support fp8 quant for inductor on CPU

  #2896 opened Aug 28, 2025

• CI ROCM tests failing with "HW exception - GPU hang"

  #2890 opened Aug 27, 2025

• Add build docs instructions to contributing guide

  #2889 opened Aug 27, 2025

• Meta kernel for an AO op is incorrect

  #2885 opened Aug 26, 2025

• [MoE fp8 rowwise training] Runtime of quantizing 3d expert weights scales worse than linearly

  #2880 opened Aug 26, 2025

• Add HuggingFace integration doc page

  #2873 opened Aug 25, 2025

• CUDA OOM When Running AWQ int4 Quantized llama3.1-8b at Batch Size 1

  #2867 opened Aug 25, 2025

• Duplicated tests in test_mx_tensor.py and test_nvfp4_tensor.py?

  #2862 opened Aug 23, 2025

• RuntimeError: use_rowwise == false INTERNAL ASSERT FAILED at "/pytorch/aten/src/ATen/cuda/CUDABlas.cpp":1971

  #2843 opened Aug 21, 2025

25 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• refactor common used toy model

  #2729 commented on Aug 28, 2025 • 24 new comments

• Make SmoothQuant more General

  #2728 commented on Aug 28, 2025 • 23 new comments

• [CPU][FP8] Support FP8 SDPA for CPU backend

  #2689 commented on Aug 28, 2025 • 19 new comments

• Add Int4TilePackedTo4dTensor

  #2791 commented on Aug 28, 2025 • 9 new comments

• Molly/enable xpu ci

  #2814 commented on Aug 22, 2025 • 6 new comments

• Xpu ut/quantization

  #2796 commented on Aug 22, 2025 • 1 new comment

• test fsdp2 Moe

  #2842 commented on Aug 22, 2025 • 0 new comments

• [test only] testing adding optioanl tensor arg to float8 tensor

  #2840 commented on Aug 22, 2025 • 0 new comments

• Add MyPy support

  #2838 commented on Aug 26, 2025 • 0 new comments

• [wip] mx: expose a fast path for casting to fp4x2

  #2832 commented on Aug 22, 2025 • 0 new comments

• Add test function for the group wise lut quantization

  #2703 commented on Aug 23, 2025 • 0 new comments

• Replace `torch.norm` with `torch.linalg.vector_norm`

  #2660 commented on Aug 25, 2025 • 0 new comments

• Update test_bitpacking.cpp. Bug fix in test.

  #2633 commented on Aug 23, 2025 • 0 new comments

• Refactor Wanda for better readability

  #2538 commented on Aug 23, 2025 • 0 new comments

• config change to enable pre compute scale for fp8

  #2536 commented on Aug 23, 2025 • 0 new comments

• [CPU] Add support for dynamic float8 act float8 weight on CPU

  #2505 commented on Aug 28, 2025 • 0 new comments

• Groupwise low bit LUT based model quantization.

  #2407 commented on Aug 23, 2025 • 0 new comments

• add recipe config in aps for fp8

  #2322 commented on Aug 23, 2025 • 0 new comments

• Benchmark AWQ and SmoothQuant within vLLM ecosystem

  #2815 commented on Aug 28, 2025 • 0 new comments

• [roadmap/tracker] Low precision MoE training

  #2147 commented on Aug 27, 2025 • 0 new comments

• Deprecate and remove subclass.py, dynamic_quant.py, and weight_only.py

  #2745 commented on Aug 27, 2025 • 0 new comments

• Deprecation for Float8DynamicActivationFloat8WeightConfig and Float8WeightOnlyConfig and the models

  #2649 commented on Aug 27, 2025 • 0 new comments

• Implement an AWQ algorithm with dynamic activation quantization for ExecuTorch

  #2388 commented on Aug 27, 2025 • 0 new comments

• [moe training] fsdp2 bug for llama4 shared experts where num_experts=1

  #2673 commented on Aug 26, 2025 • 0 new comments

• [not user facing] Split `implements` and `implements_torch_function`

  #2707 commented on Aug 24, 2025 • 0 new comments